OpenAI has launched BioBasicsBench, a 100-task benchmark designed to evaluate how AI agents perform routine wet-lab science. Developed in collaboration with biotech firm Ligo, the benchmark tests agents on interpreting raw experimental data, identifying errors, and recognizing when evidence is insufficient. Current models show progress but still struggle with image interpretation and control checks.
At the 'xhigh' reasoning setting, GPT-6 Astra scored 66.7%, outperforming GPT-5.6 Sol at 43.8%. The tasks span protein methods, imaging, quantification, nucleic acid methods, and molecular cloning, using real-world data like gel images, microscopy, and sequencing reads. Each task is reviewed by experts and graded against detailed rubrics to assess both conclusions and reasoning.
Source: OpenAI Blog · Summarized by HeadlinesBriefing