OpenAI Releases BioBasicsBench, Says GPT-6 Astra Scores 66.7%
Summary
OpenAI releases BioBasicsBench, a 100-task benchmark built with biotech firm Ligo that tests whether AI agents can read raw wet-lab data, check controls and flag unsupported conclusions. OpenAI reports GPT-6 Astra scores 66.7% of available credit at xhigh reasoning effort, against 43.8% for GPT-5.6 Sol. Astra still miscounts colonies, overcounting about 12%.
Key Points
- OpenAI says the BioBasicsBench evaluation covers seven models across 37 settings, with five attempts on each of the 100 tasks.
- Each task goes through three reviews, including an external check by a PhD-level scientist working in a wet lab.
- Astra rejects a negative control plate labeled LB Kan50, kanamycin, because the experiment uses ampicillin, OpenAI reports.