On November 6, 2026, OpenAI introduced BioBasicsBench, a benchmark of 100 routine wet-lab tasks developed with biotechnology company Ligo. OpenAI reported a 66.7% mean reward for GPT-6 Astra at the xhigh reasoning setting. That figure is average credit earned against task-specific grading rubrics—not the percentage of tasks the model solved completely. OpenAI’s benchmark description and results explain what the score measures.

A benchmark score is most useful when you know what earns credit—and what a mistake could change in the experiment.

What BioBasicsBench asks AI models to do

BioBasicsBench tests whether models can interpret raw evidence from routine lab work, spot errors, and recognize when the evidence does not support a firm conclusion. Its 100 tasks cover protein methods; imaging and cytometry; quantification and assays; nucleic acid methods; and molecular cloning.

A task can combine several kinds of evidence: gel and blot images, photographs of plates, microscopy, sequencing files, instrument exports, protocols, lab notes, or videos of bench work. OpenAI says Ligo scientists built the tasks from their experimental data and wrote the grading criteria. Each task then received review from another Ligo scientist, an external scientist with wet-lab PhD-level expertise, and a final internal reviewer.

Some task formats overlap: OpenAI lists protocol videos in 7 tasks and agar plates and colonies in 7, while 21 tasks include reference sheets. Those counts describe different features of the task set, not separate slices that add up to 100.

What the xhigh scores mean

A rubric can award partial credit for a conclusion, measurements, reasoning, or correctly recognizing that the evidence is insufficient. OpenAI calls the average credit across tasks “mean reward” and expresses it as a percentage of the available rubric credit. A score of 66.7% therefore does not mean that GPT-6 Astra fully completed 66.7 of the 100 tasks.

ModelMean reward at xhigh95% confidence interval
GPT-6 Astra66.7%60.3–73.1%
GPT-6.1 Sol62.0%55.6–68.4%
GPT-5.6 Sol43.8%37.6–50.0%

These are OpenAI-reported averages at the highest reasoning setting it reported for all seven models. The confidence intervals were calculated by resampling whole tasks while keeping the five attempts on each task together. OpenAI evaluated seven models across 37 model-and-reasoning-effort settings, with five attempts for each of the 100 tasks per setting: 18,500 results in total, but still just 100 distinct tasks.

Ambiguous signals and a mislabeled control

A turquoise-gloved hand holds a Petri dish with pale agar; an orange outline circles the rim label.

In one microscopy task, GPT-6 Astra identified overlapping detection windows as a possible source of signal interference and asked for better controls. OpenAI gives the SYTOX-labelled channel as 470–540 nm and the FUN-1 green channel as 525–580 nm. They overlap from 525–540 nm; OpenAI did not measure the amount of interference.

A different example concerns a plate label. OpenAI says the experiment used ampicillin, while the plate’s circled rim label reads “LB Kan50,” indicating kanamycin. The mismatch matters because interpreting the sample correctly depends on matching the label to the experimental context.

A lane choice can change what gets purified

A dark gel with ten numbered lanes; colored outlines and dashed lines distinguish selected lanes.

In a colony-PCR task, GPT-6 Astra selected lanes 4, 6, 7, and 8; the benchmark reference selected lanes 5, 9, and 10. OpenAI says Astra’s selection would lead to purification of the incorrect plasmid. That is a specific task outcome, not a general error rate for the model.

OpenAI also reports one colony-counting example in which Astra’s count was about 12% above the reference. With plated volume and dilution held fixed, that overcount would inflate the estimated DNA-introduction efficiency by the same proportion.