OpenAI Introduces BioBasicsBench: GPT-6 Astra Scores 66.7% on Wet-Lab Tasks at xhigh

On November 5, 2026, OpenAI published BioBasicsBench, a benchmark developed with Ligo, a biotech company focused on enzyme design, to measure how well AI agents handle routine wet-lab analysis. The benchmark contains 100 tasks across five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Its main departure from existing benchmarks is the input. Instead of charts selected for publication, tasks supply raw evidence: gel and blot images, colony-plate photographs, microscopy stacks, sequencing reads, and plate-reader exports. OpenAI's stated rationale is that real experiments are full of failures, ambiguous images and conflicting observations, precisely the material that paper figures filter out.

Design: raw evidence and expert rubrics

Ligo scientists wrote the tasks from data generated in their own experiments, pairing each question with an expert-written grading rubric. Every task went through three reviews: an internal check by a Ligo scientist who did not write it, an external review by a PhD-level scientist working in a wet lab, and a final internal review. Answers earn credit against individual rubric criteria, which lets the benchmark distinguish a correct conclusion from a partially correct analysis. Most tasks test several skills at once, including biological knowledge, image and data interpretation, error detection, and protocol understanding. One distinction matters when reading the results: the facts above come from the source, but rubric-based scoring means the benchmark measures agreement with a reference interpretation, not a unique answer dictated by nature.

OpenAI evaluated seven models across 37 model-and-reasoning-effort settings, with five attempts on each of the 100 tasks per setting, for 18,500 results. At xhigh, GPT-6 Astra averaged 66.7% of available credit (95% CI 60.3–73.1%) with 4,704 output tokens per task. GPT-6.1 Sol scored 62.0% (55.6–68.4%) and GPT-5.6 Sol 43.8% (37.6–50.0%). Confidence intervals were computed by resampling whole tasks. The intervals matter: Astra and GPT-6.1 Sol overlap substantially, so the claim that Astra is better than GPT-6.1 Sol is not statistically firm. The safer claim is that Astra clearly exceeds GPT-5.6 Sol, whose interval does not overlap Astra's.

Progress and failure in the same examples

The source uses several examples to illustrate progress. In a two-channel image of cells stained with SYTOX and FUN-1, GPT-5.6 Sol classified cells as dead based on brightness alone. Astra noted that the two detection windows overlap at 525–540 nm, that brightness could not justify the conclusion, and that better controls were needed to separate the signals. The source notes that each channel uses its own linear intensity scale, so brightness cannot be compared quantitatively between channels. In another task, a plate labelled “LB Kan50” was offered as a negative control for an ampicillin experiment; Astra noticed the label named a different antibiotic and rejected the plate as a reliable control. In a third, cells appeared to fade in and out across a microscopy series; GPT-5.6 Sol looked for a biological explanation, and Astra recognized a z-stack taken at different depths.

The same examples expose failure modes. On a colony-PCR gel, Astra grouped the bands correctly but reversed their experimental identities: it assumed a larger PCR product indicated successful cloning and selected lanes 4, 6, 7 and 8, whereas the reference used control comparisons to select lanes 5, 9 and 10. On a vacuole-contrast task, it noted that contrast varied with focus, yet kept only a “vacuoles are darker” generalization and omitted the lighter vacuoles in the same field. On a colony-counting task, it failed to resolve close neighbours, left visible colonies unmarked, and marked empty agar, producing a net overcount of roughly 12% relative to the benchmark reference. With plated volume and dilution fixed, that overcount inflates the estimated transformation efficiency by the same proportion.

Limits and verification

As evidence, the benchmark has clear boundaries. It has 100 tasks, all drawn from the experimental data of a single collaborator. All seven evaluated models are GPT-series models, and OpenAI co-developed the benchmark; the source reports no independent third-party replication. Scoring depends on expert rubrics, and the source does not report inter-rater agreement, task-difficulty distribution, or how disagreements were resolved. The evaluation is static: models interpret already-collected data rather than designing and running experiments. The source also does not say whether task data overlaps with any model's training corpus.

Testing whether agents can handle routine analysis requires more than one benchmark release. Useful checks include re-running the tasks on unpublished experimental data, reporting hit and false-positive rates for recognizing insufficient evidence rather than only aggregate credit, breaking errors down by skill tag, and having an independent lab conduct a prospective evaluation. Token efficiency is another signal: Astra used 4,704 tokens per task, fewer than GPT-5.6 Sol's 9,099, which suggests the higher score was not purchased with longer outputs; but the source provides no price or latency data, so this cannot be converted into a cost advantage. The author's inference is that the failures share a pattern: models often see the contradictory evidence but do not carry it into the final decision. If Astra's advantage replicates and transfers to new data, it would support a claim of usable routine analysis; if it mainly reflects conformity to reference-interpretation style, it is a benchmark-engineering result rather than a leap in lab capability.