Four-Factor Integrity Breakdown & Primary-Source Evidence
Access Model: Auto-Gated / Encrypted
"BioMysteryBench is provided for evaluation and benchmarking only. By requesting access you agree that you will not use this benchmark (problem statements, answer rubrics, task formulation) to train, fine-tune, reinforce, or distill any machine-learning model."
"BioMysteryBench was developed by Anthropic. Its task materials are released under CC BY 4.0 with restrictions against their use for model training, while the biological data remain subject to their original repositories’ policies."
Originality
Weight 25%
7.2/10
Reuse Profile: Low Reuse (<25%)
"Tasks are constructed by stripping real-world biological data of metadata (e.g. sample origin) and evaluating whether models can recover this information via computational analysis."
Freshness
Weight 20%
7/10
Initial Release: 2026-04Latest Refresh: 2026-07Rolling Updates: No
"Removed after a June 2026 audit that reran every problem with expert bioinformaticians and multiple model runs, and cross-checked answer keys against the underlying data."
Grading
Weight 15%
3.3/10
Integrity: Self-Reported / Live NetworkSandbox: live internet
"Network access is restricted to the scientific package repositories and biological databases required by the benchmark. Tasks were limited to 200 agent turns and four hours of wall-clock execution."
Frontier Model Standings on BioMysteryBench
Top 15 active models tracked
Benchmark maintainers can embed this live SVG badge in GitHub READMEs or documentation:
[](https://benchaudit.org/benchmarks/biomysterybench/)