BioBasicsBench: How AI Agents Score on Real Wet-Lab Data

OpenAI has released BioBasicsBench, a new benchmark that asks a refreshingly practical question: can an AI agent handle the everyday analysis a working biologist does at the bench? Instead of polished charts pulled from published papers, the benchmark uses the messy raw evidence of real wet-lab work, from smudged gel images to colony plates and microscopy stacks. OpenAI built it with Ligo, a biotech company focused on enzyme design, and published the results on November 6.

• BioBasicsBench contains 100 routine wet-lab tasks across five areas of laboratory work
• GPT-6 Astra scored 66.7% of available credit at the xhigh setting, up from 43.8% for GPT-5.6 Sol
• Every task was written from scientists' own experiments and passed three rounds of expert review
• The evaluation covered seven models, 37 settings and 18,500 graded results

What Is BioBasicsBench?

The benchmark spans protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Tasks range from reading the molecular weight of a DNA band on a gel to working through a full assay analysis. The inputs are deliberately realistic: agarose gels and Western blots, plate photographs, raw fluorescence images, long-read sequencing files, plate-reader exports, photos of instrument screens, and even short videos of bench techniques such as serial dilution.

Crucially, the benchmark rewards scientific caution. Many tasks test whether an agent can spot an error in the controls or recognize when the evidence simply cannot support a conclusion. Answers are graded against expert-written rubrics, so a partially correct analysis earns partial credit rather than a flat pass or fail.

How Did GPT-6 Astra Perform on Wet-Lab Tasks?

At the xhigh reasoning setting, GPT-6 Astra led the field at 66.7%, followed by GPT-6.1 Sol at 62.0%. GPT-5.6 Sol scored 43.8% and GPT-5.2 scored 31.2%. Astra also got there efficiently, using about 4,700 output tokens per task compared with more than 9,000 for GPT-5.6 Sol. OpenAI suggests that stronger native multimodal ability, the capacity to really read an image, is a key driver of the jump.

One example in OpenAI's write-up shows the kind of judgment the benchmark measures. Shown cells stained with two dyes, an older model labeled cells dead based on brightness alone. Astra noticed that the detector could pick up light from both dyes and asked for better controls instead of overreaching.

Where AI Lab Assistants Can Still Improve

OpenAI is candid that higher scores do not yet mean an agent can make reliable lab calls on its own. Its examples show Astra sometimes letting a familiar assumption override the controls during a cloning check, setting aside contradictory microscopy observations, and overcounting colonies on a plate by roughly 12%. Publishing these misses is valuable: it gives researchers a clear map of what to improve next.

Why a Wet-Lab Benchmark Matters for AI in Science

Everyday analysis is where scientists lose hours to troubleshooting, so an AI assistant that reads gels and plates accurately could free them to focus on which experiment to run next. OpenAI also says tracking these capabilities helps it understand biosafety and biosecurity implications. This continues a busy stretch of practical evaluations, following the LifeVal everyday-tasks benchmark and the earlier GeneBench Pro biology benchmark. For more research news, visit our AI coverage.

Sources: OpenAI — Measuring how AI agents handle everyday wet-lab science — November 6, 2026.