OpenAI GPT-6 Astra Scores 66.7% on New Wet-Lab Benchmark

OpenAI introduced BioBasicsBench on Nov. 6, 2026, a 100-task benchmark built with biotech company Ligo to measure how well AI agents interpret raw wet-lab evidence. The tasks use gels, colony plates, microscopy, sequencing files, and instrument exports rather than cleaned figures from papers. OpenAI’s stated goal is to help scientists accelerate biological research, from understanding disease to developing treatments, by testing whether models can handle the messy evidence people actually see at the bench.

The headline scores

At the xhigh reasoning setting, which allows more reasoning per task, GPT-6 Astra earned an average of 66.7% of available credit, with a 95% confidence interval of 60.3% to 73.1%. GPT-5.6 Sol scored 43.8% on the same rubric, with a confidence interval of 37.6% to 50.0%. OpenAI evaluated seven models across 37 model and reasoning-effort settings, with five attempts on each of the 100 tasks, for a total of 18,500 graded results. GPT-6.1 Sol landed in between at 62.0%, while older GPT-5.x models clustered in the low-to-mid 30s.

What the bench covers

Tasks span five routine areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Agents must interpret raw data, spot errors, and recognize when the evidence cannot support a conclusion. Each task has expert-written grading criteria and went through an internal Ligo review, an external review by a PhD-level wet-lab scientist, and a final internal review. Inputs include agarose and SDS-PAGE gels, Western blots, plate photographs, microscopy stacks, flow-cytometry files, plate-reader exports, sequencing reads, and short protocol videos.

Where models still fail

OpenAI says higher scores do not yet mean reliable lab judgment. Published examples show models still misreading images, overlooking bad controls, counting colonies incorrectly, and letting assumptions override the data. In one case GPT-5.6 Sol treated a negative-control plate as trustworthy when the rim label showed the wrong antibiotic; Astra caught the mismatch. In other cases Astra still over-selected colonies from a PCR gel or overcounted colonies by about 12%. The company frames BioBasicsBench as a target for better everyday scientific assistance and a way to track biosafety and biosecurity implications as those skills improve.