OpenAI's BioBasicsBench tests AI agents on everyday wet-lab science

Most people assume AI in biology is all about inventing dream molecules or curing rare diseases overnight. But talk to anyone working in an actual wet lab, and they will tell you the real job involves squinting at blurry gel images, debugging why a control failed, and manually counting stubborn colonies on a petri dish.

OpenAI wants to find out if AI agents can actually survive that messy reality. Teaming up with enzyme design startup Ligo, the company rolled out BioBasicsBench, a new benchmark designed specifically to test how models handle routine wet-lab grunt work.

Real bench data, flaws and all

Existing scientific benchmarks usually rely on pristine, polished figures selected for published papers. BioBasicsBench takes the opposite route by feeding models raw experimental evidence directly from active labs.

The benchmark features 100 everyday tasks spanning five key domains: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Agents are given raw gel images, microscopy stacks, sequencing files, plate-reader exports, and even bench-work video clips. To pass, a model has to interpret the messy input, spot potential errors, and know when the data simply does not justify a verdict.

OpenAI evaluated seven models across 37 reasoning settings, tallying 18,500 total runs. At the highest reasoning setting, GPT-6 Astra scored 66.7% of available credit, easily topping GPT-5.6 Sol's 43.8%.

Smarter vision, familiar blind spots

GPT-6 Astra showed noticeable improvements in multimodal reasoning, especially when it came to avoiding unwarranted leaps of faith.

In one test, GPT-5.6 Sol declared cells dead after spotting a bright dye signal. Astra caught that the detector was also picking up light from a secondary dye, correctly arguing the signal was ambiguous without better controls. In a cloning task, Astra even noticed that a negative control plate was labeled with the wrong antibiotic, tossing it out as invalid. It also correctly recognized a microscope z-stack where earlier models assumed cells had mysteriously vanished.

Yet a 66.7% score means agents still fail a third of the time, often because preconceived notions overpower hard evidence.

In a colony PCR screening, Astra identified the visual bands correctly but fell into the classic trap of assuming a larger PCR product represented the successful construct, ultimately picking the wrong lanes. On a cell vacuole task, it brushed aside contradictory contrast variations caused by changing focus depths. And when counting colonies on an agar plate, Astra struggled with overlapping spots and marked empty agar, overcounting colonies by roughly 12%.

Why it matters

We think BioBasicsBench is a refreshingly pragmatic move from OpenAI. Instead of hyping autonomous AI scientists, it forces models to confront the mundane friction that consumes most of a lab worker's week.

For researchers and biotech teams, having an AI that flags contaminated controls or catches mislabeled plates could save thousands of dollars and weeks of wasted pipetting. But nobody should hand over the keys just yet. Overcounting colonies by 12% directly skews transfection efficiency calculations, and choosing the wrong PCR band means downstream purification is doomed from the start.

For the industry, this benchmark shifts the goalposts. Reading PubMed abstracts is table stakes now. The real test of multimodal intelligence is whether an agent can look at noisy, imperfect sensor readouts and know when to say that the data simply is not good enough.