Examenos

64.8% · raw 64.8 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 111. Adaptive max, mean of five trials; Anthropic coding harness. See pp. 111–112 for evaluation-specific configuration.

83.7% · raw 83.7 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 111. Adaptive max, mean of five trials; Anthropic coding harness. See pp. 111–112 for evaluation-specific configuration.

30.7% · raw 30.7 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 111. Adaptive max, mean of five trials; Anthropic coding harness. See pp. 111–112 for evaluation-specific configuration.

46.4% · raw 46.4 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 113. Cognition official revision 1.1; Main 100 tasks, five runs/task. Claude Code for Claude; Codex for GPT. Max-effort summary, distinct from best-effort leaderboard.

45.8% · raw 45.8 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 111. Explicit xhigh score in summary table. Alternative evidence; existing headline/configuration retained. No averaging.

34.8% · raw 34.76 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 23868, "cost": 0.0596, "duration_min": 6.71, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

41.6% · raw 41.63 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 36172, "cost": 0.13, "duration_min": 8.67, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

41.9% · raw 41.89 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 55149, "cost": 0.256, "duration_min": 11.06, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

45.8% · raw 45.83 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 100847, "cost": 0.6297, "duration_min": 16.0, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

46.4% · raw 46.36 %

Alternative · max effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 181387, "cost": 1.3284, "duration_min": 23.55, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

58.4% · raw 58.4 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 114. Cognition revision 1.1; Extended 150 tasks, five runs/task; Claude Code. Max-effort result.

48.2% · raw 48.19 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 20240, "cost": 0.0486, "duration_min": 6.08, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

55% · raw 55.04 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 30398, "cost": 0.1021, "duration_min": 7.68, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

55.9% · raw 55.88 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 46347, "cost": 0.1998, "duration_min": 9.89, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

58.1% · raw 58.12 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 84943, "cost": 0.4876, "duration_min": 13.92, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

58.4% · raw 58.43 %

Alternative · max effort · Own vendor

Source/record date: 2026-10-07

Cognition primary public data.json, v1_1 weighted new_score ×100, not unweighted correct. Harness: claude-code. {"tokens": 157710, "cost": 1.0803, "duration_min": 20.95, "flagged_rate": null}. Cost is source-observed USD/task; not an API rate. Data checked 8 October; evaluation date unpublished. Anthropic launch headline retained; exact board precision/effort alternative, no averaging.

30.9% · raw 30.9 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 120. No tools; adaptive effort matrix, five trials. Published labeled chart values.

35.5% · raw 35.5 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 120. No tools; adaptive effort matrix, five trials. Published labeled chart values. Alternative evidence; existing headline/configuration retained. No averaging.

39.7% · raw 39.7 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 120. No tools; adaptive effort matrix, five trials. Published labeled chart values. Alternative evidence; existing headline/configuration retained. No averaging.

44.1% · raw 44.1 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 120. No tools; adaptive effort matrix, five trials. Published labeled chart values. Alternative evidence; existing headline/configuration retained. No averaging.

45.9% · raw 45.9 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 120. No tools; adaptive effort matrix, five trials. Published labeled chart values.

38.6% · raw 38.6 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 119. Web search/fetch and code execution; contamination-source restrictions; adaptive effort matrix, five trials. Web-search fees excluded from chart costs.

45% · raw 45 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 119. Web search/fetch and code execution; contamination-source restrictions; adaptive effort matrix, five trials. Web-search fees excluded from chart costs. Alternative evidence; existing headline/configuration retained. No averaging.

50.1% · raw 50.1 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 119. Web search/fetch and code execution; contamination-source restrictions; adaptive effort matrix, five trials. Web-search fees excluded from chart costs. Alternative evidence; existing headline/configuration retained. No averaging.

54.8% · raw 54.8 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 119. Web search/fetch and code execution; contamination-source restrictions; adaptive effort matrix, five trials. Web-search fees excluded from chart costs. Alternative evidence; existing headline/configuration retained. No averaging.

57.4% · raw 57.4 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 119. Web search/fetch and code execution; contamination-source restrictions; adaptive effort matrix, five trials. Web-search fees excluded from chart costs.

12.7% · raw 12.7 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch interactive chart labels; Haiku ten runs/task, no fallback; offline-egress Anthropic harness. Sonnet five runs/task with safeguard fallback. Do not merge independent AA fallback runs.

20.3% · raw 20.3 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch interactive chart labels; Haiku ten runs/task, no fallback; offline-egress Anthropic harness. Sonnet five runs/task with safeguard fallback. Do not merge independent AA fallback runs. Alternative evidence; existing headline/configuration retained. No averaging.

24.8% · raw 24.8 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch interactive chart labels; Haiku ten runs/task, no fallback; offline-egress Anthropic harness. Sonnet five runs/task with safeguard fallback. Do not merge independent AA fallback runs. Alternative evidence; existing headline/configuration retained. No averaging.

31.5% · raw 31.5 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch interactive chart labels; Haiku ten runs/task, no fallback; offline-egress Anthropic harness. Sonnet five runs/task with safeguard fallback. Do not merge independent AA fallback runs. Alternative evidence; existing headline/configuration retained. No averaging.

39.2% · raw 39.2 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch interactive chart labels; Haiku ten runs/task, no fallback; offline-egress Anthropic harness. Sonnet five runs/task with safeguard fallback. Do not merge independent AA fallback runs.

20.6% · raw 20.6 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 116. Haiku ten runs/task (700 trials), no fallback; 2.3% ended by safeguards. Offline-egress Anthropic harness; peers five runs/task.

43.8% · raw 43.8 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 117. 34 tasks, five runs/task, 20-hour wall-clock budget, Proximal agent harness. No unapproved Fable 5 peer attached to Fable 5.1.

82% · raw 82 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 117. Hidden-test pass rate, mini-swe-agent, no six-hour timeout. 166 of 200 tasks; remove 34 with reference pass rate below 0.9; only reference-passing tests. Distinct from full-set Almost@1.

64.3% · raw 64.3 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 121. 100 tasks; 980k-token budget, Anthropic web/code agent; final report file only. Opus 4.6 grader, five grading runs/response; not comparable to original Gemini 3 Pro grading.

72.4% · raw 72.4 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 121. 100 tasks; 980k-token budget, Anthropic web/code agent; final report file only. Opus 4.6 grader, five grading runs/response; not comparable to original Gemini 3 Pro grading. Alternative evidence; existing headline/configuration retained. No averaging.

77.8% · raw 77.8 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 121. 100 tasks; 980k-token budget, Anthropic web/code agent; final report file only. Opus 4.6 grader, five grading runs/response; not comparable to original Gemini 3 Pro grading. Alternative evidence; existing headline/configuration retained. No averaging.

80.5% · raw 80.5 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 121. 100 tasks; 980k-token budget, Anthropic web/code agent; final report file only. Opus 4.6 grader, five grading runs/response; not comparable to original Gemini 3 Pro grading. Alternative evidence; existing headline/configuration retained. No averaging.

81.5% · raw 81.5 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 121. 100 tasks; 980k-token budget, Anthropic web/code agent; final report file only. Opus 4.6 grader, five grading runs/response; not comparable to original Gemini 3 Pro grading.

3.5% · raw 3.5 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 123. Soft F1, 980k-token task budget; Anthropic harness and Opus 4.8 grader. Distinct from prior Opus 4.6/frozen-index or Perplexity GPT-5.4 grading.

12.9% · raw 12.9 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 123. Soft F1, 980k-token task budget; Anthropic harness and Opus 4.8 grader. Distinct from prior Opus 4.6/frozen-index or Perplexity GPT-5.4 grading. Alternative evidence; existing headline/configuration retained. No averaging.

37.3% · raw 37.3 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 123. Soft F1, 980k-token task budget; Anthropic harness and Opus 4.8 grader. Distinct from prior Opus 4.6/frozen-index or Perplexity GPT-5.4 grading. Alternative evidence; existing headline/configuration retained. No averaging.

45.9% · raw 45.9 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 123. Soft F1, 980k-token task budget; Anthropic harness and Opus 4.8 grader. Distinct from prior Opus 4.6/frozen-index or Perplexity GPT-5.4 grading. Alternative evidence; existing headline/configuration retained. No averaging.

49.9% · raw 49.9 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 123. Soft F1, 980k-token task budget; Anthropic harness and Opus 4.8 grader. Distinct from prior Opus 4.6/frozen-index or Perplexity GPT-5.4 grading.

46.4% · raw 46.4 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 124. 100 tasks, no tools. Claude adaptive max, mean five runs; Gemini 3.5 Flash grader. Non-Claude values reproduced from Surge board; their effort is not specified in this figure.

86.2% · raw 86.2 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 124. 100 tasks, five runs at max; image files, Python libraries and image cropping tool; Gemini 3.5 Flash grader.

67% · raw 67 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 126. Random 1,000-file subset of 17,900 Vision2Code files, five runs at max. Raw voxel IoU 0–1 multiplied by 100; tools variant has image files/Python/cropping. Distinct from full-set BenchCAD.

87% · raw 87 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 126. Random 1,000-file subset of 17,900 Vision2Code files, five runs at max. Raw voxel IoU 0–1 multiplied by 100; tools variant has image files/Python/cropping. Distinct from full-set BenchCAD.

42% · raw 42 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch chart exact labels; 82/108 offline tasks, no VM internet, 1080p, 500 actions, five attempts/task; Opus 4.8 grader. Partial checkpoint credit; distinct from strict completion and full OSWorld.

53.3% · raw 53.3 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch chart exact labels; 82/108 offline tasks, no VM internet, 1080p, 500 actions, five attempts/task; Opus 4.8 grader. Partial checkpoint credit; distinct from strict completion and full OSWorld. Alternative evidence; existing headline/configuration retained. No averaging.

61.3% · raw 61.3 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch chart exact labels; 82/108 offline tasks, no VM internet, 1080p, 500 actions, five attempts/task; Opus 4.8 grader. Partial checkpoint credit; distinct from strict completion and full OSWorld. Alternative evidence; existing headline/configuration retained. No averaging.

67.6% · raw 67.6 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch chart exact labels; 82/108 offline tasks, no VM internet, 1080p, 500 actions, five attempts/task; Opus 4.8 grader. Partial checkpoint credit; distinct from strict completion and full OSWorld. Alternative evidence; existing headline/configuration retained. No averaging.

72.4% · raw 72.4 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Official launch chart exact labels; 82/108 offline tasks, no VM internet, 1080p, 500 actions, five attempts/task; Opus 4.8 grader. Partial checkpoint credit; distinct from strict completion and full OSWorld.

37.1% · raw 37.1 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 128. 82/108 offline tasks, five attempts/task, max; all checkpoints must pass. No VM internet, 1080p, 500 actions, Opus 4.8 grader. Not partial-credit score.

73.5% · raw 73.5 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 130. Relevant documents preselected and supplied as extracted text; max, mean five runs. Pro is the 133-question subset; not end-to-end document retrieval.

60.3% · raw 60.3 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 130. Relevant documents preselected and supplied as extracted text; max, mean five runs. Pro is the 133-question subset; not end-to-end document retrieval.

1125 · raw 1125 Elo

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Vendor-reported GDPval-AA v2.1 native Elo; exact launch chart labels, rounded vendor evidence kept apart from independently sourced AA decimal values.

1277 · raw 1277 Elo

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Vendor-reported GDPval-AA v2.1 native Elo; exact launch chart labels, rounded vendor evidence kept apart from independently sourced AA decimal values. Alternative evidence; existing headline/configuration retained. No averaging.

1420 · raw 1420 Elo

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Vendor-reported GDPval-AA v2.1 native Elo; exact launch chart labels, rounded vendor evidence kept apart from independently sourced AA decimal values. Alternative evidence; existing headline/configuration retained. No averaging.

1513 · raw 1513 Elo

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Vendor-reported GDPval-AA v2.1 native Elo; exact launch chart labels, rounded vendor evidence kept apart from independently sourced AA decimal values. Alternative evidence; existing headline/configuration retained. No averaging.

1620 · raw 1620 Elo

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card / launch. Vendor-reported GDPval-AA v2.1 native Elo; exact launch chart labels, rounded vendor evidence kept apart from independently sourced AA decimal values.

1578 · raw 1578 Elo

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 131. Vendor-reported AA-Briefcase v1.1 native Elo, max; independently sourced AA decimal values remain separate.

1372 · raw 1372 Elo

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 131. Medium-effort result, native Elo. Alternative evidence; existing headline/configuration retained. No averaging.

59.9% · raw 59.9 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled. Low–xhigh one trial/level (5 Oct); max mean five trials (2 Oct). Raw score is separate.

60.2% · raw 60.2 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled. Low–xhigh one trial/level (5 Oct); max mean five trials (2 Oct). Raw score is separate. Alternative evidence; existing headline/configuration retained. No averaging.

60.3% · raw 60.3 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled. Low–xhigh one trial/level (5 Oct); max mean five trials (2 Oct). Raw score is separate. Alternative evidence; existing headline/configuration retained. No averaging.

61.1% · raw 61.1 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled. Low–xhigh one trial/level (5 Oct); max mean five trials (2 Oct). Raw score is separate. Alternative evidence; existing headline/configuration retained. No averaging.

61.6% · raw 61.6 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled. Low–xhigh one trial/level (5 Oct); max mean five trials (2 Oct). Raw score is separate.

57.9% · raw 57.9 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled; five runs/level. Low–xhigh 4 Oct; max 1–2 Oct. Raw score is separate.

59.9% · raw 59.9 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled; five runs/level. Low–xhigh 4 Oct; max 1–2 Oct. Raw score is separate. Alternative evidence; existing headline/configuration retained. No averaging.

61.3% · raw 61.3 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled; five runs/level. Low–xhigh 4 Oct; max 1–2 Oct. Raw score is separate. Alternative evidence; existing headline/configuration retained. No averaging.

61% · raw 61 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled; five runs/level. Low–xhigh 4 Oct; max 1–2 Oct. Raw score is separate. Alternative evidence; existing headline/configuration retained. No averaging.

64.8% · raw 64.8 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. Length-adjusted; Opus 4.8 grader, safeguards enabled; five runs/level. Low–xhigh 4 Oct; max 1–2 Oct. Raw score is separate.

64.2% · raw 64.2 %

Alternative · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 133. Length-adjusted repeat on 4 October, mean thirteen runs; alternative to headline five-run 1–2 October result. Alternative evidence; existing headline/configuration retained. No averaging.

64.2% · raw 64.2 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 132. Raw score before response-length adjustment; max, Opus 4.8 grader, safeguards enabled.

71% · raw 71 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 132. Raw score before response-length adjustment; max, Opus 4.8 grader, safeguards enabled.

17.8% · raw 17.8 %

Alternative · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. 100 EHR tasks; pass@1 mean five attempts/task, Opus 5 grader; safeguards enabled. Low–xhigh 5 October, max 1–2 October.

25.2% · raw 25.2 %

Alternative · medium effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. 100 EHR tasks; pass@1 mean five attempts/task, Opus 5 grader; safeguards enabled. Low–xhigh 5 October, max 1–2 October. Alternative evidence; existing headline/configuration retained. No averaging.

31.6% · raw 31.6 %

Alternative · high effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. 100 EHR tasks; pass@1 mean five attempts/task, Opus 5 grader; safeguards enabled. Low–xhigh 5 October, max 1–2 October. Alternative evidence; existing headline/configuration retained. No averaging.

35.8% · raw 35.8 %

Alternative · xhigh effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. 100 EHR tasks; pass@1 mean five attempts/task, Opus 5 grader; safeguards enabled. Low–xhigh 5 October, max 1–2 October. Alternative evidence; existing headline/configuration retained. No averaging.

43% · raw 43 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 134. 100 EHR tasks; pass@1 mean five attempts/task, Opus 5 grader; safeguards enabled. Low–xhigh 5 October, max 1–2 October.

87.8% · raw 87.8 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 135. GMMLU: 42 languages, one trial; MILU: 11 languages, five trials. Max; rare blocked examples excluded (<0.1% GMMLU / <0.2% MILU).

87.6% · raw 87.6 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 136. GMMLU: 42 languages, one trial; MILU: 11 languages, five trials. Max; rare blocked examples excluded (<0.1% GMMLU / <0.2% MILU).

67.7% · raw 67.7 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 137. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

56.2% · raw 56.2 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 137. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

25.4% · raw 25.4 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 138. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

58% · raw 58 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 138. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

33.1% · raw 33.1 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 138. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

45.3% · raw 45.3 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 138. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

43.5% · raw 43.5 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 138. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

58% · raw 58 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 139. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

64.1% · raw 64.1 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 139. Life sciences, adaptive max; Haiku biology safeguards disabled to measure underlying capability, not deployed behavior. Five runs unless specified: sequence generation and biomedical image analysis one; library ranking three. Figure 8.13.A peer values are rounded raw fractions ×100; text values take precedence where supplied. ADME 504 questions; biomedical scores normalized to competition-best results.

48% · raw 48 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 13. 350-question multimodal virology troubleshooting; raw accuracy 0–1 multiplied by 100. Biology safeguards off; capability assessment.

167.1 score · raw 167.11 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 20. Anthropic internal benchmark IRT fit; native index. Not Epoch’s public ECI.

6.6 flags · raw 6.56 flags

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 25. 41 V8 vulnerabilities from 2024 onward; 300-turn uniform author harness, five trials/arm. Cyber safeguards off, escape classifiers on. Mean flags native 0–16; Cap rate randomly sampled trials; ACE counts both plain and AutoNudge, 410 trials total. Distinct from Jun–Aug 2026 challenge set.

49% · raw 49 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 25. 41 V8 vulnerabilities from 2024 onward; 300-turn uniform author harness, five trials/arm. Cyber safeguards off, escape classifiers on. Mean flags native 0–16; Cap rate randomly sampled trials; ACE counts both plain and AutoNudge, 410 trials total. Distinct from Jun–Aug 2026 challenge set.

4 exploits · raw 4 exploits

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 25. 41 V8 vulnerabilities from 2024 onward; 300-turn uniform author harness, five trials/arm. Cyber safeguards off, escape classifiers on. Mean flags native 0–16; Cap rate randomly sampled trials; ACE counts both plain and AutoNudge, 410 trials total. Distinct from Jun–Aug 2026 challenge set.

5.2 flags · raw 5.15 flags

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 25. 41 V8 vulnerabilities from 2024 onward; 300-turn uniform author harness, five trials/arm. Cyber safeguards off, escape classifiers on. Mean flags native 0–16; Cap rate randomly sampled trials; ACE counts both plain and AutoNudge, 410 trials total. Distinct from Jun–Aug 2026 challenge set.

38% · raw 38 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 25. 41 V8 vulnerabilities from 2024 onward; 300-turn uniform author harness, five trials/arm. Cyber safeguards off, escape classifiers on. Mean flags native 0–16; Cap rate randomly sampled trials; ACE counts both plain and AutoNudge, 410 trials total. Distinct from Jun–Aug 2026 challenge set.

1% · raw 1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 24. Published 1.0% (4/410), both plain and AutoNudge arms; 41 vulnerabilities, cyber safeguards disabled.

58.2% · raw 58.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 28. 831 entrypoints, 228 OSS-Fuzz projects; August 2026 revised harness, not older OSS-Fuzz harness. Pass@1 vulnerability identification.

3 exploits · raw 3 exploits

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 28. 831 entrypoints, 228 projects; revised August 2026 harness. Published successful control-flow hijack count.

79 exploits · raw 79 exploits

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 29. 869 known vulnerabilities; successful exploits must capture dynamic secret flag and use intended vulnerability. Wall-clock budget per task; mitigations disabled as specified on p. 29. Count, not percent.

82 exploits · raw 82 exploits

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 29. 869 known vulnerabilities; successful exploits must capture dynamic secret flag and use intended vulnerability. Wall-clock budget per task; mitigations disabled as specified on p. 29. Count, not percent.

0.7% · raw 0.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 50. Q1+Q2 2026 IPI suite; probability attacker succeeds in k attempts; extended thinking, no product-specific prompt-injection protections. Do not mix with older challenge sets. Gemini/Kimi effort unspecified.

5.5% · raw 5.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 50. Q1+Q2 2026 IPI suite; probability attacker succeeds in k attempts; extended thinking, no product-specific prompt-injection protections. Do not mix with older challenge sets. Gemini/Kimi effort unspecified.

7.1% · raw 7.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 50. Q1+Q2 2026 IPI suite; probability attacker succeeds in k attempts; extended thinking, no product-specific prompt-injection protections. Do not mix with older challenge sets. Gemini/Kimi effort unspecified.

0.2% · raw 0.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 50. Domain-specific attack success rate, Q1+Q2 2026 suite; no product prompt-injection protections.

4% · raw 4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 50. Domain-specific attack success rate, Q1+Q2 2026 suite; no product prompt-injection protections.

24.4% · raw 24.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 50. Domain-specific attack success rate, Q1+Q2 2026 suite; no product prompt-injection protections.

98.4% · raw 98.39 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 31. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

99.7% · raw 99.71 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 31. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0.2% · raw 0.17 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 32. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0.8% · raw 0.82 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 32. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

99.3% · raw 99.29 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 34. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0% · raw 0.04 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 34. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

99.9% · raw 99.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 34. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 34. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

99.6% · raw 99.61 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 36. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 36. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 36. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0.4% · raw 0.41 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 36. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

97.4% · raw 97.43 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 39. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0% · raw 0.03 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 39. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

99.8% · raw 99.83 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 39. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

0.1% · raw 0.11 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 39. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

96% · raw 96 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 35. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

99% · raw 99 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 35. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

70% · raw 70 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 37. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

90% · raw 90 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 37. Anthropic internal evaluation; API without system prompt and Claude.ai default system prompt are distinct. Updated evaluation set as of Haiku 5.5 card; do not infer real-world incidence. Reported percentage, confidence intervals in source.

55.6% · raw 55.56 %

Headline · none effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 43. No system prompt; thinking disabled where supported. Accuracy, not signed bias.

99.7% · raw 99.65 %

Headline · none effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 43. No system prompt; thinking disabled where supported. Accuracy, not signed bias.

84.3% · raw 84.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 48. Product agent harness; malicious computer-use is average of with/without thinking for Haiku/Sonnet, thinking-only for Opus. Specific evaluated task set, not general safety.

98.9% · raw 98.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 48. Product agent harness; malicious computer-use is average of with/without thinking for Haiku/Sonnet, thinking-only for Opus. Specific evaluated task set, not general safety.

82.6% · raw 82.59 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 49. Product agent harness; malicious computer-use is average of with/without thinking for Haiku/Sonnet, thinking-only for Opus. Specific evaluated task set, not general safety.

0.1% · raw 0.08 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 51. Adaptive red-team attempt-level attack success rate; 200 attempts/scenario. Coding 40 scenarios, GUI 14. Haiku has no model fallback; larger Claude models may fall back under protections. Scenario counts preserved in source.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 51. Adaptive red-team attempt-level attack success rate; 200 attempts/scenario. Coding 40 scenarios, GUI 14. Haiku has no model fallback; larger Claude models may fall back under protections. Scenario counts preserved in source.

0.1% · raw 0.07 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 53. Adaptive red-team attempt-level attack success rate; 200 attempts/scenario. Coding 40 scenarios, GUI 14. Haiku has no model fallback; larger Claude models may fall back under protections. Scenario counts preserved in source.

0.1% · raw 0.07 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 53. Adaptive red-team attempt-level attack success rate; 200 attempts/scenario. Coding 40 scenarios, GUI 14. Haiku has no model fallback; larger Claude models may fall back under protections. Scenario counts preserved in source.

0% · raw 0 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 54. 110 scenarios, ten attacks/scenario; attempt-level ASR, 0/110 successful scenarios; auto mode on. Saturated evaluation scheduled for deprecation.

3.2 score · raw 3.19 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

3 score · raw 2.96 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

3 score · raw 3.02 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

4 score · raw 4 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.8 score · raw 1.83 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.134 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.2 score · raw 1.2 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.075 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.3 score · raw 1.276 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 61. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.7 score · raw 2.69 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.6 score · raw 1.56 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.125 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.061 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.017 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.028 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.004 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.017 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.047 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.054 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.099 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.2 score · raw 1.22 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 62. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.2 score · raw 2.203 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 64. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.1 score · raw 2.06 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 64. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.4 score · raw 2.36 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 64. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.4 score · raw 1.36 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.8 score · raw 1.77 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.038 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.3 score · raw 1.34 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.9 score · raw 1.88 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2 score · raw 2.04 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.7 score · raw 1.73 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.9 score · raw 1.89 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 66. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.3 score · raw 2.32 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 67. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.3 score · raw 1.34 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 67. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.076 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 67. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.034 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.03 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.2 score · raw 1.16 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.03 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.086 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.2 score · raw 1.2 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.114 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1 score · raw 1.045 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.7 score · raw 1.71 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 69. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.7 score · raw 1.65 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 70. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.1 score · raw 2.09 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 70. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

6.2 score · raw 6.22 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 71. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

6.1 score · raw 6.12 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 71. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.9 score · raw 1.85 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 71. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.6 score · raw 2.58 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 71. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

6.7 score · raw 6.74 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

7 score · raw 7.01 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

6.5 score · raw 6.45 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

6.5 score · raw 6.45 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.5 score · raw 2.49 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

7.4 score · raw 7.39 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

5.9 score · raw 5.94 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.1 score · raw 1.082 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

2.1 score · raw 2.089 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

1.6 score · raw 1.58 score

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 73. Automated internal behavioral audit; model-judged native 1–10 scale, not a rate. Up to ~4,100 investigations / ~1,850 seeds; some metrics apply only to scenario subsets. Simulated settings may not predict deployment; see pp. 58–76.

8.7% · raw 8.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 75. 28 handmade scenarios ×14 trials (392 runs). Attempt severity is judged; freshly re-evaluated on current version, distinct from 1,753-scenario opportunistic sandbox test.

6.6% · raw 6.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 75. 28 handmade scenarios ×14 trials (392 runs). Attempt severity is judged; freshly re-evaluated on current version, distinct from 1,753-scenario opportunistic sandbox test.

2% · raw 2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 75. 28 handmade scenarios ×14 trials (392 runs). Attempt severity is judged; freshly re-evaluated on current version, distinct from 1,753-scenario opportunistic sandbox test.

30% · raw 30 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 78. Same 1,753 deliberately weak sandboxes, one attempt/scenario. Haiku 524 attempted crossings; incomplete instrumentation may underestimate. Distinct from 28-scenario severity evaluation.

96.8% · raw 96.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 81. Public MASK split; fraction of scored pressured answers consistent with elicited belief. Evasions/inconsistent beliefs excluded.

99% · raw 99 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 82. Public MASK scored subset by pressure category; figure rounded to whole percent except explicit text precision for Haiku disinformation. Not separate global honesty score.

88.5% · raw 88.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 82. Public MASK scored subset by pressure category; figure rounded to whole percent except explicit text precision for Haiku disinformation. Not separate global honesty score.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 82. Public MASK scored subset by pressure category; figure rounded to whole percent except explicit text precision for Haiku disinformation. Not separate global honesty score.

98% · raw 98 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 82. Public MASK scored subset by pressure category; figure rounded to whole percent except explicit text precision for Haiku disinformation. Not separate global honesty score.

97% · raw 97 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 82. Public MASK scored subset by pressure category; figure rounded to whole percent except explicit text precision for Haiku disinformation. Not separate global honesty score.

17.3% · raw 17.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 83. Held-out coding tasks with reachable leaked solutions; substantial use without disclosure, as judged by Anthropic.

95.9% · raw 95.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 84. Summary request on transcripts containing concealed repository changes; spontaneous disclosure rate.

57% · raw 57 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 92. 39 simulated confidential jobs; Opus 5.5 counterpart presses for up to 20 turns. End-of-conversation fraction without leak. Reply-only and reasoning-visible conditions use different environments.

3% · raw 3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card, p. 92. 39 simulated confidential jobs; Opus 5.5 counterpart presses for up to 20 turns. End-of-conversation fraction without leak. Reply-only and reasoning-visible conditions use different environments.

94% · raw 94 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

89% · raw 89 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

98% · raw 98 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

78% · raw 78 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

78% · raw 78 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

99% · raw 99 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

98% · raw 98 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

87% · raw 87 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

88% · raw 88 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

96% · raw 96 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

94% · raw 94 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

95% · raw 95 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

98% · raw 98 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

90% · raw 90 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

90% · raw 90 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 33. Appropriate throughout conversation; API without system prompt versus Claude.ai default prompt; synthetic user Opus 4.6, category-specific rubrics. Updated set; scores are not comparable across categories.

97.7% · raw 97.7 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 41. Pairwise political evaluation, 1,350 paired prompts across 150 topics and nine tasks; surfaces have distinct system prompts. Combined Fable/Mythos bar not assigned to either model.

99.1% · raw 99.1 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 41. Pairwise political evaluation, 1,350 paired prompts across 150 topics and nine tasks; surfaces have distinct system prompts. Combined Fable/Mythos bar not assigned to either model.

35.9% · raw 35.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 41. Pairwise political evaluation, 1,350 paired prompts across 150 topics and nine tasks; surfaces have distinct system prompts. Combined Fable/Mythos bar not assigned to either model.

67.8% · raw 67.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 41. Pairwise political evaluation, 1,350 paired prompts across 150 topics and nine tasks; surfaces have distinct system prompts. Combined Fable/Mythos bar not assigned to either model.

3.3% · raw 3.3 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 42. Pairwise political evaluation, 1,350 paired prompts across 150 topics and nine tasks; surfaces have distinct system prompts. Combined Fable/Mythos bar not assigned to either model.

5% · raw 5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 42. Pairwise political evaluation, 1,350 paired prompts across 150 topics and nine tasks; surfaces have distinct system prompts. Combined Fable/Mythos bar not assigned to either model.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 45. Single-turn election evaluation; 300 harmful and 300 benign prompts; API without prompt versus Claude.ai default prompt. Specific evaluation, not general election capability.

0.3% · raw 0.33 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 45. Single-turn election evaluation; 300 harmful and 300 benign prompts; API without prompt versus Claude.ai default prompt. Specific evaluation, not general election capability.

100% · raw 100 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 45. Single-turn election evaluation; 300 harmful and 300 benign prompts; API without prompt versus Claude.ai default prompt. Specific evaluation, not general election capability.

0.3% · raw 0.33 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 45. Single-turn election evaluation; 300 harmful and 300 benign prompts; API without prompt versus Claude.ai default prompt. Specific evaluation, not general election capability.

99% · raw 99 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 45. Appropriate throughout multi-turn election conversations; synthetic users/rubric. API without system prompt versus Claude.ai default prompt.

98% · raw 98 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 45. Appropriate throughout multi-turn election conversations; synthetic users/rubric. API without system prompt versus Claude.ai default prompt.

5.9% · raw 5.9 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 88. Successful harmful side task without Opus 4.8 monitor flag, best optimized prompt per model/setting; Haiku one run/task pair, peers five. Hidden reasoning withheld from monitor; other conditions expose full trajectory. Revised prompt optimization may change earlier scores.

0.4% · raw 0.4 %

Headline · none effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 88. Successful harmful side task without Opus 4.8 monitor flag, best optimized prompt per model/setting; Haiku one run/task pair, peers five. Hidden reasoning withheld from monitor; other conditions expose full trajectory. Revised prompt optimization may change earlier scores.

0.1% · raw 0.1 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 88. Successful harmful side task without Opus 4.8 monitor flag, best optimized prompt per model/setting; Haiku one run/task pair, peers five. Hidden reasoning withheld from monitor; other conditions expose full trajectory. Revised prompt optimization may change earlier scores.

1.2% · raw 1.2 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 90. 24 task pairs, >=25 attempts/pair, best optimized prompt per setting, Opus 4.8 monitor. Haiku API safety interventions on; ended episodes count unsuccessful. Different from Shade IPI attacks.

5.2% · raw 5.2 %

Headline · none effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 90. 24 task pairs, >=25 attempts/pair, best optimized prompt per setting, Opus 4.8 monitor. Haiku API safety interventions on; ended episodes count unsuccessful. Different from Shade IPI attacks.

2.3% · raw 2.3 %

Headline · low effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 90. 24 task pairs, >=25 attempts/pair, best optimized prompt per setting, Opus 4.8 monitor. Haiku API safety interventions on; ended episodes count unsuccessful. Different from Shade IPI attacks.

0% · raw 0 %

Headline · max effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 90. 24 task pairs, >=25 attempts/pair, best optimized prompt per setting, Opus 4.8 monitor. Haiku API safety interventions on; ended episodes count unsuccessful. Different from Shade IPI attacks.

12% · raw 12 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 80. Anthropic evaluation on public split, no tools; not independent AA private evaluation. Published fractions ×100; net = correct minus incorrect in percentage points. Chart precision retained.

44% · raw 44 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 80. Anthropic evaluation on public split, no tools; not independent AA private evaluation. Published fractions ×100; net = correct minus incorrect in percentage points. Chart precision retained.

32% · raw 32 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-07

Anthropic Haiku 5.5 system card p. 80. Anthropic evaluation on public split, no tools; not independent AA private evaluation. Published fractions ×100; net = correct minus incorrect in percentage points. Chart precision retained.

43.4 · raw 43.3950199670746 index

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

41.2 · raw 41.2489304677024 index

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

37.8 · raw 37.8240211671094 index

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

34.5 · raw 34.4646519474657 index

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

29.4 · raw 29.4494981304947 index

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1577.7 · raw 1577.72 Elo

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1532.7 · raw 1532.67 Elo

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1442.6 · raw 1442.55 Elo

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1371.7 · raw 1371.7 Elo

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1111.6 · raw 1111.58 Elo

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1620.1 · raw 1620.05 Elo

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1510.8 · raw 1510.76 Elo

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1420 · raw 1420.03 Elo

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1276.8 · raw 1276.84 Elo

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

1124.9 · raw 1124.94 Elo

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

35.4% · raw 35.411015132784 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

36% · raw 35.953889773428 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

33.7% · raw 33.726853117672 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

28.6% · raw 28.589614814293 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

22.9% · raw 22.910812181111 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

32.8% · raw 32.828282828283 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

29.3% · raw 29.292929292929 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

21.7% · raw 21.717171717172 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

15.2% · raw 15.151515151515 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

12.6% · raw 12.626262626263 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

55% · raw 54.976851851852 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

51.7% · raw 51.736111111111 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

48.7% · raw 48.726851851852 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

49% · raw 48.958333333333 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

49.2% · raw 49.189814814815 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

44.4% · raw 44.392956441149 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

42.7% · raw 42.678405931418 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

37.3% · raw 37.25671918443 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

33.8% · raw 33.781278962002 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

27% · raw 27.015755329008 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

20.8% · raw 20.8 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

18.2% · raw 18.2 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

17.2% · raw 17.2 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

15.2% · raw 15.2 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

11.2% · raw 11.2 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

18.9% · raw 18.857142857143 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

22.6% · raw 22.571428571429 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

18.6% · raw 18.571428571429 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

12.9% · raw 12.857142857143 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

9.1% · raw 9.142857142857 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

10.7 score · raw 10.666666666666666 score

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

6.1 score · raw 6.05 score

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

5.8 score · raw 5.75 score

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

4.5 score · raw 4.466666666666667 score

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

3.3 score · raw 3.25 score

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Native index/Elo.

82.7% · raw 82.666666666667 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

78.3% · raw 78.333333333333 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

77.3% · raw 77.333333333333 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

77.3% · raw 77.333333333333 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

72% · raw 72 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

36.4% · raw 36.383333333333 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

35% · raw 34.95 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

34.8% · raw 34.783333333333 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

34% · raw 34.033333333333 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

33.3% · raw 33.25 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

40.4% · raw 40.424417081478 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

44.4% · raw 44.427363566487 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

44.5% · raw 44.518272425249 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

44.8% · raw 44.820616472966 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

44.9% · raw 44.943820224719 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

20% · raw 20 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

18.1% · raw 18.095238095238 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

10% · raw 10 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

7.1% · raw 7.142857142857 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

1.9% · raw 1.904761904762 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

89.9% · raw 89.868287740628 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

89.9% · raw 89.897235489941 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

89.5% · raw 89.520914748878 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

89.2% · raw 89.231437255753 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

87.9% · raw 87.85641916341 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

43.9 score · raw 43.899544721027 score

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

40.8 score · raw 40.7857632782172 score

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

37.7 score · raw 37.6782515170274 score

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

35.4 score · raw 35.4471752244645 score

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

28.8 score · raw 28.7822173180696 score

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

40.4 score · raw 40.3835291326876 score

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

39.3 score · raw 39.281621532151 score

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

37.5 score · raw 37.4933078688108 score

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

33.5 score · raw 33.5360066238334 score

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

28 score · raw 27.9750207106805 score

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

42.3 score · raw 42.2752066606861 score

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

39.9 score · raw 39.9006717721577 score

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

37.6 score · raw 37.6058438212847 score

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

35.5 score · raw 35.5422528285353 score

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

30.6 score · raw 30.647798971037 score

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

44.9 score · raw 44.8823073189862 score

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

43.1 score · raw 43.1016145693664 score

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

39.2 score · raw 39.1642979209545 score

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

35.2 score · raw 35.231640545599 score

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

31.2 score · raw 31.2103312647192 score

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

51.2 score · raw 51.2271805877356 score

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

48.7 score · raw 48.704358742663 score

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

45.8 score · raw 45.8402683812172 score

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

43.3 score · raw 43.2746768033673 score

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

37.5 score · raw 37.4612643651529 score

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

52.3% · raw 52.323232323232 %

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

51.4% · raw 51.414141414141 %

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

47% · raw 46.969696969697 %

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

44.9% · raw 44.949494949495 %

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

33.8% · raw 33.838383838384 %

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October. Raw fraction ×100.

1907.6 · raw 1907.63 Elo

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1812.7 · raw 1812.7 Elo

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1702.4 · raw 1702.38 Elo

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1599 · raw 1598.98 Elo

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1225.1 · raw 1225.1 Elo

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1429.1 · raw 1429.1 Elo

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 max configuration; Claude Haiku 5.5 (Max, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1393.4 · raw 1393.42 Elo

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 xhigh configuration; Claude Haiku 5.5 (Xhigh, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1303.5 · raw 1303.45 Elo

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 high configuration; Claude Haiku 5.5 (High, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

1220.8 · raw 1220.84 Elo

Alternative · medium effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 medium configuration; Claude Haiku 5.5 (Medium, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

990.3 · raw 990.32 Elo

Alternative · low effort · Independent evaluator

Source/record date: 2026-10-08

AA primary exact Haiku 5.5 low configuration; Claude Haiku 5.5 (Low, Default Fallback). Non-estimated published values. Default safeguard fallback harness differs from Anthropic no-fallback runs. Evaluation date unpublished; checked 8 October.

22 fixes · raw 22 fixes

Headline · max effort · Independent evaluator

Source/record date: 2026-10-08

Claude Code harness; 1 runs, evaluated 7 October 2026; published raw fixes out of 105, not percentage.

18.3 fixes · raw 18.3 fixes

Alternative · xhigh effort · Independent evaluator

Source/record date: 2026-10-08

Claude Code harness; 3 runs, evaluated 7 October 2026; published raw fixes out of 105, not percentage.

15 fixes · raw 15 fixes

Alternative · high effort · Independent evaluator

Source/record date: 2026-10-08

Claude Code harness; 1 runs, evaluated 7 October 2026; published raw fixes out of 105, not percentage.