BioBasicsBench Puts Messy Laboratory Evidence at the Center of AI Evaluation

Chinese

OpenAI and Ligo introduced a 100-task wet-lab analysis benchmark on November 6. The published scores measure rubric credit in a developer-run evaluation, rather than successful experiments or independently verified laboratory reliability.

AI-generated concept illustration: a magnifying glass examines abstract specimens beside blank observation cards, representing evidence checking.

OpenAI introduced BioBasicsBench with enzyme-design company Ligo on November 6, 2026, moving the evaluation question toward ordinary laboratory evidence. The benchmark contains 100 wet-lab analysis tasks, using inputs such as gel images, microscopy and instrument outputs. It asks agents to interpret evidence, identify errors and recognize when a conclusion is unsupported. [S1]

In OpenAI’s reported evaluation at xhigh reasoning effort, GPT-6 Astra received 66.7% of available rubric credit, compared with 43.8% for GPT-5.6 Sol. These are average scores with partial credit, not percentages of experiments completed successfully. The release is a benchmark announcement, not a new model launch or a deployed autonomous laboratory service. [S1]

OpenAI describes both improvements and remaining failures, including cases where assumptions override experimental controls. The Independent AI’s November 7 coverage likewise identifies the scores as OpenAI’s figures. AIYXL has not independently reproduced them or conducted laboratory testing. [S1][S2]

For research teams, the useful editorial question is what should happen after an uncertain answer. A polished explanation can still leave the next experimental choice unjustified. An assistant should make it possible for a scientist to trace a conclusion back to the observation, the comparison being made and the evidence that might overturn it.

That suggests a practical acceptance standard for future deployments: retain the original input, record which evidence the assistant used, and distinguish a proposed interpretation from a decision approved by a researcher. A team could then review errors at the point where they affect work, rather than treating a single aggregate score as permission to automate an entire workflow.

Independent reproduction would answer another question: whether the reported gains survive different tools, data preparation and operating conditions. Until that evidence exists, the announcement is best read as a more concrete testing agenda for scientific assistants. The next meaningful milestone is evidence that better analysis produces dependable decisions in the setting where a laboratory actually works.

Sources:

[S1] OpenAI publication, November 6 — Primary evidence for design and developer-reported scores; not independent validation

[S2] The Independent AI, November 7 — Secondary corroboration of announcement; explicitly labels scores as company figures, not a reproduction