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
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Max | 46.5 | #5 |
| 2 | GPT-6 Astra Pro | Max | 25.6 | #3 |
| 3 | GPT-6.1 Pro Sol | Max | 21.7 | #7 |
| 4 | Kimi K3 | Max | 18.8 | #56 |
| 5 | Gemini 4 Argon | High | 17.9 | #12 |
| 6 | GLM-5.3 | Max | 16.6 | #68 |
| 7 | Grok 4.7 | xHigh | 14.8 | #47 |
| 8 | Gemini 3.8 Flash | High | 6.7 | #36 |
| 9 | DeepSeek V4 Pro (0813) | Max | 5.5 | #90 |
| 10 | Qwen3.8 2.4T A95B | xHigh | 5 | #83 |
| 11 | Qwen 3.8 27B | xHigh | 3.1 | #209 |
| 12 | Muse Glimmer 30B | xHigh | 0.21 | #391 |
| 13 | Qwen 3.5 9B | Max | 0.11 | #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.