OpenAI Scores Its Top Model at 66.7% on Wet-Lab Tasks

GPT-6 Astra earned an average of 66.7 per cent of the available credit on 100 routine laboratory analysis tasks, according to a benchmark OpenAI published on 6 November 2026. The same evaluation gave GPT-5.6 Sol 43.8 per cent. OpenAI built the test itself, with a biotech company called Ligo, and tested only its own models on it.

The point of BioBasicsBench is the messiness. Existing science benchmarks tend to draw on charts and images selected for publication, which strips out the ambiguity bench scientists actually face. These tasks use raw evidence instead: agarose gels and Western blots in 31 of them, photographed handwritten lab notes, raw flow cytometry files, long-read FASTQs, plate-reader exports, and seven tasks built on videos of bench work such as streaking and serial dilution. The questions span protein methods, imaging and cytometry, quantification and assays, nucleic acid methods and molecular cloning.

Scientists at Ligo wrote the tasks from their own experiments and paired each with grading criteria. OpenAI says every task passed an internal review by a Ligo scientist who had not written it, an external review by a PhD-level scientist working in a wet lab, then a final internal review.

What 66.7 per cent of the credit actually counts

The reported figure is mean reward, the average task score as a share of available credit, with answers marked against individual rubric criteria rather than pass or fail. That lets the benchmark separate a correct conclusion from a partially correct analysis. OpenAI evaluated seven models across 37 model and reasoning-effort settings, five attempts on each of the 100 tasks per setting, for 18,500 results.

The token column is the more interesting one. Astra scored highest while spending 4,704 output tokens per task, and GPT-5.2 scored lowest while spending 27,620. On this test the older models were not thinking their way to the answer so much as talking around it, and OpenAI attributes the gain to stronger native multimodal handling rather than to more reasoning.

The three ways Astra still got it wrong

OpenAI published the failures in detail, which is the most useful part of the write-up. On a colony PCR gel, Astra grouped the bands correctly but reversed what the groups meant: assuming the larger product marked successful cloning, it nominated colonies in lanes 4, 6, 7 and 8, where the reference answer reads the controls and selects 5, 9 and 10. In the lab that choice sends the wrong plasmid to purification. On a second task it concluded that vacuoles are generally darker than the surrounding cell, having noticed that contrast varies with focus and then dropped the lighter examples in the same field. On a third it miscounted a plate, merging close neighbours and marking empty agar, for an overcount of about 12 per cent.

The wins are specific in the same way. One cloning task hinged on a negative control plate; GPT-5.6 Sol saw no colonies and drew its conclusion, while Astra read the handwritten label on the plate rim, "LB Kan50", recognised kanamycin rather than the ampicillin used in the experiment, and rejected the plate as a control. Another task showed cells fading in and out of view across a series of microscope frames. GPT-5.6 Sol looked for a biological cause. Astra identified a z-stack.

That caveat is the company's own, and it is the right one, but it does not cover the structural problem with the numbers. OpenAI commissioned the tasks, wrote the rubrics, chose the field of entrants and graded the results. No Anthropic, Google or open-weight model appears in the table, so there is no way to tell from this work whether 66.7 per cent is a frontier figure or a middling one. The confidence intervals are also wide enough that Astra and GPT-6.1 Sol overlap substantially.

What the benchmark does establish, on OpenAI's own evidence, is a floor well below reliability on work the company says wet-lab scientists should complete every time. A model that reads a rim label most humans would miss and then overcounts a plate by an eighth is not yet a second pair of eyes you would leave alone with a cloning experiment. Whether that floor rises because models improve or because the rubrics loosen is the thing worth watching, and it is the thing a vendor-run benchmark cannot settle on its own.