OpenAI BioBasicsBench Tests Whether AI Agents Can Read Everyday Wet-Lab Data

OpenAI just published BioBasicsBench, a benchmark built with biotech company Ligo that asks a blunt question: can AI agents actually interpret the messy evidence scientists see every day in a wet lab?

Paper-ready charts are easy. Real benches are not. Tasks here use gel images, colony plates, microscopy stacks, instrument exports, notebook photos and short protocol videos. Agents have to spot bad controls, read ambiguous signals and say when the data simply cannot support a conclusion.

What is in the bench

The suite has 100 tasks across protein methods, imaging and cytometry, quantification, nucleic acid work and molecular cloning. Scientists wrote the prompts from their own experiments, then sent each task through internal and external PhD-level review with a grading rubric. Credit is awarded criterion by criterion, so a lucky guess on the final answer is not enough.

At the highest shared reasoning setting, GPT-6 Astra scored a mean 66.7% of available credit. GPT-5.6 Sol landed at 43.8%. Other GPT-6 and GPT-5 variants trailed further behind. OpenAI highlights gains in multimodal reading and in refusing overconfident calls when dyes overlap or a plate is labeled with the wrong antibiotic.

Why it matters for fans of the tech

This is not a sci-fi lab-bot announcement. It is a measuring stick for routine troubleshooting — the kind of work that burns hours before anyone designs a new therapy. OpenAI is clear that higher scores still do not mean reliable lab judgment: the same write-up shows Astra miscounting colonies, flipping plasmid identities and discarding contradictory vacuoles. Progress is real; trust is not automatic.

If you follow how frontier models are tested outside chat demos, BioBasicsBench is worth a bookmark. Full details, tables and failure cases are on the OpenAI announcement.