OpenAI Made a Test to See if AI Can Handle Messy Lab Work
When people test AI on science, they usually use clean, polished data from published papers. But real lab work is nothing like that. It’s messy, things go wrong, and results don’t always make sense. So on November 6, 2026, OpenAI introduced BioBasicsBench, a new benchmark with 100 tasks, built together with an enzyme-design startup called Ligo. It’s meant to check whether AI can actually read raw, messy lab evidence.
What’s in the test?
The tasks cover five areas of everyday lab work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. The AI has to deal with raw gel and blot images, microscope stacks, sequencing reads, plate-reader exports, and even protocol videos. That means it has to work through unclear data, failed experiments, and results that don’t agree with each other.
To keep the tasks fair, they used three layers of checking. A Ligo scientist who didn’t write the problem reviewed it first. Then an outside PhD-level lab scientist looked at it. Finally, there was a last internal check of the evidence, prompts, and grading rubrics.
How did the AI do?
OpenAI tested 37 different model and effort settings. These results are reported by the company itself. Its top model, GPT-6 Astra, got a 66.7% average partial-credit score at the highest reasoning setting, called “xhigh.” That’s a big jump from GPT-5.6 Sol, which got 43.8% under the same conditions while using almost twice as many output tokens per task.
OpenAI says the improvement comes from better built-in multimodality and better reasoning about what happened in an experiment. They gave a few examples:
- Cell imaging: Older models mixed up dying cells because of the raw dye signals. Astra noticed the detection channels were overlapping and asked for control experiments to separate the interference.
- DNA cloning: A negative-control plate was labeled with the wrong antibiotic. Astra caught it and rejected it as an invalid control, while older models just accepted the result.
- Microscopy: Astra figured out it was looking at a z-stack depth scan, instead of making up biology reasons for why cells were drifting out of focus.
But don’t get too excited
OpenAI itself warned that a higher score doesn’t mean the AI has good lab judgment. The grading gives partial credit for finding correct data points or logical steps, so a model can score pretty high and still make the wrong final call. The models often make correct observations and then draw the wrong conclusion. Even worse, they still keep misreading experimental controls and baseline conditions. Those are basic mistakes that could wreck real bench research if nobody is watching.
The big catch
BioBasicsBench is still reported by the vendor only. Nobody else has replicated it, and OpenAI only tested its own models. There are no comparisons with rival AI models or with human lab technicians. Since the test uses complicated multimodal inputs, researchers will also need good human baselines across different lab fields before the scores really mean something.
So the benchmark is a useful new way to test multimodal reasoning, but a good score isn’t the same as a working lab assistant. Re-running a code test can’t copy how unpredictable it is to actually do experiments with your hands. There’s still a big gap between reading evidence on a screen and doing reliable science at the bench.
