Summary:
BioBasicsBench measures interpretation of laboratory evidence. Read the seven-model comparison with its uncertainty and practical limits.OpenAI and Ligo introduced BioBasicsBench on November 6: a test of AI agents’ ability to interpret raw wet-lab evidence. Its 100 tasks use expert grading criteria rather than a simple right-or-wrong score.
What the comparison measures
The chart shows all seven tested models at xhigh reasoning effort, with five attempts per task. Higher mean reward is better; it measures the share of available credit, not the percentage of experiments successfully completed. These are developer-reported results, not an independent ranking.
GPT-6 Astra leads this comparison, but uncertainty intervals overlap with GPT-6.1 Sol. The study also documents errors interpreting controls and counting colonies. Better scores therefore do not establish reliable autonomous lab work.
How to use the evidence
For teams considering a lab assistant, a useful next step is a small, supervised evaluation on their own routine data. Record the model version and reasoning setting, keep human checks on consequential conclusions, and examine failure cases alongside averages. A leaderboard cannot substitute for validation in a particular workflow.
This is a benchmark announcement, not a new model release. For another perspective on measuring AI’s research impact, see our coverage of METR’s Anthropic assessment.
Sources and image credit
OpenAI’s November 6 report and results, checked November 7, 2026. Original PromptsZone chart; the official OpenAI provider wordmark is used as a family-logo fallback and belongs to OpenAI, under its brand guidelines. No endorsement implied.
