OpenAI has published a test of how well its AI models read the messy data that comes off a real laboratory bench, and its newest model, GPT-6 Astra, scored 66.7 percent of the available credit on average. Ligo, a biotech company that designs enzymes, co-developed the test, which is called BioBasicsBench. OpenAI posted the results on its own site on 6 November 2026, which means the company is grading its own models on an exam it built with a partner.

The benchmark holds 100 routine tasks, written by Ligo scientists from their own experiments, each with a grading rubric written by experts. Most inputs are raw files and photos: 31 tasks use images of gels and blots, 27 use protocols and notes (some of them photographed handwriting), and the rest cover microscope time-lapses, flow cytometry exports, agar plates, and sequencing files. A gel is a slab of jelly where DNA or protein samples spread into bands, and where a band lands tells you how large the molecule is. Answers earn partial credit against the rubric, and a second Ligo scientist and an outside PhD-level wet-lab scientist reviewed every task.

OpenAI ran seven of its own models through 37 model and effort settings, with five attempts per task, for 18,500 results. At the highest reasoning setting, Astra averaged 66.7 percent, followed by GPT-6.1 Sol at 62.0 and then GPT-5.6 Sol at 43.8, and the oldest model, GPT-5.2, at 31.2. Astra’s margin of error runs from 60.3 to 73.1 percent, so the 62.0 from GPT-6.1 Sol sits inside it and the gap between the top two is hard to call. The jump from GPT-5.6 Sol to Astra, 22.9 points, is not in doubt. OpenAI adds that Astra outscored GPT-5.6 Sol with the reasoning dial turned down, and suggests better built-in understanding of images and video is the reason.

The examples show where the gain came from. In a DNA cloning task, a negative control is a plate that should grow nothing, because growth would mean the experiment is contaminated. GPT-5.6 Sol accepted a blank plate as a valid control. Astra read the label on the plate’s rim, “LB Kan50”, spotted that it named kanamycin when the experiment used a different antibiotic, ampicillin, and threw the control out.

Astra also caught a trap in a cell-staining task. A dye meant to mark dead cells can, on some detectors, pick up a second dye as well. GPT-5.6 Sol counted cells as dead anyway. Astra pointed out that the signal was ambiguous and requested better controls. In a microscope series where cells kept sliding out of focus and back, GPT-5.6 Sol hunted for a biological explanation for why they vanished, while Astra recognized a z-stack, a set of images taken at different depths.

The misses are just as concrete. In a colony PCR task, Astra let an assumption outrank the controls and picked lanes 4 and 6 through 8 where the reference answer was 5, 9, and 10, a mistake that would send a lab off to purify the wrong plasmid. It dismissed lighter-looking vacuoles because its own rule said vacuoles look darker. And it overcounted colonies on a plate by about 12 percent, which would inflate the estimate of how efficiently DNA got into cells by a matching 12 percent. OpenAI says plainly: “Higher benchmark scores do not yet mean models can make reliable judgments in the lab.”

The company also says tracking these skills will help it understand biosafety and biosecurity implications, though the post gives no detail on how. The post also includes no results from any other company’s models, so the 66.7 has no outside yardstick yet.

For a lab thinking about putting an AI assistant on its gel photos, a third of the credit went unclaimed by the best model, and the unclaimed part includes wrong lane picks that cost days of bench work, so a person still has to check the controls.

Reported by OpenAI on 6 November 2026.