MysteryMechanism
Whether a model can infer an unknown mechanism from evidence, in the style of a research problem rather than an exam question.
Scientific reasoningVals AI21 models run by Vals AIUpdated 29 September 2026
How it is run
A small board — nineteen models — so rank positions move a lot between releases.
What it cannot tell you
Small sample sizes mean single-digit differences are not meaningful.
- Leader
- GPT-6 Astra53.15% · Vals AI's own list
- Top-ten spread
- 19.82%Leader minus tenth place, or last place in smaller fields.
- Median
- 30.63%of the 21 rows published
- No published result
- 25of 46 models on this site
Independent leaderboard · Vals AI
Top 21 of 21 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | GPT-6 Astra | 53.15% |
| 2 | Claude Opus 5.5 | 49.55% |
| 3 | Claude Sonnet 5.5 | 49.10% |
| 4 | Claude Fable 5.1 | 47.75% |
| 5 | GPT-6.1 Sol | 46.40% |
| 6 | Gemini 4 Argon | 45.49% |
| 7 | Claude Opus 5 | 37.39% |
| 8 | Gemini 3.8 Flash | 36.49% |
| 9 | Muse Spark 1.3 Max | 36.04% |
| 10 | GPT-5.6 Sol | 33.33% |
| 11 | Grok 4.6 | 30.63% |
| 12 | GPT-6 Sol | 30.18% |
| 13 | Grok 4.7 | 25.23% |
| 14 | Qwen3.8 Max | 23.87% |
| 15 | GLM 5.3 | 22.97% |
| 16 | MiMo-V2.6-Flash | 21.62% |
| 17 | DeepSeek V4.1 Flash | 21.17% |
| 18 | Gemini 3.1 Pro | 20.72% |
| 19 | GPT-6 Luna | 19.37% |
| 20 | MiMo V2.6 Pro | 15.31% |
| 21 | GPT-5.6 Luna | 14.41% |
No published result tracked
25 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Step 5 PreviewGPT-5.6 TerraGLM 5.3 FlashQwen3.8 2.4T A95BQwen3.8-Flash-NextQwen3.8 Max (0902)Claude Fable 5Claude Opus 4.8Claude Opus 4.7Gemini 3.7 FlashMuse Spark 1.2GPT-5.4Grok 4.5GPT-5.5DeepSeek V4.1 Flash 1017GLM-LegalClaude Haiku 5.5BeamClaude Sonnet 5DeepSeek V4 ProKimi K3MiniMax M3Hunyuan Hy4 PreviewInklingMistral Medium 3.5
Read this with
Benchmarks in the same category, which usually disagree with this one on purpose.
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes itwww.vals.ai/benchmarks/mysterymechanism
- Independent resultswww.vals.ai/benchmarks/mysterymechanism