BioBasicsBench Puts GPT-6 Astra at 66.7% on Wet-Lab Analysis. Only OpenAI Models Are on the Board, the Same Catch as LifeVal.
OpenAI published BioBasicsBench on November 6, built with the biotech company Ligo: 100 wet-lab analysis tasks over raw gels, plates, microscopy, instrument exports, and videos, graded on expert rubrics. At the highest reasoning setting GPT-6 Astra scores 66.7% mean reward with a 95% interval of 60.3 to 73.1, using 4,704 output tokens per task; GPT-6.1 Sol scores 62.0% and GPT-5.6 Sol 43.8%. Seven OpenAI models and no one else, self-built and self-graded, four days after LifeVal did the same. OpenAI's own failure cases show Astra picking the wrong colonies, overcounting by about 12%, and letting assumptions override controls. We read the post.
By T.W. Ghost
OpenAI published BioBasicsBench today, November 6, 2026, with the biotech company Ligo: 100 wet-lab analysis tasks across five areas, built from the raw material a bench scientist actually looks at, gels, plates, microscopy images, instrument exports, and videos, and graded against rubrics written by experts. OpenAI ran seven of its own models at 37 settings, five attempts each, for 18,500 results. At the highest reasoning setting, GPT-6 Astra scores 66.7% mean reward, with a 95% confidence interval of 60.3 to 73.1 and 4,704 output tokens per task. GPT-6.1 Sol scores 62.0%, GPT-5.6 Sol 43.8%, and the older models fall away from there.
It is the twin of LifeVal, published four days earlier, and it has the twin's catch. OpenAI built the benchmark, OpenAI graded it, and only OpenAI models are on it. There is no Anthropic row, no Google row, no open-weight row, so 66.7% is a position on OpenAI's own ladder and nothing else. The more useful part of the post is the section OpenAI titled "where models still fall short," because the failures are OpenAI's own examples: Astra assuming a larger PCR product means success and picking the wrong colonies; overcounting colonies by about 12%; letting an assumption override what the controls showed; discounting a contradictory vacuole contrast. We read the post. Here is what held up.
The Board, All of It OpenAI's
| Model, highest reasoning setting (OpenAI's table) | Mean reward | 95% interval | Output tokens per task |
|---|---|---|---|
| GPT-6 Astra | 66.7% | 60.3 to 73.1 | 4,704 |
| GPT-6.1 Sol | 62.0% | 55.6 to 68.4 | 5,238 |
| GPT-5.6 Sol | 43.8% | 37.6 to 50.0 | 9,099 |
| GPT-6 Luna | 38.3% | ||
| GPT-5.5 | 35.4% | ||
| GPT-5.4 | 33.2% | ||
| GPT-5.2 | 31.2% | ||
| Models from other labs | None |
Progress on OpenAI's ladder is real and not the claim that matters. From 31.2% on GPT-5.2 to 66.7% on Astra is a steep curve. It says nothing about whether Astra beats a Claude or a Gemini at reading a gel, because nobody ran one.
The failure cases are the honest part. Wrong colonies, a 12% overcount, assumptions overriding controls. Those are OpenAI's own words about its best model, and they describe a tool that needs a scientist checking it, not one that replaces the check.
The setting is not a cheap call. The headline score is at the highest reasoning effort, 4,704 output tokens for every task. That is the benchmark configuration, not what a production pipeline would run per image.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| BioBasicsBench published Nov 6, 2026; 100 tasks with Ligo; five wet-lab areas | Verified (OpenAI post) |
| Seven OpenAI models; 37 settings; five attempts; 18,500 results | Vendor |
| Astra 66.7% (60.3 to 73.1); Sol 62.0%; 5.6 Sol 43.8%; token counts | Vendor (OpenAI's table) |
| Failure cases: wrong colonies, about 12% overcount, assumptions over controls | Vendor (OpenAI's own examples) |
| Models from other labs on the board | None |
| Independent replication | None |
What To Do
Cite the post for what the benchmark is and label every percentage as OpenAI's; this is an eval, not a leaderboard.
Do not read 66.7% as "AI can run the wet lab." OpenAI's own failure section is the better summary of where the model stands: useful on routine analysis, wrong in ways a technician would catch.
The frontier is elsewhere: the biology model OpenAI sells is Rosalind-6, which is Astra with fewer refusals at a higher price, and the board that scores more than one lab is Artificial Analysis's. Our model comparison has the current standings side by side.
Sources
*Not sure whether a vendor's own benchmark should move your lab's model choice? Take the free 2-minute quiz and get matched. Then read the everyday-tasks eval this one twins: OpenAI's LifeVal.*