# OpenAI unveils BioBasicsBench, a benchmark testing how well AI interprets messy wet-lab data > OpenAI's BioBasicsBench tests AI on 100 raw wet-lab tasks, and top model GPT-6 Astra scored 66.7% while still overcounting colonies by 12% and recommending the wrong clones. - Type: AI news - Page: https://completeaitraining.com/news/openai-unveils-biobasicsbench-a-benchmark-testing-how-well/ - Published: 2026-11-08 - Updated: 2026-11-08 - Author: Jeroen Erne - Jobs: Education, Healthcare, IT and development, Product development, Science and research - Reading time: 3 min OpenAI released BioBasicsBench on Nov. 6, a benchmark built with enzyme-design startup Ligo Biosciences that tests how well AI models read raw experimental data from 100 routine wet-lab tasks. The evaluation pushes models to interpret messy, real-world lab evidence-failed experiments, ambiguous images, conflicting observations-rather than the polished charts that make it into published papers. #### What the benchmark actually tests Scientists turned data from their own experiments into 100 test questions, each paired with detailed scoring criteria. Every question went through 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 accurate measurements, sound reasoning, and recognizing uncertainty. The reported scores represent the average percentage of available points earned across the tasks. The tasks cover ground familiar to anyone who has spent time at a lab bench. DNA cloning, for instance, underpins work from studying genes to engineering proteins, and recovering the intended DNA often requires painstaking screening. Choosing the wrong bacterial colony 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-and up to half the colonies retained the original plasmid in a typical workflow. #### Where models stumbled on real lab data OpenAI tested seven of its own models across two generations, running 37 model-and-setting combinations with five attempts per task, for 18,500 graded answers. At the highest reasoning-effort setting, GPT-6 Astra earned 66.7% of the available credit while using about half the output tokens of GPT-5.6 Sol. OpenAI said stronger native multimodal capabilities "may be key" to the improvement. Even the top-scoring model made errors that would derail real lab work. In a colony PCR task, Astra sorted colony lanes into the correct two groups, then assumed the larger PCR product meant successful cloning and picked 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. In another cloning task, GPT-5.6 Sol based its conclusions on an empty negative control plate. 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-a detail a rushed graduate student might miss. #### OpenAI's growing science push The benchmark arrives amid a broader science focus at OpenAI. In a Nov. 3 podcast interview, CEO Sam Altman said the company's recent math results "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. Benchling, which sells laboratory data software, has also built tests of AI's lab reasoning, including one on planning cloning experiments where 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. #### Why this matters for science and research professionals BioBasicsBench exposes a specific gap: models can read a gel correctly and still recommend the wrong colonies. For researchers evaluating AI tools for lab work, the takeaway is not that these systems are useless-it's that their errors often pair a correct observation with a wrong call. A model that overcounts colonies by 12% will inflate transformation efficiency estimates by the same margin. Until benchmarks like this one show models catching their own flawed controls and recognizing ambiguous evidence, the human in the loop remains the one making the final call at the bench. Professionals looking to understand where AI fits into laboratory workflows can explore [AI for Scientists Courses](https://completeaitraining.com/tag/science-and-research/) and [AI for Biochemists Training](https://completeaitraining.com/learningpath/ai-for-biochemists/) to build the evaluation skills these benchmarks demand. --- This is the Markdown form of https://completeaitraining.com/news/openai-unveils-biobasicsbench-a-benchmark-testing-how-well/. The whole site is mapped at https://completeaitraining.com/llms.txt.