Measuring how AI agents handle everyday wet-lab science
GPT‑6 Astra earned 66.7% of available credit on routine lab analysis, but still made errors that could derail experiments.
OpenAI developed BioBasicsBench with enzyme-design company Ligo, using 100 tasks drawn from scientists’ own experimental data. At the xhigh reasoning setting, Astra outscored GPT‑5.6 Sol’s 43.8%, showing better recognition of ambiguous signals, faulty controls and microscope depth scans. It still selected incorrect cloning samples, discounted contradictory observations and overcounted colonies on one plate by about 12%. OpenAI says higher benchmark scores do not yet establish reliable judgment in the lab.
OpenAI's note
OpenAI developed BioBasicsBench with enzyme-design company Ligo, using 100 tasks drawn from scientists’ own experimental data. At the xhigh reasoning setting, Astra outscored GPT‑5.6 Sol’s 43.8%, showing better recognition of ambiguous signals, faulty controls and microscope depth scans. It still selected incorrect cloning samples, discounted contradictory observations and overcounted colonies on one plate by about 12%. OpenAI says higher benchmark scores do not yet establish reliable judgment in the lab.
OpenAI's note
score 5