OpenAI releases BioBasicsBench, a 100-task benchmark for AI agents doing routine wet-lab analysis
On November 6 OpenAI published BioBasicsBench, a test that checks whether AI agents can read raw results from biology labs. It was built with the enzyme-design company Ligo and has 100 tasks in five areas, from protein and DNA methods to microscopy and molecular cloning, using real gel photos, plate images, sequencing files and lab videos. At the highest reasoning setting shared by all seven tested models, GPT-6 Astra earned 66.7% of the available credit, GPT-6.1 Sol 62.0% and GPT-5.6 Sol 43.8%.
Read the original on OpenAI — Measuring how AI agents handle everyday wet-lab science ↗
Why it matters
Most science benchmarks use clean figures taken from papers, while real lab data is messy and often inconclusive. This test also rewards a model for saying when the evidence is not enough, which is what a scientist needs before trusting an AI helper. OpenAI's own examples show the best model still picks the wrong colonies, ignores data that contradicts it and overcounts colonies by about 12% on one plate.
What changes
- For developers
- Teams building lab or research agents get a public reference point that scores partial answers with expert rubrics, run five times per task across 37 model and effort settings.
- For businesses
- Biotech companies can see that newer models are much better at images and lab context, but also that human review of every conclusion is still required.
Sources
- OpenAI — Measuring how AI agents handle everyday wet-lab science ↗ openai.com Primary source
Summarized by AI from official sources on November 7, 2026. Spotted a mistake? Let us know. How it works