OpenAI's BioBasicsBench Tests AI on Real Wet-Lab Work — and GPT-6 Astra Still Gets It Wrong

What OpenAI Published on 6 November

OpenAI published BioBasicsBench on 6 November 2026, a benchmark of 100 tasks built with enzyme-design company Ligo that tests whether AI agents can handle everyday wet-lab science. At the highest reasoning setting reported for all seven models, GPT-6 Astra earned 66.7% mean rubric reward, ahead of GPT-6.1 Sol at 62.0% and GPT-5.6 Sol at 43.8%. BioBasicsBench is a benchmark of 100 tasks built from real wet-lab experiments that tests whether AI agents can interpret raw experimental evidence, spot errors, and recognise when the evidence cannot support a conclusion.

The most important number in that paragraph is also the most misunderstood, so let us be clear about it at once. 66.7% is mean rubric reward — the average share of available credit a model earned across the tasks, with partial credit for getting parts of an answer right. It is NOT the share of tasks the model solved. A model can earn most of the credit on a task while reaching the wrong conclusion, and BioBasicsBench is designed so that exactly that kind of near-miss shows up in the score. Anyone who reports this story as "Astra solved two-thirds of lab tasks" has got it wrong.

BioBasicsBench is built the other way round. Ligo scientists took data from their own experiments and turned it into 100 tasks, each paired with an expert-written grading rubric. The inputs are the raw material of lab work: gel and blot images, colony plates, microscopy stacks, sequencing reads, plate-reader exports, bench photographs and protocol videos. The tasks span five areas of routine work — protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning — and they range from simple reads, such as the molecular weight of a DNA sample in a gel, to detailed assay analysis. Every task went through three reviews: one by a Ligo scientist who had not written it, one by an outside PhD-level scientist working in a wet lab, and a final internal review of the prompt, the evidence, the rubric, the author notes and a graded model attempt.

Illustration created for INORYVA

The Results: Astra Leads, and the Gap Is Large

OpenAI evaluated seven of its own 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. The evaluation covered 37 model and reasoning-effort settings, with five attempts on each of the 100 tasks per setting, for a total of 18,500 scored results. The headline figures below are at "xhigh", the highest reasoning-effort setting reported for all seven models. Mean reward is the average task score as a percentage of available credit, and the error bars are 95% confidence intervals calculated by resampling whole tasks.

Model (xhigh effort)	Mean reward	95% CI	Output tokens per task
GPT-6 Astra	66.7%	60.3–73.1%	4,704
GPT-6.1 Sol	62.0%	55.6–68.4%	5,238
GPT-5.6 Sol	43.8%	37.6–50.0%	9,099
GPT-6 Luna	38.3%	32.3–44.5%	13,120
GPT-5.5	35.4%	29.7–41.0%	11,135
GPT-5.4	33.2%	28.3–38.4%	13,717
GPT-5.2	31.2%	26.2–36.3%	27,620

The confidence intervals matter too. Astra's 95% interval runs from 60.3% to 73.1%, and GPT-6.1 Sol's runs from 55.6% to 68.4%. The two intervals overlap, so the gap between the top two models is not as firm as the headline suggests — on a different set of 100 tasks, their order could narrow or even flip. The gap between the GPT-6 leaders and the GPT-5 models, by contrast, is wide and clear: even the bottom of Astra's interval sits above the top of GPT-5.6 Sol's. That is a genuine generational step, not a statistical wobble.

Where Astra Still Failed

OpenAI also shows the failures, and they are just as specific. In a colony-PCR task, Astra could see the two groups of bands on the gel, but it assumed that a larger PCR product would identify successful cloning, and on that assumption it proposed lanes 4, 6, 7 and 8. The reference interpretation, comparing against the controls, selects lanes 5, 9 and 10. Astra's visual grouping was correct and its experimental conclusion was reversed — and following its answer would have meant purifying the wrong plasmid. In a colony-counting task, Astra failed to separate close neighbours, missed visible colonies, and marked empty agar as a colony, producing an overcount of about 12% against the reference. An inaccurate colony count directly affects estimates of how efficiently DNA was introduced into cells, so a 12% overcount inflates that estimate by the same proportion.

OpenAI's own summary of these failures is blunt: "Higher benchmark scores do not yet mean models can make reliable judgments in the lab." The pattern in the failures is the point. Astra's observations are often right and its conclusions wrong. It sees the bands, the colonies and the contrast — and then lets an assumption, an expected answer, or a supplied label override what the evidence actually supports. That is a harder problem than reading an image correctly, and it is the problem that will decide whether AI assistants can be trusted with real lab decisions.

Illustration created for INORYVA

What the Benchmark Does Not Tell Us

Second, the benchmark cannot yet be independently checked. At drafting, OpenAI's announcement contained no public link to the dataset, the rubrics, or the scoring code, and an independent check on 6 November found the same absence. Outside labs cannot yet run their own models on BioBasicsBench, reproduce OpenAI's scores, or test whether the rubrics reward the right things. Until that link appears, every figure in this article is OpenAI's own report of its own test, and it should be described that way. The benchmark connects to wider AI-in-biology work, including Claude AI drug discovery, the UK AI biosecurity lab, the State of AI Report 2026 and earlier GPT-6 Astra benchmark scores. If OpenAI publishes the dataset later, the scores can be re-checked — and they should be.

Illustration created for INORYVA

Key Takeaways

• BioBasicsBench is a 100-task benchmark from OpenAI and Ligo, published 6 November 2026, testing AI agents on raw wet-lab evidence — gels, colony plates, microscopy, sequencing files and instrument exports.
• At xhigh reasoning effort, GPT-6 Astra scored 66.7% mean rubric reward (95% CI 60.3–73.1%), ahead of GPT-6.1 Sol at 62.0% and GPT-5.6 Sol at 43.8% — mean rubric reward is partial credit, not the share of tasks solved.
• The evaluation covered 7 OpenAI models, 37 model/effort settings and 5 attempts per task, for 18,500 scored results; Astra also used roughly half the output tokens of GPT-5.6 Sol (4,704 vs 9,099 per task).
• Astra still made serious errors: it picked the wrong colony-PCR lanes (4, 6, 7, 8 instead of 5, 9, 10) and overcounted colonies by about 12% — OpenAI says higher scores do not yet mean reliable lab judgment.
• Only OpenAI's own models were tested, no human baseline was published, and the announcement contained no public link to the dataset, rubrics or scoring code, so the scores cannot yet be independently verified.

Conclusion

BioBasicsBench is the right kind of test, and its first results are both encouraging and sobering. Encouraging, because GPT-6 Astra's jump over the GPT-5 models is large, clear, and efficient — it reads messy lab evidence far better than its predecessors, and it does so while writing less. Sobering, because the same model that spots a mislabelled control plate will then pick the wrong colonies from a gel and overcount the ones it does find. A laboratory assistant that is right about the evidence and wrong about the conclusion is not yet a laboratory assistant. The benchmark's real value is not the 66.7% headline. It is the 100 tasks, the expert rubrics, and the honest failure cases — a measuring stick the whole field can use, once OpenAI publishes it for everyone to run. Until then, the fair summary is OpenAI's own: progress on routine lab analysis, with room for further improvement, and no claim yet to reliable judgment at the bench.

Sources

• OpenAI — Measuring how AI agents handle everyday wet-lab science, 6 November 2026 — https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science (primary source, opened and read in full at drafting)
• Unite.AI — OpenAI Introduces BioBasicsBench to Test AI Agents on Wet-Lab Data, 6 November 2026 — https://www.unite.ai/openai-introduces-biobasicsbench-to-test-ai-agents-on-wet-lab-data/
• Kingy.ai — BioBasicsBench: OpenAI's Wet-Lab AI Results Explained, 6 November 2026 — https://kingy.ai/blog/biobasicsbench-openai-wet-lab-ai-guide/ (no public link to dataset, rubrics or scoring code found on 6 November)
• R&D World — OpenAI's BioBasicsBench tests AI on everyday wet-lab data, 6 November 2026 — https://www.rdworldonline.com/openai-unveils-biobasicsbench-a-benchmark-to-gauge-how-well-ai-agents-read-raw-experimental-data-across-100-routine-wet-lab-tasks/