BioBasicsBench by OpenAI: A benchmark for how AI agents read raw lab data

BioBasicsBench is OpenAI’s new benchmark for how AI agents handle the routine analysis work of a wet lab (hands-on experiments with real samples and instruments). OpenAI developed it with Ligo, a biotech company focused on enzyme design.

Most science benchmarks use clean, curated figures from papers. Real lab work is messier, with failed experiments, ambiguous images and results that need troubleshooting. BioBasicsBench tries to test that. It has 100 tasks built by scientists from their own experiments, and agents have to interpret raw lab data, spot errors and make decisions when results don’t look right.

What’s in it:

Task areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning
Evidence types: gel and blot images, colony plates, microscopy stacks, sequencing reads, plate-reader exports, bench photos and protocol videos
Scientists wrote the tasks and the grading rubrics, which give credit for correct numbers, supporting reasoning and recognizing when the evidence isn’t enough
Tasks were reviewed by a Ligo scientist who didn’t write them and by an outside PhD-level wet-lab scientist

On the results, GPT-6 Astra scored 66.7% mean reward at its highest reasoning effort, against 43.8% for GPT-5.6 Sol, using about half the output tokens. OpenAI’s own caveat is that higher scores “do not yet amount to reliable laboratory judgment”. Even the best model made mistakes, like picking the wrong gel lanes and overcounting colonies by about 12%.

Good fit if you work on AI for science, lab automation or evaluation, and want a benchmark based on real bench work instead of tidy textbook cases.

OpenAI’s announcement

What’s in it:

Task areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning
Evidence types: gel and blot images, colony plates, microscopy stacks, sequencing reads, plate-reader exports, bench photos and protocol videos
Scientists wrote the tasks and the grading rubrics, which give credit for correct numbers, supporting reasoning and recognizing when the evidence isn’t enough
Tasks were reviewed by a Ligo scientist who didn’t write them and by an outside PhD-level wet-lab scientist

Good fit if you work on AI for science, lab automation or evaluation, and want a benchmark based on real bench work instead of tidy textbook cases.

OpenAI’s announcement