THE ESSENTIALS IN 30 SECONDS

  • BioBasicsBench has 100 tasks written by Ligo scientists from their own experiments, using raw inputs such as gels, blots, colony plates, microscopy and sequencing reads.
  • At the highest reasoning setting, GPT-6 Astra earned 66.7% of available rubric credit and GPT-5.2 earned 31.2%; only seven OpenAI models were evaluated.
  • OpenAI says higher scores do not yet amount to reliable lab judgment, and coverage found no human or rival-lab baseline.

OpenAI announced BioBasicsBench on 6 November, a benchmark meant to measure how well AI agents handle the routine analysis that follows a wet-lab experiment. It was developed with Ligo, a biotech company that works on enzyme design. Rather than drawing on cleaned-up figures from published papers, the 100 tasks use the kind of messy raw evidence a bench scientist actually looks at.

How the benchmark is built

Ligo scientists wrote the tasks from their own experiments and attached a grading rubric to each one. According to coverage of OpenAI's post, every task then went through three checks: a Ligo scientist who had not written it, an outside wet-lab researcher with a PhD, and a final internal review. The tasks fall into five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Inputs include agarose and SDS-PAGE gels, Western blots, colony photos, microscopy stacks, sequencing reads, plate-reader exports and videos of bench work. Difficulty ranges from reading a DNA size off a gel to analysing a full assay, and agents are also expected to spot errors and say when the data cannot support a conclusion. Answers are scored criterion by criterion, so partial credit is possible.

Results so far

OpenAI tested seven of its own models across 37 model and reasoning-effort settings, with five attempts per task. At the highest setting, GPT-6 Astra earned 66.7% of available credit, followed by GPT-6.1 Sol at 62.0%, GPT-5.6 Sol at 43.8% and GPT-5.2 at 31.2%. The newer models also used far fewer output tokens: about 4,700 per task for Astra against roughly 27,600 for GPT-5.2.

The post also lists failures. In one colony PCR task, Astra grouped gel bands correctly but gave them the wrong sample identities, which would send the wrong plasmid downstream. In another case, a colony count came out about 12% too high, and a focus shift in a microscopy stack was read as a biological event.

Limits

The benchmark comes from one partner, and every model compared is an OpenAI model, with no human or competitor baseline reported. OpenAI itself says better scores do not yet mean dependable lab judgment. The coverage reviewed here does not say whether the task set and rubrics are publicly available. Physical lab work such as pipetting is outside its scope.