Examenos

71.1% · raw 71.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. OpenCode default agent harness; report Figure 8 unmodified-harness results have separate run settings. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

67.7% · raw 67.72 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. OpenCode default agent harness; report Figure 8 unmodified-harness results have separate run settings. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

51.8% · raw 51.78 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. OpenCode default agent harness; report Figure 8 unmodified-harness results have separate run settings. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

49.5% · raw 49.45 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Claude Code 2.1.126. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

47.4% · raw 47.4 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Mini-SWE-agent 2.4.5. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

47.3% · raw 47.26 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Codex 0.144.6. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

47.2% · raw 47.2 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Hermes 0.18.2. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

46.9% · raw 46.92 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Pi 0.80.10. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

41% · raw 41.04 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: OpenCode 1.17.8. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

69.2% · raw 69.24 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

76.5% · raw 76.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

73.2% · raw 73.19 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

18.9% · raw 18.88 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

19.9% · raw 19.85 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

60.7% · raw 60.67 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Terminus 2, three-hour task timeout. Report Tables 10/23 and chart agree on 60.67 (~60.7); HF card table says 66.70, retained as an alternative rather than silently reconciled. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

61.9% · raw 61.94 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

66.7% · raw 66.7 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact released BF16 card table. Conflicts/rounding versus detailed report remain non-headline alternatives; detailed report supplies the precise protocol. Raw card label: Terminal-Bench 2.1. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

61.8% · raw 61.8 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Codex 0.144.6. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

61.4% · raw 61.42 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Hermes 0.18.2. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

59.6% · raw 59.55 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Claude Code 2.1.126. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

55.1% · raw 55.06 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: OpenCode 1.17.8. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

54.7% · raw 54.68 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Mini-SWE-agent 2.4.5. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

48.3% · raw 48.31 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Harness: Pi 0.80.10. Unmodified harness system prompts and standard configurations. SWE-Pro 731 tasks / 2 repeats, no git history and explicit no-cheating/no-online-resources instruction; Terminal-Bench 2.1 89 tasks / 3 repeats. Distinct from the main NeMo Gym setup. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

1211.2 · raw 1211.2 Elo

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Stirrup agent, NVIDIA reference committee loosely following AA v2.1; not independent AA Elo. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1211 · raw 1211 Elo

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact released BF16 card table. Conflicts/rounding versus detailed report remain non-headline alternatives; detailed report supplies the precise protocol. Raw card label: GDPval-AA v2. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

35.6% · raw 35.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Stirrup agent, NVIDIA reference committee loosely following AA v2.1; not independent AA Elo. Raw source 0.356/0.532 normalized values ×100; not a pass rate. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

78.9% · raw 78.91 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

71.5% · raw 71.5 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

78% · raw 78 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact released BF16 card table. Conflicts/rounding versus detailed report remain non-headline alternatives; detailed report supplies the precise protocol. Raw card label: BrowseComp. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

72.4% · raw 72.35 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

47.1% · raw 47.13 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

47.3% · raw 47.27 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

15.1% · raw 15.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

88.2% · raw 88.22 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

89.5% · raw 89.52 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

90.8% · raw 90.78 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

10.9% · raw 10.86 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

31.8% · raw 31.79 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

31.8% · raw 31.84 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

46% · raw 46.01 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

88.4% · raw 88.36 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

26.2% · raw 26.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

26% · raw 25.98 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

70% · raw 69.96 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

69.2% · raw 69.24 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

29.6% · raw 29.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

80.8% · raw 80.82 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

78.2% · raw 78.16 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

65.1% · raw 65.08 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

68.1% · raw 68.12 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

59.2% · raw 59.15 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

32.9% · raw 32.85 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

85.6% · raw 85.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

29.2% · raw 29.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

62.8% · raw 62.82 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

52.1% · raw 52.14 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

51.1% · raw 51.14 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact released BF16 card table. Conflicts/rounding versus detailed report remain non-headline alternatives; detailed report supplies the precise protocol. Raw card label: OSWorld-Verified. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

68.6% · raw 68.63 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

68.8% · raw 68.81 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

68.6% · raw 68.62 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact released BF16 card table. Conflicts/rounding versus detailed report remain non-headline alternatives; detailed report supplies the precise protocol. Raw card label: AA-LCR v1.1. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

89.7% · raw 89.71 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

70.9% · raw 70.86 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

52.7% · raw 52.67 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

76.9% · raw 76.91 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

84.8 score · raw 84.8 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

34.5 · raw 34.48 index

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

90.3% · raw 90.25 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

51.2% · raw 51.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

57.4% · raw 57.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

73% · raw 73.04 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Bolt report Table 7: explicit BF16 columns; all other metrics inherit Super parent evaluation setup. GDPVal is a published index, not Elo or percent; its mapping to Table 10 normalization is not asserted. MCP-Atlas uses the official evaluator on 500 public tasks with reasoning on, not the full 1,000-task suite. WebArena uses 812 public tasks, screenshot vision, vLLM temp 1.0/top_p 0.95, 100 steps, internal string/URL/programmatic checks. BrowseComp uses Exa. Super values differ from its main launch run and remain alternatives. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

30% · raw 30 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact released BF16 card table. Conflicts/rounding versus detailed report remain non-headline alternatives; detailed report supplies the precise protocol. Raw card label: AA-Omniscience hallucination rate. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4 · raw 4.03 index

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact released BF16 card table. Conflicts/rounding versus detailed report remain non-headline alternatives; detailed report supplies the precise protocol. Raw card label: AA-Omniscience index. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

78% · raw 78 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 78. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

77.7% · raw 77.72 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 77.72. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1137 score · raw 1137 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 1137 (-) | 1252 (-) | 40. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1252 score · raw 1252 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 1137 (-) | 1252 (-) | 40. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

40 score · raw 40 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 1137 (-) | 1252 (-) | 40. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

32.4% · raw 32.44 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 32.44. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

6.6% · raw 6.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 6.6 (AA 6.6). Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1% · raw 1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 1. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

24.7% · raw 24.65 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Exact card offline snapshot; full source text: 24.65. Tuple labels and parentheses are not explained; no quantity/denominator inferred. Reported external AA values are quotes, not independent verification. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

22353 tokens/trajectory · raw 22353 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

29710 tokens/trajectory · raw 29710 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

25630 tokens/trajectory · raw 25630 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

5784 tokens/trajectory · raw 5784 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4028 tokens/trajectory · raw 4028 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

5426 tokens/trajectory · raw 5426 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

12173 tokens/trajectory · raw 12173 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

65560 tokens/trajectory · raw 65560 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

50% · raw 49.98 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Final BF16 reference, Table 23. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

49705 tokens/trajectory · raw 49705 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

6621 tokens/trajectory · raw 6621 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

6621 tokens/trajectory · raw 6621 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

15449 tokens/trajectory · raw 15449 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

59.6% · raw 59.58 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Final BF16 reference, Table 23. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

637 tokens/trajectory · raw 637 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

14383 tokens/trajectory · raw 14383 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

68050 tokens/trajectory · raw 68050 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

39253 tokens/trajectory · raw 39253 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4373 tokens/trajectory · raw 4373 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4373 tokens/trajectory · raw 4373 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

7347 tokens/trajectory · raw 7347 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

11839 tokens/trajectory · raw 11839 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4157 tokens/trajectory · raw 4157 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

7434 tokens/trajectory · raw 7434 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

5362 tokens/trajectory · raw 5362 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1860 tokens/trajectory · raw 1860 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

2688 tokens/trajectory · raw 2688 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

43.5% · raw 43.53 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Final BF16 reference, Table 23. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

236 tokens/trajectory · raw 236 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4579 tokens/trajectory · raw 4579 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

855 tokens/trajectory · raw 855 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

36564 tokens/trajectory · raw 36564 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

3365 tokens/trajectory · raw 3365 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4601 tokens/trajectory · raw 4601 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

10635 tokens/trajectory · raw 10635 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

10016 tokens/trajectory · raw 10016 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

3835 tokens/trajectory · raw 3835 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

3478 tokens/trajectory · raw 3478 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

52.3% · raw 52.32 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Final BF16 reference, Table 23. NeMo Gym; report Table 10 and §§3.8–3.9. AA Omniscience uses 600 public questions / 10 repeats with GPT-5.6 Luna judge. SciCode uses scientist background / 8 repeats and distinguishes subtask vs whole-problem accuracy. GPQA averages 8 repeats; APEX Shortlist 16 repeats, no tools. HLE text-only has 2,158 questions; full HLE 2,500 with one sample. Video-MME v2 has no subtitles; V* has no search tool. OSWorld excludes Google Drive tasks (361), screenshot-only pyautogui, 100 steps and partial-credit mean over 3 runs. GraphWalks averages F1 across BFS and parents on the same 950/1,150 prompts fitting ≤1M. IFBench is mean fraction of constraints satisfied under strict verification, not prompt all-pass. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1175 tokens/trajectory · raw 1175 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1175 tokens/trajectory · raw 1175 tokens/trajectory

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Mean BF16 completion tokens per trajectory from Table 23; exact benchmark and harness defined by the corresponding accuracy observation. No cash cost is inferred. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

62.7% · raw 62.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; empirical pass@1 averaged across candidate rollouts; 5 candidate trajectories, 89 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

71.9% · raw 71.9 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; LAV selected accuracy; 5 candidate trajectories, 89 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

71.3% · raw 71.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; empirical pass@1 averaged across candidate rollouts; 3 candidate trajectories, 500 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

72.4% · raw 72.4 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; LAV selected accuracy; 3 candidate trajectories, 500 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

52.2% · raw 52.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; empirical pass@1 averaged across candidate rollouts; 3 candidate trajectories, 731 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

54.6% · raw 54.6 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; LAV selected accuracy; 3 candidate trajectories, 731 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

69.8% · raw 69.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; empirical pass@1 averaged across candidate rollouts; 3 candidate trajectories, 300 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

71.7% · raw 71.7 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; LAV selected accuracy; 3 candidate trajectories, 300 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

24.6% · raw 24.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; empirical pass@1 averaged across candidate rollouts; 5 candidate trajectories, 113 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

34.5% · raw 34.5 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; LAV selected accuracy; 5 candidate trajectories, 113 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

71.5% · raw 71.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; empirical pass@1 averaged across candidate rollouts; 3 candidate trajectories, 400 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

79.8% · raw 79.8 %

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 9 final post-trained checkpoint; LAV selected accuracy; 3 candidate trajectories, 400 tasks. Same model verifies candidates using three criteria, both presentation orders and probabilistic pivot tournament. BrowseComp uses only a 400-problem tracking subset. Baseline is headline; selected result is an explicit alternative. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

5.6 tokens · raw 5.563 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

5.4 tokens · raw 5.441 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

4 tokens · raw 3.975 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

3.8 tokens · raw 3.817 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

4.8 tokens · raw 4.841 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4.8 tokens · raw 4.825 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

5.5 tokens · raw 5.548 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

5.4 tokens · raw 5.358 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

4.5 tokens · raw 4.526 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4.4 tokens · raw 4.378 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

5.5 tokens · raw 5.538 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

5.5 tokens · raw 5.496 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

4.8 tokens · raw 4.791 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4.8 tokens · raw 4.751 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

3.8 tokens · raw 3.781 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

3.7 tokens · raw 3.703 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

4.1 tokens · raw 4.102 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4 tokens · raw 4.045 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

5.2 tokens · raw 5.235 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

5 tokens · raw 5.041 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

4 tokens · raw 4.001 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

3.9 tokens · raw 3.873 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

4.7 tokens · raw 4.718 tokens

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 0 greedy decoding. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

4.6 tokens · raw 4.612 tokens

Alternative · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Table 9 Boosted MTP, SPEED-Bench qualitative split, draft length 7, reasoning on, temperature 1 probabilistic draft sampling. Native/pre-boosting checkpoint and unlabeled Figure 6 coordinates excluded. Greedy configuration is headline. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated. Alternative source/configuration; does not replace the existing headline.

522 tok/s · raw 522 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 10 BF16 bars; vLLM, 50K shared prefix, 20 turns of 2,048 input / 1,024 output tokens, ReplaySSM disabled. Aggregate throughput is output tokens/s/GPU subject to ≥50 tokens/s/user; minimum-latency point reports tokens/s/user, not latency seconds. Not OpenRouter API speed. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

509.1 tok/s · raw 509.131 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 10 BF16 bars; vLLM, 50K shared prefix, 20 turns of 2,048 input / 1,024 output tokens, ReplaySSM disabled. Aggregate throughput is output tokens/s/GPU subject to ≥50 tokens/s/user; minimum-latency point reports tokens/s/user, not latency seconds. Not OpenRouter API speed. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

1778 tok/s · raw 1778 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 10 BF16 bars; vLLM, 50K shared prefix, 20 turns of 2,048 input / 1,024 output tokens, ReplaySSM disabled. Aggregate throughput is output tokens/s/GPU subject to ≥50 tokens/s/user; minimum-latency point reports tokens/s/user, not latency seconds. Not OpenRouter API speed. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

587.5 tok/s · raw 587.453 tok/s

Headline · unknown effort · Own vendor

Source/record date: 2026-10-15

Harness: NVIDIA evaluation of the released BF16 post-trained model. Figure 10 BF16 bars; vLLM, 50K shared prefix, 20 turns of 2,048 input / 1,024 output tokens, ReplaySSM disabled. Aggregate throughput is output tokens/s/GPU subject to ≥50 tokens/s/user; minimum-latency point reports tokens/s/user, not latency seconds. Not OpenRouter API speed. Checked 15 October 2026; execution date and selectable model effort unpublished unless explicitly stated.

3.84 acronyms/response · raw 3.84 acronyms/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Acronyms: Two or more capital letters in a row (AI, USA) or dotted forms (U.S.). Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 3.84.

14.60 words/sentence · raw 14.6 words/sentence

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Avg. sentence length: Words divided by sentences. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 14.6.

14.90 bold spans/response · raw 14.9 bold spans/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Bold text: Uses of bold: each unbroken run of bold text counts once, so a single bolded word and a fully bolded sentence both count as 1. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 14.9.

4158 characters/response · raw 4158 characters/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Characters: Characters in the prose, spaces included, after code and URLs are removed. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 4,158.

10.40 colons/response · raw 10.4 colons/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Colons: Colons. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 10.4.

9.94 contractions/response · raw 9.94 contractions/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Contractions: Contractions such as “don’t”, “you’re” and “it’s”; possessives such as “John’s” do not count. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 9.94.

0.15 ellipses/response · raw 0.15 ellipses/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Ellipses: Three periods (...) or the ellipsis character (…). Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 0.15.

3.41 em dashes/response · raw 3.41 em dashes/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Em dashes: The em dash character (—); double hyphens do not count. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 3.41.

0.02 emojis/response · raw 0.02 emojis/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Emojis: Each emoji as the reader sees it. An emoji built from several characters, such as a family (👨‍👩‍👧), still counts as one emoji. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 0.02.

2% · raw 2 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Ends with a question: Whether the answer’s last character is a question mark, reported as the share of responses that do. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 2%.

0.07 exclamations/response · raw 0.07 exclamations/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Exclamations: Exclamation marks. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 0.07.

5.81 headings/response · raw 5.81 headings/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Headings: Headings of any level. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 5.81.

2.39 italic spans/response · raw 2.39 italic spans/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Italics: Uses of italics: each unbroken run of italic text counts once, so a single italicized word and a fully italicized sentence both count as 1. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 2.39.

3.78 lists/response · raw 3.78 lists/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Lists: Top-level bulleted or numbered lists; nested sub-lists do not count separately. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 3.78.

11.60 paragraphs/response · raw 11.6 paragraphs/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Paragraphs: Paragraphs of prose; headings, code and the items of a single-spaced list do not count. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 11.6.

7.56 parenthesis pairs/response · raw 7.56 parenthesis pairs/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Parentheses: Closed pairs of parentheses, nested pairs included; unmatched parentheses do not count. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 7.56.

3.14 question marks/response · raw 3.14 question marks/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Questions: Question marks. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 3.14.

13.50 quotation marks/response · raw 13.5 quotation marks/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Quotes: Quotation marks of any style, such as straight, curly, « » and „ “; apostrophes do not count. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 13.5.

45.70 sentences/response · raw 45.7 sentences/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Sentences: Sentences found by a rule-based splitter that handles abbreviations such as “Dr.”. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 45.7.

0.41 tables/response · raw 0.41 tables/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Tables: Tables. Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 0.41.

634 words/response · raw 634 words/response

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Programmatically measured Words: Tokens that contain a letter or digit; each Chinese, Japanese or Korean character counts as one word (but all questions are in English). Source-native average over responses to the same900questions; heatmap colors use pool-relative percentiles, but the visible cell numbers here are native quantities, not those color percentiles. The Characters definition explicitly removes code and URLs; exact feature-specific handling follows the methodology. Average sentence length is words per sentence, not per response. Ends with a question is the share of responses (%); bold/italic count spans, not words or characters. Descriptive communication style, not a quality ranking. Source label: 634.

24 percentile · raw 24 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Direct, confident positions and constructive pushback. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 24.

62 percentile · raw 62 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Rich, varied, well-crafted language. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 62.

48 percentile · raw 48 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Creative, exploratory, open-ended thinking. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 48.

59 percentile · raw 59 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Practical depth, evidence, actionable content. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 59.

97 percentile · raw 97 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Well-structured, scannable, answer-first. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 97.

63 percentile · raw 63 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Use of filler and generic language. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 63.

53 percentile · raw 53 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Truthful, acknowledges uncertainty, non-sycophantic. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 53.

36 percentile · raw 36 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Friendly, humorous, sincere, emotionally attuned. Source index is the mean of member-trait Elo z-scores followed by percentile ranking across the October2026 evaluated pool. Sycophantic is inverted only when contributing to Transparency; this does not invert the separately reported Sycophantic trait. A higher number means more of this named communication pattern, not better model quality. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 36.

66 percentile · raw 66 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Actionable Guidance writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 66.

52 percentile · raw 52 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Admits Uncertainty writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 52.

43 percentile · raw 43 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Answer First writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 43.

89 percentile · raw 89 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Bottom-Line Summary writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 89.

58 percentile · raw 58 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Confident writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 58.

49 percentile · raw 49 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Constructive Pushback writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 49.

46 percentile · raw 46 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Creative writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 46.

29 percentile · raw 29 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Direct writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 29.

49 percentile · raw 49 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Disclaims Assumptions writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 49.

72 percentile · raw 72 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Eloquent writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 72.

26 percentile · raw 26 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Emotionally Intelligent writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 26.

26 percentile · raw 26 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Evidence-Based writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 26.

55 percentile · raw 55 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Flags Nuances writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 55.

46 percentile · raw 46 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Friendly writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 46.

55 percentile · raw 55 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Humorous writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 55.

46 percentile · raw 46 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Information Dense writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 46.

55 percentile · raw 55 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Lexically Varied writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 55.

72 percentile · raw 72 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Literate writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 72.

69 percentile · raw 69 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Metaphorical writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 69.

35 percentile · raw 35 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Natural Sounding writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 35.

38 percentile · raw 38 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Open-Ended writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 38.

15 percentile · raw 15 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Opinionated writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 15.

38 percentile · raw 38 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Pragmatic writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 38.

66 percentile · raw 66 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Resource-Specific writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 66.

94 percentile · raw 94 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Scannable writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 94.

38 percentile · raw 38 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Sincere writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 38.

63 percentile · raw 63 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Slop writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 63.

92 percentile · raw 92 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Structured writing trait, judged pairwise among responses to the same questions. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 92.

41 percentile · raw 41 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Sycophantic writing trait, judged pairwise among responses to the same questions. Native trait direction: higher means more sycophancy; only its contribution to Transparency is inverted. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 41.

43 percentile · raw 43 percentile

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-19

Public rendered chart labels manually captured by the root agent, then transcribed once from those captures. Published 19 October 2026; source checked 20 October 2026; execution dates unpublished. No system prompt: one response to each of 900 English questions, 30 questions per each of 30 traits. This describes default communication style and trait prominence, not objective answer quality or accuracy. No AA API, internal app-state access, automated collector, paid export or inference was used. Prominence of the Truthful writing trait, judged pairwise among responses to the same questions. A judged writing/personality trait in this question population, not a general factual-accuracy benchmark. Blind pairwise LLM judging by Gemini3.8Flash(high), ClaudeSonnet5.5(medium), GLM5.3Flash(max), one randomly assigned judge per comparison with balanced assignments and randomized answer order; reasoning text hidden. Firth-penalized Bradley–Terry trait ratings; ties count as half-win. GPT-6.1Sol(max) anchors Elo1000. Published values here are rounded 1–100 pool-relative percentiles, NOT Elo, raw accuracy or a percentage of tasks passed. The current capture contains 36 judged configurations even though UI selection labels say38of38; the two extra communication-only models have no captured judged values. Underlying Elo and 95%CI values were not captured, and are not inferred from chart widths/colors; methodology uses1000bootstrap refits. The pool-relative scale can move when the evaluated pool changes without a new model execution. Source label: 43.

1162 · raw 1162 Elo

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-20

Manually observed public GDPval-AA v2.1 leaderboard table, full 298-configuration pool checked 20 October 2026. Exact source configuration: Nemotron 3.5 Super VL; creator: NVIDIA; source rank 127; source release-month label Oct 2026 is model release metadata, not execution date. Published rounded native Elo 1162; literal displayed CI -19 / +19. Native Elo is not a percentage and is not the normalized (Elo−500)/2000 chart quantity. This versioned metric is separate from older generic GDPval-AA Elo and vendor GDPval/GDPval-AA results; no old evidence is overwritten. Execution date, inference precision, unprinted effort budgets and task costs are unpublished in this captured table. Elo is relative to the dated comparison pool; drift from an earlier observation does not prove a new evaluation run. Source does not label an effort; unknown is preserved rather than copying a vendor default. AA exact model-page links to the approved reference checkpoint were verified in prior curated evidence; a BF16 reference link does not prove BF16 inference. Headline: highest published native Elo for this exact approved model; tied values prefer existing featured effort then source order.

26.5 · raw 26.4910658661689 index

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

-4.2 · raw -4.216666666666667 index

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

23.3% · raw 23.283333333333335 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

35.8% · raw 35.84618726917228 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

1162.4 · raw 1162.41 Elo

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

0% · raw 0 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

47.8% · raw 47.8009259259259 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

69% · raw 69 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

30.7% · raw 30.722891566265098 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

10.6% · raw 10.5714285714286 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

8% · raw 8 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.

14.8% · raw 14.818345798643588 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100. Partial reward, not strict all-assertion pass rate.

27.2 · raw 27.1898695604114 index

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

22.6 · raw 22.5653667608007 index

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

28.1 · raw 28.1134112990134 index

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

30.1 · raw 30.1143480206541 index

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

38.5 · raw 38.4726370481928 index

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget.

998.8 · raw 998.76 Elo

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. 95% CI 989.23 to 1008.29.

996.7 · raw 996.67 Elo

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. 95% CI 979.1 to 1014.24.

997.8 · raw 997.76 Elo

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. 95% CI 980.29 to 1015.22.

26.9% · raw 26.86868686868687 %

Headline · unknown effort · Independent evaluator

Source/record date: 2026-10-15

Artificial Analysis independent evaluation; page explicitly links the requested BF16 weights repository. Inference provider/weight precision and exact run date are not disclosed; this is not a claim that AA served BF16. Reasoning enabled but selectable effort is unspecified; the metadata reasoningTokens field is not treated as a verified inference budget. Published fraction on 0–1 scale ×100.