On this page

OpenAI and enzyme-design biotech company Ligo have introduced BioBasicsBench, a benchmark built from 100 tasks involving raw experimental data. OpenAI’s article is dated November 6, 2026. It asks whether AI agents can interpret evidence, detect errors and recognize when a conclusion is not justified—not whether they can carry out experiments autonomously.

What the benchmark measures

The tasks use evidence such as gel images, microscopy and instrument readouts. They span five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Scientists paired questions with expert-written grading rubrics and reviewed the tasks internally and externally.

A response earns credit against individual rubric criteria, including conclusions, numerical results, reasoning and recognition of insufficient evidence. A partly sound analysis can therefore receive partial credit. Data-type and skill categories overlap; adding their counts does not produce the task total.

What 66.7% means

At xhigh reasoning effort, OpenAI reports 66.7% mean reward for GPT-6 Astra, with a 95% confidence interval of 60.3–73.1%. GPT-5.6 Sol scored 43.8%, with a 37.6–50.0% interval.

Mean reward is the average task score expressed as a share of available credit. It is not the percentage of tasks completed successfully. The full evaluation covers seven models and 37 model–reasoning settings, each with five attempts on all 100 tasks: 18,500 results, not 18,500 separate tasks.

Higher scores still leave reliability gaps

OpenAI’s examples include assumptions overriding controls, contradictory observations being discounted and colonies being miscounted. Such errors explain why a higher benchmark score does not establish reliable judgment in a working laboratory. The study does not establish that AI can replace experimental scientists.

The separate Orchestra Bio acquisition and Rosalind story provides context on OpenAI’s biology work; it is not this benchmark announcement.

OpenAI reports these evaluation results; we haven’t independently reproduced them.