Artificial intelligence / news brief
OpenAI introduces BioBasicsBench for interpreting raw wet-lab evidence
- Event
- Published

OpenAI introduced BioBasicsBench, a 100-task evaluation built with Ligo around raw wet-lab images, videos and instrument files. OpenAI reports GPT-6 Astra earned 66.7% mean partial credit at xhigh effort across five attempts per task; the tasks, rubrics, code and raw results were not publicly linked, and the scores have not been independently reproduced.
Why it matters
The benchmark targets error detection and unsupported conclusions in laboratory evidence, where a plausible answer can still lead to a costly experiment.
Sources
- Measuring how AI agents handle everyday wet-lab science (external site)OpenAI · primary source ·
- BioBasicsBench: OpenAI’s Wet-Lab AI Results Explained (external site)Kingy AI · coverage ·