OpenAI's BioBasicsBench finds GPT-6 Astra earns 66.7% of credit on routine wet-lab analysis tasks

OpenAI · Research report

Read the source (opens in a new tab)
What happened
OpenAI released BioBasicsBench on November 6, 2026, a 100-task benchmark developed with biotech company Ligo that tests whether AI agents can interpret raw data from routine wet-lab experiments.
Key points
  • At "xhigh" reasoning effort, GPT-6 Astra earned 66.7% of available credit, against 62.0% for GPT-6.1 Sol and 43.8% for GPT-5.6 Sol.
  • Tasks use scientists' own gels, colony plates, microscopy and instrument readouts, and test reading data, catching errors and recognizing when evidence cannot support a conclusion.
  • OpenAI says "Higher benchmark scores do not yet mean models can make reliable judgments in the lab," showing cases where expected answers overrode the evidence.
  • OpenAI says "Tracking these capabilities will also help us understand their implications for biosafety and biosecurity."
Evidence
OpenAI's own benchmark, run on seven of its own models, with tasks reviewed by Ligo and outside scientists. Its biosafety content is that one sentence; it reports no risk finding or safeguard.
Uncertain
Scores carry 95% confidence intervals of about six points either way. OpenAI tested no other developers' models and reports no human baseline.

See it in the latest record