OpenAI introduced BioBasicsBench with biotech company Ligo on November 6. Its 100 tasks assess interpretation of raw experimental data, error recognition and whether evidence supports a conclusion.

Mean rubric credit, not a pass rate

At xhigh reasoning effort, OpenAI reports GPT-6 Astra earned 66.7% of available rubric credit, versus GPT-5.6 Sol’s 43.8% at xhigh. Each setting included five attempts on all 100 tasks. Reported 95% confidence intervals are 60.3–73.1% and 37.6–50.0%, respectively.

Reliability remains a limitation

OpenAI warns that higher scores do not yet establish reliable laboratory judgments. This is a vendor-run evaluation, not an independently reproduced leaderboard or a new model release. No experiments were performed for this report.