KWBench: leaderboard

Unprompted problem recognition on KWBench (Knowledge Work Bench, 2026): 223 practitioner-sourced knowledge-work tasks in which the model must recognize, unprompted, the game-theoretic structure of a professional situation from raw inputs; a Gemini 3 Flash judge checks rubric criteria, best of three runs per model. Share of tasks on which every mandatory criterion is met, in percent; higher is better.

Metric: Mandatory-gate pass rate (%): share of evaluated tasks on which every mandatory criterion is met, over KWBench's 223 practitioner-sourced knowledge-work tasks (acquisitions, negotiations, clinical pharmacy, organizational politics, fraud analysis, incentive design), each encoding a game-theoretic pattern the model must recognize unprompted from raw inputs, with a code interpreter; a Gemini 3 Flash judge checks 15 binary rubric criteria per task, any failed mandatory criterion scores the task zero; best of three runs per model at provider-recommended temperature; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.627.9#113
2GPT-5.421.1#135
3GLM-5-Turbo20.4#255
4Gemini 3.1 Pro (Preview)15.7#99
5Qwen 3.5 Plus14.9#242
6Kimi K2.514.3#266
7KWBench Qwen 3.5 MoE (checkpoint unspecified)14
8MiniMax-M2.713.9#539
9Gemini 3 Flash13.5#175
10MiniMax-M2.511.8#600
11GPT-OSS-120B10.8#704

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=kwbench · How It Works · Data refreshed daily, snapshot 2026-11-21.