# BioMysteryBench Contamination Audit

> BioMysteryBench scores 4.5/10 (Tier C, High Risk) on BenchAudit as of November 2026, reflecting auto-approved click-through or encrypted public access and self-reported or live-network evaluation.

- **Canonical URL:** https://benchaudit.org/benchmarks/biomysterybench/
- **Benchmark ID:** `biomysterybench`
- **Hub:** Vals AI
- **Category:** Science
- **Domains:** Science
- **Skills:** Agentic, Reasoning, Tool Use, Coding
- **BenchAudit Score:** 4.5/10 (Tier C - High Risk)
- **Last Audited:** 2026-11-22 (Ref 2026-11)

## Factor Scores

| Factor | Score (/10) | Weight |
| --- | ---: | ---: |
| Privacy | 2 | 40% |
| Originality | 7.2 | 25% |
| Freshness | 7 | 20% |
| Grading | 3.3 | 15% |

## Primary-Source Evidence

### Privacy
> "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."
- Source: https://huggingface.co/datasets/Anthropic/BioMysteryBench-full

### Originality
> "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."
- Source: https://www.vals.ai/benchmarks/biomysterybench

### Freshness
> "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."
- Source: https://huggingface.co/datasets/Anthropic/BioMysteryBench-preview/blob/main/CHANGELOG.md

### Grading
> "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."
- Source: https://www.vals.ai/benchmarks/biomysterybench

## Top Frontier Models on BioMysteryBench

| Rank | Model | Creator | Percentile |
| ---: | --- | --- | ---: |
| #1 | [Claude Fable 5.2](https://benchaudit.org/models/claude-fable-5-2__max/) | Anthropic | 100.0p |
| #2 | [Claude Sonnet 5.5](https://benchaudit.org/models/claude-sonnet-5-5__max/) | Anthropic | 100.0p |
| #3 | [GPT-6.2 Astra](https://benchaudit.org/models/gpt-6-2-astra__max/) | OpenAI | 96.6p |
| #4 | [GPT-6.1 Sol](https://benchaudit.org/models/gpt-6-1-sol__max/) | OpenAI | 92.6p |
| #5 | [Claude Opus 5.5](https://benchaudit.org/models/claude-opus-5-5__max/) | Anthropic | 86.6p |
| #5 | [GPT-6 Astra](https://benchaudit.org/models/gpt-6-astra__max/) | OpenAI | 86.6p |
| #7 | [GPT-6.2 Sol](https://benchaudit.org/models/gpt-6-2-sol__max/) | OpenAI | 80.6p |
| #8 | [Gemini 4 Argon](https://benchaudit.org/models/gemini-4-argon__high/) | Google | 76.7p |
| #9 | [Kimi K3.1 Pro](https://benchaudit.org/models/kimi-k3-1-pro__max/) | Kimi | 72.6p |
| #10 | [GPT-6 Sol](https://benchaudit.org/models/gpt-6-sol__max/) | OpenAI | 68.7p |
