OpenAI introduces BioBasicsBench for routine wet-lab analysis

Its 100 tasks test whether agents can interpret raw experimental data, spot errors and recognize when the evidence cannot support a conclusion.[1] The benchmark spans protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning.[1] Its inputs include gel images, colony plates, microscopy and instrument readouts.[1] This gives teams evaluating agents a concrete set of lab-analysis capabilities to consider.

Scientists created the tasks using data from their own experiments, and each task is paired with expert-written grading criteria.[1] Each task received an internal review from a Ligo scientist who did not write it, an external review from a wet-lab scientist with a PhD, and a final internal review.[1] Agents return structured answers that may include a requested conclusion, measurement or recommendation, with supporting reasoning when required.[1] The benchmark awards credit against individual rubric criteria, so it can distinguish a correct conclusion from a partially correct analysis.[1] For teams, that scoring detail is useful when assessing more than a single final answer.

At the xhigh reasoning setting, GPT-6 Astra earned an average of 66.7% of available credit; GPT-5.6 Sol earned 43.8%.[1] OpenAI characterizes the results as evidence of progress in routine lab analysis, while identifying image interpretation, control checking and the use of experimental context as areas needing improvement.[1] CellCog separately reports that GPT-6 Astra used 4,704 output tokens per task, compared with 9,099 for GPT-5.6 Sol.[3] Teams comparing these figures should keep the sources and measures distinct.

OpenAI says routine wet-lab work can involve failed experiments, ambiguous images and conflicting observations.[1] It also says benchmarks based on charts and images selected for scientific papers may miss ambiguity and the troubleshooting that it requires.[1] CellCog cautions that higher benchmark scores do not yet mean models can make reliable judgments in the lab.[3] Its reported examples include GPT-6 Astra catching a control plate labeled for the wrong antibiotic, but also picking the wrong colonies from a cloning gel and overcounting a colony plate by about 12%.[3] These results counsel against treating an overall score as proof of dependable performance on every task.