Can autonomous artificial intelligence models reliably execute the routine analytical workload of a biological wet lab? At the maximum reasoning capacity designated as the xhigh setting, GPT-6 Astra earned an average of 66.7% of the available credit, compared with 43.8% for GPT-5.6 Sol, according to an evaluation published on the OpenAI Blog. The figures reveal measurable progress alongside a persistent performance gap in handling physical experimental workflows.
The benchmark, BioBasicsBench, measures how well artificial intelligence agents interpret raw experimental data across 100 routine wet-lab tasks. OpenAI developed the framework in collaboration with Ligo, a biotech company focused on enzyme design. The underlying question confronting operational leaders is whether frontier agents can process noisy laboratory outputs as effectively as human staff without generating compounding operational errors.
Auditing the Laboratory Workflow
BioBasicsBench evaluates performance across five areas of routine wet-lab work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Real-world laboratory execution generates failed runs, ambiguous visuals, and conflicting readouts that rarely appear in polished academic literature. The evaluation reflects these conditions by feeding models uncurated artifacts, including blot and gel images, plate photographs, sequencing reads, microscopy stacks, plate-reader exports, and bench videos.
To ensure rigorous evaluation, scientists built each problem from internal experimental data and attached specialized rubrics. Each individual task underwent an internal review by a Ligo scientist who had not authored it, an external review by a PhD-level scientist working in a wet lab, and a final internal review. Model submissions provided structured conclusions, numerical readings, or procedural recommendations scored directly against discrete rubric criteria.
How do these technical evaluation mechanics translate into day-to-day error margins across physical benches?
The Mechanics of Diagnostic Error
The gap between paper-ready data and raw readouts exposes how easily automated agents misinterpret basic biological signals. As the authors of the study note, agents can misinterpret control samples and instrument readouts.
Misreading an image or overlooking a problem with the controls used to check an experiment's results can lead scientists to draw the wrong conclusion and spend time on experiments that won't help answer their question.
When an automated system fails to register optical bleed-through across instrumentation channels, downstream teams execute protocols based on invalid premises. OpenAI evaluated seven models on BioBasicsBench: GPT‑5.2, GPT‑5.4, GPT‑5.5, GPT‑5.6 Sol, GPT‑6.1 Sol, GPT‑6 Luna and GPT‑6 Astra. Across 37 distinct model and reasoning-effort configurations, testing reached 18,500 total results by running five attempts per task.
While GPT-6 Astra marks an advance over GPT-5.6 Sol, a 66.7% success rate leaves one-third of benchmarked routine tasks unresolved or incorrectly handled. Deploying models into wet-lab data interpretation currently requires structured verification layers to catch misread controls and ambiguous assays before they trigger wasted labor and misallocated reagent runs. As reasoning models scale toward operational biology, organizations face the strategic choice of whether autonomous throughput offsets the cost of ongoing human validation.
Why it matters
A score of 66.7% across one hundred routine lab tasks looks like measurable progress, but it remains shaky ground for physical biological workflows. Any missed spectral bleed-through or misread blotting image immediately triggers cascade errors in downstream protocols, forcing human PhDs to double-check every inference. Even running at the maximum xhigh reasoning level, the system still fumbles a third of typical wet-lab edge cases. How many costly experimental runs will Ligo and OpenAI burn through before an autonomous agent stops misreading every third unedited assay?