OpenAI and Ligo publish BioBasicsBench for routine wet-lab analysis

OpenAI partnered with enzyme-design company Ligo to build BioBasicsBench, a benchmark of 100 tasks drawn from scientists' actual lab work. The tasks cover protein methods, imaging and cytometry, assays, nucleic acids, and cloning, the kind of routine analysis that fills a biologist's workday. OpenAI tested seven models across 37 different settings, running five attempts per task, producing 18,500 total results. The standout was GPT-6 Astra at its highest reasoning setting, which averaged 66.7 percent of available credit, with a 95 percent confidence interval of 60.3 to 73.1 percent, while using 4,704 output tokens per task. That compares to 43.8 percent for GPT-5.6 Sol, which needed far more tokens, 9,099 per task, and 62.0 percent for GPT-6.1 Sol. The benchmark is meant to measure how well AI models can handle the basic, repetitive scientific analysis that underlies lab work, rather than flashy discovery claims.

Why it matters

It gives researchers a grounded, real-world way to measure whether AI models can actually help with everyday lab work, not just headline-grabbing science problems.

Source: OpenAI