OpenAI benchmark finds top model reads raw lab evidence 67% correctly

The latest:

AI agents correctly interpreted raw wet-lab evidence in two of three cases at best, according to a benchmark OpenAI published on November 6, 2026. BioBasicsBench, built with enzyme-design company Ligo, puts 100 tasks to models using gel images, colony plates, microscope stacks and plate-reader files. OpenAI said the highest scores still do not amount to trustworthy laboratory judgment.

Details:

• The scores: At the xhigh reasoning level, the highest reported across seven models, GPT-6 Astra scored 66.7%, with a 95% confidence interval of 60.3–73.1%, according to OpenAI. GPT-5.6 Sol scored 43.8%, interval 37.6–50.0%. Scores are mean reward: average task score against available credit.
• Compute gap: Astra averaged 4,704 output tokens per task against Sol’s 9,099, per OpenAI’s figures — the stronger model scored higher on roughly half the output. The published comparison covers OpenAI models only, with no external systems included.
• Scale of testing: The evaluation spanned 37 model-and-reasoning-level configurations, with 5 attempts per task per configuration, producing 18,500 results. Confidence intervals were computed by resampling whole tasks while keeping each set of five attempts together, OpenAI said.
• The rationale: OpenAI argued that figures lifted from published papers miss the mess of daily bench work, which it said “also involves failed experiments, ambiguous images, conflicting observations and troubleshooting.” The benchmark is built to test evidence that is incomplete or contradictory.
• Task design: Tasks cover protein methods, imaging and flow cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Inputs include gel and blot images, colony plates, microscope z-stacks, sequencing reads, bench photos, protocol videos, and references such as ladder diagrams and plasmid maps.
• Grading and build: Agents return structured conclusions, measurements or recommendations, with credit for numerical results, the reasoning behind them, and for recognizing insufficient evidence. Scientists wrote tasks from their own experiments; each passed internal Ligo review, external review by a PhD-level wet-lab scientist, then final internal review.
• Where Astra gained: In cell imaging it flagged that overlapping detection windows could allow a second dye to bleed through and asked for separated controls, while Sol called cells dead from the dye signal alone. In cloning, Astra rejected a negative control plate labeled kanamycin rather than the ampicillin in use.
• Documented failures: In colony PCR, Astra picked lanes 4, 6, 7 and 8 on the assumption that the larger product meant successful cloning; the reference answer was 5, 9 and 10. It also over-counted colonies by about 12%, merging adjacent ones, missing visible ones and marking empty agar.
• Safety framing: OpenAI said it will use the trajectory of these capabilities to inform its assessment of biosafety and biosecurity consequences. The announcement did not state whether the benchmark will be released open source.
• Also: Labs weighing AI for routine analysis now have a measure of where agents misread gels and colonies.

Between the lines:

The failure cases cluster around counting and reading physical images rather than scientific reasoning: miscounting colonies by 12%, misreading gel lanes. The gains, by contrast, came from catching procedural errors — a mislabeled control plate, a dye bleed-through risk. That split suggests perception, not inference, is the current bottleneck.

What’s next

Watch whether OpenAI publishes scores for non-OpenAI models, releases the 100 tasks openly, and what its stated biosafety and biosecurity assessment concludes once capability trends are folded in.