MysteryMechanism
Working out a hidden scientific mechanism from sparse observations. Run by Vals AI.
As of 19 Sept 2026, GPT-6 Astra leads MysteryMechanism on BenchLeader with 53.1%, ahead of Claude Fable 5.1 at 47.8%, across 11 model configurations with a published result.
- Published by
- Vals AI
- Category
- Reasoning
- Index weight
- Reference only
- Models
- 11
- Data as of
- 19 Sept 2026
Vals AI (vals.ai).
What the test looks like
The model can request observations of an unknown system and must infer the mechanism behind them, a test of active scientific reasoning.
How it is scored
Percent of mechanisms identified, run by Vals AI.
What to keep in mind
Small and unusual; a good complement to static exams.
- 1GPT-6 Astra (max)53.1%
- 2Claude Fable 5.147.8%
- 3Claude Opus 537.4%
- 4Gemini 3.8 Flash (high)36.5%
- 5Muse Spark 1.3 (max)36.0%
- 6GPT-5.6 Sol (max)33.3%
- 7Grok 4.6 (high)30.6%
- 8Qwen3 8 (max)23.9%
- 9GLM-5.3 (max)23.0%
- 10DeepSeek V4.1 Flash (high)21.2%
- 11GPT-5.6 Luna (max)14.4%
11 of 11
| # | ||||
|---|---|---|---|---|
| 1 |  GPT-6 AstramaxOpenAI | 53.1% | 71.8 | 2026-09-17 |
| 2 |  Claude Fable 5.1Anthropic | 47.8% | 64.5 | 2026-09-17 |
| 3 | Claude Opus 5Anthropic | 37.4% | 66.7 | 2026-09-17 |
| 4 |  Gemini 3.8 FlashhighGoogle | 36.5% | 64.5 | 2026-09-17 |
| 5 |  Muse Spark 1.3maxMeta | 36.0% | 69.3 | 2026-09-17 |
| 6 | GPT-5.6 SolmaxOpenAI | 33.3% | 68.8 | 2026-09-17 |
| 7 |  Grok 4.6highxAI | 30.6% | 64.5 | 2026-09-17 |
| 8 |  Qwen3 8maxAlibaba | 23.9% | 66.5 | 2026-09-17 |
| 9 |  GLM-5.3maxZhipu AIopen ↗ | 23.0% | 65.8 | 2026-09-17 |
| 10 |  DeepSeek V4.1 FlashhighDeepSeekopen ↗ | 21.2% | – | 2026-09-17 |
| 11 | GPT-5.6 LunamaxOpenAI | 14.4% | 60.2 | 2026-09-17 |
Cite as: BenchLeader, “MysteryMechanism leaderboard”, https://www.benchleader.com/benchmarks/vals_mystery_mechanism, data as of 19 Sept 2026.
MysteryMechanism: questions
- What does MysteryMechanism measure?
- The model can request observations of an unknown system and must infer the mechanism behind them, a test of active scientific reasoning. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads MysteryMechanism?
- GPT-6 Astra leads MysteryMechanism with 53.1% as of 19 Sept 2026, ahead of Claude Fable 5.1 at 47.8%.
- How many models have MysteryMechanism results?
- 11 model configurations have a MysteryMechanism result on BenchLeader, all taken from Vals AI.
- Who runs MysteryMechanism and how often is it updated?
- MysteryMechanism is published by Vals AI. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does MysteryMechanism count toward the BenchLeader Index?
- No. MysteryMechanism is shown for reference but left out of the composite index.