By Hamza Faisal
OpenAI has introduced BioBasicsBench, a new benchmark that tests whether AI agents can make sense of the messy, unpolished evidence scientists actually produce in wet laboratories — DNA gel images, colony plates, microscope photos and instrument readouts — rather than the clean charts that appear in published papers. Built with enzyme-design company Ligo, the benchmark was published on November 6.
The headline result: GPT-6 Astra, OpenAI's newest top-tier model, earned 66.7% of the available rubric credit at the highest reasoning setting tested across all seven models, well ahead of GPT-5.6 Sol's 43.8% and GPT-6.1 Sol's 62.0%. The gap — nearly 23 percentage points over the previous generation's flagship — is the clearest signal yet of how quickly OpenAI's newest models are pulling away on scientific reasoning.
What the benchmark actually measures
BioBasicsBench contains 100 tasks spanning five areas of everyday bench work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Ligo scientists built the tasks from data in their own experiments, pairing each with a detailed scoring rubric written by experts and reviewed by another Ligo scientist and an external wet-lab PhD.
Unlike most science benchmarks, which grade models on polished figures selected for publication, these tasks ask models to interpret raw evidence, spot errors and recognise when the data cannot justify a conclusion — including failed experiments and ambiguous images.
The evaluation ran seven OpenAI models — GPT-5.2, GPT-5.4, GPT-5.5 and GPT-5.6 Sol, plus GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra — across 37 combinations of model and reasoning effort, with five attempts on every task, producing 18,500 scored responses. Scores are mean rubric reward, including partial credit, not the share of tasks fully solved. Astra also did it efficiently, averaging 4,704 output tokens per task against 9,099 for GPT-5.6 Sol.
Where the models still stumble
OpenAI's own worked examples show the gap between a good score and reliable judgement. In a colony-PCR task, Astra sorted the lanes into the right two groups but assumed a larger PCR product meant successful cloning, picking colonies 4, 6, 7 and 8 while the reference answer was 5, 9 and 10. In a colony-counting task it overcounted by about 12%, missing colonies that sat close together and marking empty agar.
GPT-5.6 Sol fared worse on one cloning task, basing its conclusions on an empty negative control plate. Astra caught the error: it read the handwritten "LB Kan50" label on the plate's rim, saw the plate contained kanamycin while the experiment used ampicillin, and rejected it as a valid control.
OpenAI is careful with the framing. "Higher benchmark scores do not yet mean models can make reliable judgments in the lab," the company says, and it notes that tracking these capabilities will also inform its assessment of their biosafety and biosecurity consequences.
Why it matters beyond the leaderboard
Most AI science tests measure whether a model can answer a textbook question or summarise a paper. BioBasicsBench measures something closer to whether it can work in a laboratory: read a gel, count a plate, notice a control that does not add up. If AI agents are going to help run experiments rather than just describe them, that kind of judgement matters more than another point on a quiz.
There are caveats: all seven models tested are OpenAI's own, so the leaderboard says nothing about rivals, and no human baseline or dataset was released with the announcement. Even so, a 23-point jump between generations on genuinely messy data is hard to ignore — the lab bench is becoming AI's next proving ground.



