OpenAI unveils BioBasicsBench, a benchmark to gauge how well AI agents read raw experimental data across 100 routine wet-lab tasks.

DNA cloning underpins work from studying genes to engineering proteins, and recovering the intended DNA often requires painstaking screening. Choosing the wrong bacterial colony, for instance, is one pitfall that can cost a researcher a day spent culturing cells and purifying a plasmid the lab already had. In one Yale study, colonies retaining the original plasmid glowed green, while those carrying the replacement DNA appeared white. The researchers checked the plates with a blue LED flashlight and filtered glasses. In their typical workflow, up to half the colonies retained the original plasmid.

In a test OpenAI published Nov. 6, its top-scoring model encountered a similar sorting problem, this time using an agarose gel, which separates DNA fragments by size into visible bands. It distinguished the band patterns correctly, then reversed the interpretation supported by the controls and recommended colonies still carrying the original plasmid, a small circular DNA molecule used to carry genetic material into cells. That example comes from BioBasicsBench, a new evaluation OpenAI built with enzyme-design startup Ligo Biosciences. It asks models to interpret raw data from 100 routine lab tasks, catch experimental errors and recognize when the evidence is too ambiguous to support a conclusion.

BioBasicsBench tests how well AI models interpret the messy evidence scientists encounter at the bench. “Everyday wet-lab work involves failed experiments, ambiguous images and conflicting observations,” OpenAI wrote in an announcement. Charts and images selected for publication can leave much of that uncertainty out of view. The benchmark uses raw experimental data to test whether models can read a result, catch a flawed control or recognize when the evidence is insufficient to draw a conclusion.

To build the benchmark, scientists turned data from their own experiments into 100 test questions for AI models, each paired with detailed scoring criteria. Every question underwent three reviews: one by another Ligo scientist, one by a PhD-level scientist at an outside wet lab and a final internal review. Models earned points for elements such as accurate measurements, sound reasoning and recognizing uncertainty. The reported scores represent the average percentage of available points earned across the tasks.

In the scope of testing, OpenAI tested seven of its own models from two generations: GPT-5.2, GPT-5.4, GPT-5.5 and GPT-5.6 Sol from the GPT-5 line, and its more recent models GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra. GPT-5.6 introduced the Sol, Terra and Luna tier names, and GPT-6 added Astra, released in early September, as the top tier. Each model ran at several reasoning-effort settings, which control how much the model works through a problem before it answers. In all, OpenAI ran 37 model-and-setting combinations, with five attempts per task, for 18,500 graded answers. The post compares only OpenAI models.

The benchmark is an example of a growing focus on science at OpenAI. In a Nov. 3 podcast interview, CEO Sam Altman said the company’s recent math results, including its claimed Navier–Stokes proof, “really made me think that we are closer to this milestone than I expected us to be,” referring to a machine for scientific discovery. The same week, OpenAI said Astra designed the enclosure, circuit board and firmware for a talking Codex desk device, and an MIT team reported that Astra could improve other AI models by editing their weights, a result critics said relied on standard machine-learning engineering.

At “xhigh,” the highest setting run on all seven models, Astra earned 66.7% of the available credit while using about half the output tokens of GPT-5.6 Sol. OpenAI says stronger native multimodal capabilities “may be key” to the improvement.

In one cloning task, GPT-5.6 Sol based its conclusions on an empty negative control plate, the parallel sample that tells researchers whether something other than the intended procedure could produce a result. Astra read the handwritten “LB Kan50” label on the plate’s rim, saw that the plate contained kanamycin while the experiment used ampicillin, and rejected it as a valid control.

LLMs exhibit uneven performance in wet lab settings at present, and the failures OpenAI highlights often pair a correct observation with a wrong call. In the colony PCR task from the opening, Astra sorted the colony lanes into the right two groups, then assumed the larger PCR product meant successful cloning and picked the colonies that still carried the old plasmid. It also overcounted one plate’s colonies by about 12%, enough to inflate an estimate of how efficiently DNA entered the cells by the same proportion. “Higher benchmark scores do not yet mean models can make reliable judgments in the lab,” OpenAI wrote.

Benchling, which sells laboratory data software, has also built its tests of AI’s lab reasoning, including one on planning cloning experiments in which Astra also scored highest. “LLMs can work in the wet lab with the right harness and a human in the loop,” Nicholas Larus-Stone, the company’s head of AI, told R&D World in August.