BioBasicsBench: how agents handle everyday wet-lab work

What BioBasicsBench measures

OpenAI introduced BioBasicsBench on November 6, 2026, in collaboration with Ligo, a biotech company focused on enzyme design. The benchmark asks whether AI agents can interpret raw experimental data from everyday wet-lab work: spotting problems with controls, and recognizing when the evidence cannot support a conclusion.

Existing benchmarks, OpenAI argues, often use charts and images selected for papers, and can miss the ambiguity scientists actually troubleshoot at the bench. The 100 tasks cover five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning.

Scientists wrote the tasks from data from their own experiments, each paired with expert grading criteria. Every task went through 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. Agents return a structured answer, and credit is assigned against individual rubric criteria, so a correct conclusion can be separated from a partly correct analysis.

What the agent sees

The inputs keep the forms of evidence scientists work with, and a task can include more than one kind. Categories overlap. Reference sheets such as ladder charts and plasmid maps are supporting material.

• Gels and blots: 31 tasks (agarose DNA gels, SDS-PAGE, Western blots).
• Protocols, notes, and specs: 27.
• Tables and instrument exports: 20, including CSV, Excel, and raw flow cytometry FCS files.
• Bench, apparatus, and reagent photos: 18.
• Microscopy: 14, including raw brightfield and fluorescence images, z-stacks, and time-lapse movies.
• Instrument screens and plots: 13.
• Sequencing files: 12 (FASTQ, FASTA, GenBank maps).
• Protocol videos: 7; agar plates and colonies: 7; reference sheets: 21.

How the models scored

OpenAI evaluated seven 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 run covers 37 model and reasoning-effort settings, with five attempts on each of the 100 tasks, for 18,500 results. The reported metric is mean reward: the average task score as a percentage of available credit.

At the xhigh setting reported for all seven models, GPT-6 Astra earned 66.7% (95% CI 60.3 to 73.1%) with 4,704 output tokens per task. GPT-6.1 Sol scored 62.0% (55.6 to 68.4%). GPT-5.6 Sol scored 43.8% (37.6 to 50.0%) with 9,099 tokens. GPT-6 Luna scored 38.3%, GPT-5.5 35.4%, GPT-5.4 33.2%, and GPT-5.2 31.2%, the last of those using 27,620 tokens per task.

OpenAI says Astra improves on GPT-5.6 Sol even at low reasoning effort, and suggests that stronger native multimodal capabilities may be a key reason.

We have written before about Astra on a scientific eval, TheoPhysEval, where it was tested on theoretical physics problems: TheoPhysEval: Astra solves 49 of 50 physics problems.

Where Astra improved, and where it still fails

In one example, GPT-5.6 Sol classified cells as dead based on the signal from a dye used to detect dead cells. The detector could also pick up light from a second dye, because the detection windows overlapped at 525 to 540 nm, so brightness alone could not justify that conclusion. Astra recognized the ambiguity and called for better controls to separate the signals.

In a DNA cloning task, GPT-5.6 Sol saw no colonies on the negative control plate and based its conclusions on that observation. Astra noticed that the plate rim was labeled "LB Kan50," identifying kanamycin, while the experiment used ampicillin, and correctly rejected the plate as a reliable control.

In a series of microscope images, cells appeared to come into and out of view. GPT-5.6 Sol looked for a biological explanation. Astra correctly recognized a z-stack: images of the same location taken at different depths.

Higher scores still do not mean reliable lab judgment. On a colony PCR readout, Astra correctly grouped two sets of bands but assumed a larger PCR product would identify successful cloning. It proposed lanes 4, 6, 7, and 8. The reference interpretation, based on the controls, selected lanes 5, 9, and 10.

In another task it concluded that vacuoles are generally darker than the surrounding cell, even though some appear lighter in the same image. Changing focus can reverse the contrast, so the darker examples do not establish a general rule. On a routine colony count it overcounted by about 12%: it failed to resolve close neighbors, missed visible colonies, and marked empty agar as a colony.

Why it matters

Everyday wet-lab analysis (reading raw images, checking controls, and noticing when the data cannot support a conclusion) is now a measurable target rather than an anecdote. OpenAI says progress here could help scientists spend less time troubleshooting and more time deciding which experiments to run next. The same capabilities also matter for tracking biosafety and biosecurity implications.

If you were putting an agent next to the bench, the benchmark's message is practical: the model is getting better at fine visual detail, but a familiar assumption or a misread label can still override the controls. 66.7% is progress, not a license for unsupervised lab calls.