OpenAI's top AI still gets a third of routine lab work wrong

OpenAI's new BioBasicsBench hands AI agents 100 routine wet-lab jobs — reading gels, counting colonies, checking controls — built with enzyme-design startup Ligo from real experiments. Its best model, GPT-6 Astra, earned 66.7% of the available credit at the highest reasoning setting, against 43.8% for GPT-5.6 Sol.

OpenAI's own write-up dwells on where Astra still fails: it sorted a colony PCR gel's bands correctly, then picked the wrong colonies anyway, and it overcounted one plate's colonies by about 12%. "Higher benchmark scores do not yet mean models can make reliable judgments in the lab," the company wrote.

The test ran seven OpenAI models across 37 settings, five attempts per task, for 18,500 graded answers — and only OpenAI's own models were tested, so the scoreboard is the company's own.