BioBasicsBench: GPT-6 Astra beats GPT-5.6 Sol in the lab but still doesn't know when to hold back

By Unite.AI · November 6, 2026.

On November 6, OpenAI unveiled BioBasicsBench, a 100-task benchmark built with Ligo, an enzyme design company. It measures whether an AI agent can interpret raw evidence from routine wet-lab experiments (labs that work with real biological material, not simulations).

According to OpenAI's own data, at the highest reasoning level (xhigh), GPT-6 Astra achieved an average reward of 66.7% (95% confidence interval: 60.3–73.1%). GPT-5.6 Sol came in at 43.8% (37.6–50.0%). Astra also used fewer tokens per task: 4,704 on average versus 9,099 for Sol. Seven models were evaluated across 37 configurations, with five attempts per task: 18,500 results in total.

What sets it apart from other tests is the material. Many scientific benchmarks use charts and images selected for academic papers. Here, the tasks come from scientists' real experiments, complete with failures, ambiguous images and contradictory observations. Inputs include gels, colony plates, microscopy, sequencing reads, plate-reader exports and protocol videos. The rubric rewards numerical results and reasoning, as well as recognizing when the evidence is insufficient to draw a conclusion. Each task went through three reviews, one of them external, conducted by a PhD scientist.

The examples OpenAI provides are the most interesting part. In a cell-imaging task, Sol concluded that some cells were dead based on a dye's signal. Astra noticed that the detection windows overlapped and that a second dye could interfere, so it asked for controls to separate the signals. In a DNA cloning task, Sol focused its conclusion on a plate with no colonies, while Astra noticed that the plate was kanamycin rather than ampicillin, the experiment's antibiotic, and discarded the control. In another, Sol looked for a biological explanation for cells drifting in and out of focus; Astra identified a z-stack (a series of images taken at different depths).

OpenAI itself is candid about the limitations. In a colony PCR task, Astra picked lanes 4, 6, 7 and 8, assuming a larger product indicated correct cloning; the reference answer was 5, 9 and 10. That mistake would lead to purifying the wrong plasmid. In a colony count, according to OpenAI, Astra overcounted by about 12% by failing to separate neighboring colonies, missing visible colonies and marking empty agar, which inflates the estimated efficiency of DNA introduction by the same proportion. It also generalized that vacuoles are darker than their surroundings despite seeing the opposite in the same image.

Our take: the thesis is that the leap is real, but the failure that matters has changed in nature. Astra no longer stumbles as much at “looking”; it stumbles at letting what it observes govern its conclusion. The fact that its intermediate observations recorded that contrast varied with focus, yet it still generalized the opposite, is a pattern worth watching: it is not ignorance, it is inconsistency between perception and decision. In a lab, that costs weeks and reagents.

There are caveats. This is a benchmark designed and published by OpenAI with a partner, and the results are its own; independent replications remain to be seen. The 66.7% also means a third of the credit is lost on “basic biology” tasks. And the rubric rewards flagging insufficient evidence, a human virtue that has rarely been measured until now and is highly valuable.

More broadly, the lab is the bottleneck of biomedicine: science advances at the pace at which experiments can be run and read. An agent that reliably interprets gels, plates and microscopy speeds up that cycle, and therein lies the concrete path toward the eradication of disease that we champion as a long-term goal. In the short term, however, the value lies in supervised assistance, not autonomy, and also in an area OpenAI acknowledges it must assess: the biosecurity implications of these capabilities.

Winners: teams that use these agents as a second pair of eyes with human checks. Losers: anyone who mistakes a 66.7% aggregate for reliable lab judgment. The right yardstick is not the average score, but how often the agent carries its own observation through to the conclusion.