GPT-6 Astra Hits 66.7% on OpenAI's New Wet-Lab Benchmark

OpenAI's GPT-6 Astra scored 66.7% mean reward on BioBasicsBench, a new 100-task benchmark the company introduced on November 6, 2026, to test whether AI agents can interpret raw evidence from routine wet-lab experiments. The benchmark was developed with enzyme-design company Ligo, and the top result still leaves a third of the available credit on the table.

The comparison model, GPT-5.6 Sol, managed 43.8% with a 95% confidence interval of 37.6–50.0%. Astra's interval ran 60.3–73.1%. Astra also worked leaner: it averaged 4,704 output tokens per task versus 9,099 for Sol. OpenAI reports results as mean reward — the average task score expressed as a percentage of available credit — at xhigh, the highest reasoning-effort setting reported for all seven models tested.

The takeaway: the newest model wins decisively, but neither clears the bar for reliable lab judgment.

Why build a benchmark from failed experiments?

BioBasicsBench attacks a limitation OpenAI identifies in evaluations built around charts and images selected for scientific papers. Everyday lab work also involves failed experiments, ambiguous images, conflicting observations and troubleshooting. The benchmark's tasks ask agents to interpret experimental data, spot errors and recognize when the evidence cannot justify a conclusion.

The tasks span five areas:

• protein methods
• imaging and cytometry
• quantification and assays
• nucleic acid methods
• molecular cloning

Scientists created the tasks from their own experiments and wrote the grading rubrics. Inputs include gel and blot images, colony plates, microscopy stacks, sequencing reads, plate-reader exports, bench photographs and protocol videos. Individual tasks can combine several evidence types plus supporting reference sheets such as ladder charts and plasmid maps.

Agents return structured conclusions, measurements or recommendations. Rubrics award credit for numerical results, supporting reasoning, and — critically — recognition that the evidence is insufficient.

In plain terms: the benchmark grades whether an agent knows what the data can't tell it.

How rigorous is the grading?

Each task passed three reviews: an internal review by a Ligo scientist who had not written it, an external review by a PhD-level wet-lab scientist, and a final internal review covering the prompt, evidence, rubric, author notes and a graded model attempt.

The evaluation covered 37 model and reasoning-effort settings, with five attempts on every task per setting — 18,500 results in total. Confidence intervals came from resampling whole tasks while keeping each task's five attempts together.

That's a statistically careful design for a field where benchmarks often get one-shot runs.

What drove Astra's improvement?

The task examples show concrete gaps between the two models. In a cell-imaging task, GPT-5.6 Sol classified cells as dead from a dye signal. Astra recognized that overlapping detection windows could allow interference from a second dye and called for controls that could separate the signals.

In a DNA-cloning task, Sol based its conclusions on seeing no colonies on a negative-control plate. Astra noticed the plate was labeled for kanamycin rather than the ampicillin used in the experiment — and rejected it as a reliable control.

A third task showed cells moving in and out of focus. Sol sought a biological explanation for their apparent disappearance. Astra identified a z-stack: a series of images captured at different depths.

OpenAI reads these examples as evidence of better multimodal handling, use of experimental context, and recognition of what data can and cannot establish — and as an indication that stronger native multimodal capabilities may matter for routine wet-lab analysis.

The takeaway: Astra reads the experiment, not just the image.

Where does the benchmark expose failures?

OpenAI cautions that higher scores do not yet amount to reliable laboratory judgment. In a colony-PCR task, Astra selected lanes 4, 6, 7 and 8 after assuming that a larger PCR product indicated successful cloning. The reference interpretation selected lanes 5, 9 and 10. That mistake would lead to purification of the incorrect plasmid.

Other examples show correct observations failing to govern the final answer. Astra generalized that vacuoles — compartments inside cells — are generally darker than surrounding cell material, despite lighter vacuoles in the same image. Changing focal depth can reverse the contrast, and Astra's own intermediate observations noted that contrast varied with focus.

In a colony-counting task, OpenAI reports that Astra failed to separate close neighbors, missed visible colonies and marked empty agar, producing an approximately 12% overcount relative to the benchmark reference. With plated volume and dilution fixed, such an overcount inflates the estimated efficiency of DNA introduction by the same proportion.

The pattern: the model sees the evidence but doesn't always let it drive the conclusion.

What's the practical bar now?

OpenAI says following the development of these capabilities will also inform its assessment of their biosafety and biosecurity consequences.

The practical bar, per the company, is higher than a better aggregate score: an agent must carry accurate observations through to its conclusion, check controls and withhold conclusions the evidence cannot support. BioBasicsBench makes those requirements explicit targets for improvement — and its worked failures show why progress toward laboratory assistance still depends on closing the gap between observing evidence and making a sound experimental decision.