OpenAI introduces BioBasicsBench: 100 routine wet-lab tasks, GPT-6 Astra scores 66.7%
On November 6, OpenAI and biotech firm Ligo introduced BioBasicsBench, 100 routine wet-lab analysis tasks built from raw evidence such as gel images, agar plates and microscopy. At the highest reasoning setting GPT-6 Astra earned 66.7% of available credit versus 43.8% for GPT-5.6 Sol, and OpenAI openly lists failures like miscounting colonies and letting assumptions override controls. My take: the progress is real, but missing a third of the credit means AI shouldn't be signing off on conclusions yet. Same goes for businesses using AI on reports or inspection data: let it flag what looks off, and keep a human on the final call.