BioMysteryBench
Whether an agent can recover information that real biological data has had stripped from it — sequencing, expression analysis, variant calling, epigenomics, metagenomics, proteomics and metabolomics — by computation rather than recall.
How it is run
Anthropic builds each task by removing the metadata (sample origin and the like) from a real dataset and asking the agent to work it back out with command-line tools, analysis code and database queries. Up to five domain experts attempted every problem, which splits the board into human-solvable and human-difficult sets; Vals runs the agents and scores both.
What it cannot tell you
Biology content filters refuse some tasks and those refusals are scored as failures — Sonnet 5.5 refused the same three tasks in each of its three runs. The split matters: Opus 5.5 still leads the human-difficult set (47.1%) even though Sonnet 5.5 leads overall, and most failures are confident wrong answers rather than flagged uncertainty.
- Leader
- Claude Sonnet 5.581.11% · Vals AI's own list
- Top-ten spread
- 10.00%Leader minus tenth place, or last place in smaller fields.
- Median
- 69.26%of the 21 rows published
- No published result
- 27of 46 models on this site
Independent leaderboard · Vals AI
Top 21 of 21 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | Claude Sonnet 5.5 | 81.11% |
| 2 | GPT-6.1 Sol | 79.63% |
| 3 | GPT-6 Astra | 79.26% |
| 3 | Claude Opus 5.5 | 79.26% |
| 3 | Claude Opus 5 | 79.26% |
| 6 | Gemini 4 Argon | 76.30% |
| 7 | GPT-6 Sol | 74.81% |
| 8 | Grok 4.6 | 72.22% |
| 9 | Kimi K3 | 71.48% |
| 10 | GPT-5.6 Sol | 71.11% |
| 11 | Grok 4.7 | 69.26% |
| 11 | Hunyuan Hy4 Preview | 69.26% |
| 11 | MiMo-V2.6-Flash | 69.26% |
| 14 | DeepSeek V4.1 Flash | 67.78% |
| 15 | Muse Spark 1.2 | 64.81% |
| 16 | Deepseek V4 Flash 0731 | 64.44% |
| 17 | Gemini 3.8 Flash | 62.22% |
| 18 | GPT-6 Luna | 61.48% |
| 18 | GPT-5.6 Luna | 61.48% |
| 20 | Gemini 3.1 Pro | 60.00% |
| 21 | Gemini 3.6 Flash | 58.52% |
No published result tracked
27 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.
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/biomysterybench
- Independent resultswww.vals.ai/benchmarks/biomysterybench