BioBasicsBench

BioBasicsBench is OpenAI’s wet-lab agent benchmark: about 100 tasks drawn from Ligo Biosciences enzyme-design experiments, testing whether agents can read gel images and instrument readouts, catch mistakes, and admit when evidence is inconclusive.

Write-up: https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science/

Public notes (X/news + site):

• Aimed at tracking progress on routine science tasks relevant to biosafety;
• Some tasks specifically reward flagging weak data;
• Newer models (e.g. GPT-6 Astra at high reasoning) score well above older GPT-5 series in coverage.

Why it matters

Moves agent evals from pure text exams toward real experimental artifacts, more relevant to lab automation and safety guardrails.

Caveats

• Scores cited in coverage need confirmation against official materials;
• Not a substitute for formal biosafety compliance.

1. Read OpenAI’s write-up for task design;
2. Watch for open tasks/scoring scripts if released;
3. Regression-test your lab agents against the failure modes described;
4. Keep humans in the loop for safety-critical calls.