OpenAI Introduces 100-Task Benchmark for AI in Wet-Lab Science

OpenAI introduced BioBasicsBench on Nov. 4 to assess how AI agents interpret raw laboratory evidence, with the 100-task evaluation emphasizing ambiguous results and experimental troubleshooting in wet-lab science.

Ligo, a biotechnology company focused on enzyme design, worked with OpenAI on the benchmark. Its tasks cover protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning.

Exercises draw on gel images, colony plates, microscopy data and instrument readouts. They are designed to include failed experiments, unclear results and troubleshooting situations. The benchmark uses raw experimental material rather than relying only on charts and images selected for scientific publications.

Ligo scientists built tasks from the company’s experiments and prepared grading rubrics. Each task went through review by a Ligo scientist, an external scientist with PhD-level wet-lab expertise and an internal reviewer.

OpenAI assessed seven models across 37 combinations of model and reasoning effort. Each configuration attempted every task five times, producing 18,500 results. The models included GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 Sol, GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra.

At the xhigh reasoning setting, GPT-6 Astra recorded a 66.7% mean reward, with a 95% confidence interval of 60.3% to 73.1%. GPT-6.1 Sol scored 62.0% (55.6% to 68.4%), while GPT-5.6 Sol scored 43.8% (37.6% to 50.0%). GPT-6 Luna posted 38.3% (32.3% to 44.5%), and GPT-5.5 scored 35.4% (29.7% to 41.0%).

The evaluation examines whether models notice experimental errors and recognize when the evidence is not strong enough to support a conclusion. Misreading raw results or missing a problem with a control can lead to incorrect interpretations and wasted resources in follow-up experiments.

OpenAI’s life-sciences model GPT-Rosalind comes with Rosalind Workbench, a set of tools spanning medicinal chemistry, genomics and wet-lab work.