OpenAI
Measuring routine laboratory work
Developed by OpenAI and enzyme design company Ligo, BioBasicsBench evaluates AI agents on 100 routine laboratory tasks.
The tasks include raw evidence encountered in real experiments, from gel and microscope images to instrument outputs, protocols, sequence files and laboratory videos.
Testing the evidence as well as the answer
The evaluation measures not only whether agents reach the right conclusion, but whether they detect problems with experimental controls, recognize ambiguous data and provide supporting reasoning.
In results shared by OpenAI, GPT‑6 Astra scored an average of 66.7% at the xhigh reasoning setting. The work shows room for improvement, particularly in interpreting images and applying experimental context.
Anthropic