Examenos

81.2% · raw 81.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Claude Code 2.1.126; parser=json, temperature 1, top_p 1, max_new_tokens 131072; five-run average; Qwen chat template modified. Not Terminus2, another agent harness, or Terminal-Bench4. As reported by Ornith in the 122B card; source table label: TerminalBench 2.1. Exact released 122B checkpoint; no effort magnitude is published.

80% · raw 80 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: CyberGym as labeled by Ornith; revision, agent harness, task population and run count unpublished. Not CyberGym-E2E. As reported by Ornith in the 122B card; source table label: CyberGym. Exact released 122B checkpoint; no effort magnitude is published.

51.2% · raw 51.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Temperature1, top_p 1, 400K context, 48K output; specified GitHub repositories and pip packages blocked to discourage reward hacking. Not MiniMax's corrected task population. As reported by Ornith in the 122B card; source table label: NL2Repo. Exact released 122B checkpoint; no effort magnitude is published.

76.9% · raw 76.9 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: OpenHands; temperature 1, top_p 0.95, 256K context. Models informed of reward-hacking behaviors; any remaining hacking solution is marked incorrect. Pro revision/task count unpublished. Not v 2-Hard or AA. As reported by Ornith in the 122B card; source table label: SWE-bench Pro (no-hack). Exact released 122B checkpoint; no effort magnitude is published.

84.5% · raw 84.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: OpenHands; temperature 1, top_p 0.95, 256K context. Models informed of reward-hacking behaviors; any remaining hacking solution is marked incorrect. Not minimal-bash or an independent evaluator run. As reported by Ornith in the 122B card; source table label: SWE-bench Verified (non-hack). Exact released 122B checkpoint; no effort magnitude is published.

76.4% · raw 76.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: OpenHands; temperature 1, top_p 0.95, 256K context. Models informed of reward-hacking behaviors; any remaining hacking solution is marked incorrect. Exact suite revision/task population unpublished. As reported by Ornith in the 122B card; source table label: SWE-bench Multilingual (non-hack). Exact released 122B checkpoint; no effort magnitude is published.

48% · raw 48 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Tool-enabled HLE with Claude 4.6 Opus judge; exact task population, tool budget and effort unpublished. Not a no-tools result or independent evaluator run. As reported by Ornith in the 122B card; source table label: HLE (with tools). Exact released 122B checkpoint; no effort magnitude is published.

37.7% · raw 37.68 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: No-tools HLE with Claude 4.6 Opus judge; exact text/full task population and effort unpublished. Table precision preserved rather than replaced with figure truncations. As reported by Ornith in the 122B card; source table label: HLE (w/o tools). Exact released 122B checkpoint; no effort magnitude is published.

53.8% · raw 53.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: SciCode as labeled; whole-problem versus subtask aggregation, background condition, run count and harness unpublished. Do not assume the AA component convention. As reported by Ornith in the 122B card; source table label: SciCode. Exact released 122B checkpoint; no effort magnitude is published.

75.2% · raw 75.2 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Vendor-reported WideSearch; exact source revision, population, task budget and effort unpublished. As reported by Ornith in the 122B card; source table label: WideSearch. Exact released 122B checkpoint; no effort magnitude is published.

78.5% · raw 78.5 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Vendor-reported BrowseComp; exact source revision, browsing harness, task budget and effort unpublished. Not BrowseComp+ or another offline protocol. As reported by Ornith in the 122B card; source table label: BrowseComp. Exact released 122B checkpoint; no effort magnitude is published.

74.6% · raw 74.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: 500-task public subset; all models in thinking mode; 10-minute timeout/task; Claude 4.8 Opus judge. Distinct from Scale's Gemini2.5-Pro judge and full1, 000-task results. As reported by Ornith in the 122B card; source table label: MCP-Atlas. Exact released 122B checkpoint; no effort magnitude is published.

70.8% · raw 70.8 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Vendor-reported official evaluation service with 128K maximum tokens. Not an independent score merely because the vendor used that service. As reported by Ornith in the 122B card; source table label: Toolathlon-Verified. Exact released 122B checkpoint; no effort magnitude is published.

79.4% · raw 79.4 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Agentic coding over real-user task distributions; temperature 0.6, 256K context; dataset revision and run dates unpublished. As reported by Ornith in the 122B card; source table label: ClawEval. Exact released 122B checkpoint; no effort magnitude is published.

37.6% · raw 37.6 %

Headline · unknown effort · Own vendor

Source/record date: 2026-10-21

Harness: Unspecified HLE figure condition. Manually read from official pixels; source labels only HLE. Do not infer a rounding/truncation rule or transfer this value into the exact no-tools table. Beam36 is present only in the figure. As reported by Ornith.