WitnessBench: leaderboard

An agent plays 26 unseen line-puzzle games of 20 levels each, 3 seeds per game, without being told the rules and must infer them from the game's responses; the score rewards solving levels in close to the fewest actions, from Fleet's WitnessBench v2.0 runs. It tests rule discovery by experiment, unlike puzzle boards that state the rules.

Metric: OAE-L20 score on 26 games x 3 seeds (0-100). Source: fireworks.ai. Saturation forecast: Around October 2027. 13 models tracked. Updated 16 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 5.5Max46.5#5
2GPT-6 Astra ProMax25.6#3
3GPT-6.1 Pro SolMax21.7#7
4Kimi K3Max18.8#56
5Gemini 4 ArgonHigh17.9#12
6GLM-5.3Max16.6#68
7Grok 4.7xHigh14.8#47
8Gemini 3.8 FlashHigh6.7#36
9DeepSeek V4 Pro (0813)Max5.5#90
10Qwen3.8 2.4T A95BxHigh5#83
11Qwen 3.8 27BxHigh3.1#209
12Muse Glimmer 30BxHigh0.21#391
13Qwen 3.5 9BMax0.11#814

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #6 GPT-6 Astra.

Interactive version: theaggregate.ai/benchmark?slug=witnessbench · How It Works · Data refreshed daily, snapshot 2026-11-16.