OpenAI's New Test Asks AI Agents to Read a Gel, Pick Colonies and Spot a Bad Control. The Best Score Was 66.7%.

Picture a first-year lab tech holding a photo of a gel. Three bands, a ladder on the left, a smudge near the bottom. What size is the DNA? It's a routine question. It's also the kind of thing that, answered wrong, quietly ruins a week of work.

OpenAI built a test around questions like that. It's called BioBasicsBench, and it asks AI agents to handle everyday wet-lab science. The headline number is 66.7%. The more useful numbers are in the mistakes.

What the Test Is

BioBasicsBench has 100 tasks, built with Ligo, a biotech firm focused on enzyme design. Ligo scientists wrote each task from their own experiments and paired it with an expert-written rubric. Every task was reviewed by a second Ligo scientist, an outside PhD-level wet-lab scientist, and a final internal check.

The tasks span five areas: protein methods, imaging and cytometry, quantification and assays, nucleic-acid methods, and molecular cloning. What sets them apart is the input. Instead of tidy figures pulled from published papers, the agents get raw evidence: gel images, colony plates, microscopy, instrument readouts, protocol notes and bench photos. The test doesn't cover pipetting, running instruments or robotic execution. It tests reading and judgment, not hands.

How the Models Did

Coverage of OpenAI's results says seven OpenAI models were tested across 37 model-and-effort settings, with five attempts per task, for about 18,500 results. Scores are partial rubric credit averaged across tasks.

At the top reasoning setting, GPT-6 Astra scored 66.7%, followed by GPT-6.1 Sol at 62.0%. GPT-5.6 Sol came in at 43.8%, GPT-6 Luna at 38.3%, GPT-5.5 at 35.4%, GPT-5.4 at 33.2% and GPT-5.2 at 31.2%. The error bars matter. Astra's and GPT-6.1 Sol's ranges overlap, so the ordering at the top isn't clear. The gap down to the older models is. Coverage says OpenAI credits Astra's gains to stronger native multimodal capabilities, and that it used about half the output tokens of GPT-5.6 Sol.

Where the Models Went Wrong

The failures are the good part. The write-ups describe a few:

• Colony PCR. Asked which bacterial colonies were correctly cloned, Astra assumed a larger PCR product meant success. It picked colonies 4, 6, 7 and 8, while the reference answer was 5, 9 and 10.
• Counting. One model overcounted colonies by about 12%, which would inflate a transformation-efficiency estimate by the same proportion.
• Microscopy. Models misread changes in focus as biology and ignored lighter regions of the same image.
• A bad control. In one cloning task, a negative-control plate was labeled for one antibiotic while the experiment used another. GPT-5.6 Sol accepted the empty control. Astra caught the mismatch.

These aren't exotic mistakes. They're what a rushed beginner does, and what a careful one learns to catch.

OpenAI's own conclusion, as quoted in coverage: higher benchmark scores don't yet mean models can make reliable judgments in the lab.

Our Take

Two-thirds credit sounds close. In a lab it isn't a pass. And the errors that matter aren't random. They're confident misreadings of the evidence, which is exactly the failure that's hard to catch if you're trusting an agent to run unsupervised.

The benchmark is useful anyway, for a plain reason: it measures something close to the real work. Many AI biology tests ask about textbook knowledge. This one hands over the messy photo.

Coverage says OpenAI plans to use the results to inform its biosafety and biosecurity tracking. That makes sense, since the same evidence-reading skill that helps with enzyme design could help with harmful work. How the scores feed into safeguards hasn't been spelled out.

What We Cannot Conclude

• We don't know how humans score. The coverage reports no human or expert baseline, so we can't say how far 66.7% is from a trained technician.
• We can't compare with other companies' models. Only OpenAI's models were tested.
• The score isn't a success rate. It's partial rubric credit averaged across tasks.
• The tasks aren't public. Outside researchers can't check the benchmark or try to reproduce the results.
• We haven't read OpenAI's post directly. Our account comes from coverage that quotes it, and outlets differ on whether the results date from Nov. 4 or Nov. 6.
• A rumor circulated about a hidden higher-scoring model. One outlet says it was a hoax and that OpenAI's results list only seven models.

Why It Matters

• It measures judgment, not recall. Reading raw lab evidence is where agents will have to prove themselves.
• The errors are the signal. Confident misreads of control plates and gels are the failure modes to watch.
• It's a stepping stone, not a green light. OpenAI says the scores don't mean models are ready for the lab.
• Safety teams will use it too. The same skills can cut both ways, so tests like this help track risk.