OpenAI said on 6 November that GPT-6 Astra earned an average of 66.7% of the available credit on BioBasicsBench, a new benchmark of 100 routine wet-lab tasks it built with Ligo, a biotech company focused on enzyme design. That is 2.1 times the 31.2% scored by GPT-5.2, the lowest of the seven OpenAI models tested at the same setting, by our arithmetic.
OpenAI says the benchmark tests “whether agents can interpret raw experimental data, identify errors, and recognize when the evidence cannot support a conclusion”. It reports results only for its own models and gives no human baseline.
Business Pill · AI Agents Explained
A short explainer on AI agents. It teaches the general idea only and does not discuss this benchmark.
The key insight: As we read it, GPT-6 Astra scored 2.1 times GPT-5.2 on OpenAI’s own wet-lab benchmark while using about 83% fewer output tokens per task, by our arithmetic; OpenAI reports no human baseline and no other labs’ models.
According to OpenAI’s post, the 100 tasks span five areas of routine wet-lab work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. The inputs are raw evidence such as gel and blot images, plate photographs, microscopy stacks, sequencing reads, plate-reader exports and videos of bench work.
OpenAI says scientists wrote the tasks from their own experiments, each with an expert-written grading rubric, and that each task was reviewed internally by a Ligo scientist who had not written it, externally by a PhD-level scientist working in a wet lab, and in a final internal review. Answers earn partial credit against individual rubric criteria.
Mean reward on BioBasicsBench at the ‘xhigh’ reasoning-effort setting, as listed in OpenAI’s post of 6 November 2026: 100 tasks, five attempts each, percentage of available credit. OpenAI reports results only for its own models and no human baseline. The 2.1 times ratio is our arithmetic.
The Scores
OpenAI evaluated seven models, GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 Sol, GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra, across 37 model and reasoning-effort settings, with five attempts on each of the 100 tasks per setting, for 18,500 results. At the ‘xhigh’ setting, the highest effort reported for all seven, the post lists mean reward of 66.7% for GPT-6 Astra, 62.0% for GPT-6.1 Sol, 43.8% for GPT-5.6 Sol, 38.3% for GPT-6 Luna, 35.4% for GPT-5.5, 33.2% for GPT-5.4 and 31.2% for GPT-5.2.
The same table lists average output tokens per task: 4,704 for GPT-6 Astra and 27,620 for GPT-5.2. Astra used about 83% fewer tokens while scoring 35.5 points higher, by our arithmetic. Against GPT-5.6 Sol, Astra’s gain is 22.9 points. OpenAI writes that the gains “suggest that stronger native multimodal capabilities may be key to improving performance on everyday wet-lab” analysis.
Where OpenAI Says the Models Still Fail
The post is explicit about limits: “Higher benchmark scores do not yet mean models can make reliable judgments in the lab.” It gives three Astra failures. In one, Astra “lets a familiar assumption override the interpretation supported by the experimental controls”, picking colony lanes 4, 6, 7 and 8 where the reference selects lanes 5, 9 and 10.
In another, OpenAI says Astra “discounts observations that contradict its preferred interpretation” when describing cell vacuoles. In a third, its colony counting produced “an overcount of approximately 12% relative to the benchmark reference” on one plate.
On safety, the post says only that “Tracking these capabilities will also help us understand their implications for biosafety and biosecurity”. It does not describe specific safeguards for the benchmark.
Two Labs, One Day
Anthropic published a related post the same day, a guest essay by its AI for Science grantees at Tübingen on AI systems that build scientific models. That essay cautions that such systems “have mostly been tested on rediscovering models that scientists already know”. We added it as an update to our piece on Anthropic’s 207,049 science sessions.
As we read the post, the test is built around messy everyday evidence, from gel photos to bench videos, rather than figures chosen for papers.
As we read the table, GPT-6 Astra and GPT-6.1 Sol, at 66.7% and 62.0%, sit well above GPT-5.6 Sol at 43.8%, while GPT-6 Luna scores 38.3%, so the newer generation is not uniformly ahead.
As we read the failure examples, they mix judgement errors, such as letting an assumption override the controls, with image-reading errors, such as missed colonies and empty agar marked as a colony.
OpenAI, in its BioBasicsBench post
“Higher benchmark scores do not yet mean models can make reliable judgments in the lab.”
Three Implications
FOR LAB TEAMSOpenAI lists three Astra failure cases, including an overcount of approximately 12% on one colony plate.
FOR BENCHMARKSOpenAI says the 100 tasks were written by scientists from their own experiments and reviewed three times, once externally.
FOR COMPARISONS The post reports only OpenAI models; no other lab or human baseline is scored.
The analysis says: “Model capability is exponential; product capability is approximately linear.”
As we read it, a benchmark of routine lab analysis measures one step between a model score and a lab that can rely on it. The analysis does not discuss this benchmark.
What Is Not Established
The benchmark is OpenAI’s own, built with one partner, and the post reports scores only for OpenAI models; it gives no results for other labs’ models and no human baseline. We have not seen the task set or rubrics, which the post does not say are public. The 2.1 times ratio, the 83% token figure and the point gaps are our arithmetic from OpenAI’s table.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
OpenAI’s new wet-lab benchmark puts GPT-6 Astra at 66.7% of available credit, 2.1 times GPT-5.2, while the post itself says higher scores do not yet mean reliable lab judgment and lists three failure cases.
A note on sourcing. The benchmark design, scores and failure examples are from OpenAI’s post of 6 November 2026; the benchmark is OpenAI’s own and reports only OpenAI models. The Anthropic quote is from its guest essay of the same day. Ratios and point gaps are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
