On November 6, 2026, OpenAI introduced BioBasicsBench, developed with the enzyme design company Ligo, to measure how well AI agents interpret raw experimental data from everyday wet-lab work. Its 100 tasks span five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Inputs are the evidence scientists actually handle, including gel and blot images, plate photographs, microscopy stacks, sequencing reads, plate-reader exports and videos of bench work. Scientists wrote the tasks from their own experiments with expert rubrics, and each went through internal and external PhD-level review.
OpenAI evaluated seven of its own models across 37 model and reasoning-effort settings, five attempts per task, for 18,500 results. At the xhigh setting GPT-6 Astra earned a mean 66.7% of available credit (95% CI 60.3 to 73.1%), GPT-6.1 Sol 62.0%, GPT-5.6 Sol 43.8% and GPT-5.2 31.2%. OpenAI credits better native multimodal handling: in one example Astra noticed a negative control plate was labeled with the wrong antibiotic, which GPT-5.6 Sol missed.
Why it matters: Benchmarks built from figures chosen for papers miss the failed runs, ambiguous images and broken controls of real lab work. A test that rewards recognising when evidence cannot support a conclusion targets the judgment an AI lab assistant would need.
What it does not show: Only OpenAI models were evaluated, and OpenAI says higher scores do not yet mean models can make reliable judgments in the lab; its own examples show Astra letting assumptions override the controls. Whether the tasks are public is not established here.