OpenAI introduced BioBasicsBench on 6 November 2026, a benchmark developed with enzyme-design company Ligo to test AI agents on 100 routine wet-lab tasks. Its five areas span protein work, visual and cell-based analysis, measurement and testing, work with DNA or RNA, and the assembly of genetic constructs. Agents work from experimental material including gel images, colony plates, microscopy, instrument readouts and sequencing files.
Scientists built the tasks from their own experiments and wrote criteria for grading each answer. Each task received review from a Ligo scientist who had not written it, a PhD-level scientist working in a wet lab, and a final internal reviewer. Credit can depend on a conclusion, a measurement, the reasoning behind it or recognition that the evidence is insufficient.
Checking whether an experiment worked could therefore require more than reading a result: an ambiguous image or a problem with the controls would change what conclusion the evidence could support.
OpenAI tested seven models across 37 combinations of model and reasoning effort, making five attempts at each task per setting. At the highest reasoning-effort setting reported for all seven, GPT-6 Astra earned 66.7% of the available credit on average, against 43.8% for GPT-5.6 Sol. In one task, Astra identified that a detector could register light from two dyes, so brightness alone could not establish whether cells were dead.
