OpenAI has unveiled BioBasicsBench, a new benchmark designed to measure the ability of AI agents to interpret raw experimental data and navigate the complexities of everyday wet-lab science. Announced on November 4, 2026, the benchmark was developed in collaboration with Ligo, a biotechnology company specializing in enzyme design.

Addressing Gaps in AI Evaluation

BioBasicsBench consists of 100 tasks that span five key areas of laboratory work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Unlike prior benchmarks that rely on publication-selected charts and images, BioBasicsBench incorporates raw experimental evidence such as gel images, colony plates, microscopy data, and instrument readouts. This approach aims to reflect the real-world challenges of wet-lab science, including failed experiments, ambiguous results, and troubleshooting scenarios.

To ensure rigor, scientists from Ligo constructed tasks based on their own experiments and paired them with expert-written grading rubrics. Each task underwent a three-stage review process involving a Ligo scientist, an external PhD-level wet-lab scientist, and an internal reviewer.

Performance Results of AI Models

OpenAI evaluated seven models—GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 Sol, GPT-6.1 Sol, GPT-6 Luna, and GPT-6 Astra—across 37 model/reasoning-effort configurations. Each configuration attempted five runs per task, generating a total of 18,500 results. At the highest reasoning setting ('xhigh'), GPT-6 Astra achieved a mean reward of 66.7% (95% confidence interval: 60.3–73.1%), outperforming GPT-6.1 Sol at 62.0% (55.6–68.4%) and GPT-5.6 Sol at 43.8% (37.6–50.0%). Lower-performing models included GPT-6 Luna at 38.3% (32.3–44.5%) and GPT-5.5 at 35.4% (29.7–41.0%).

Implications for Scientific Research

The benchmark emphasizes the importance of AI agents identifying experimental errors and recognizing when evidence is insufficient to support a conclusion. Misinterpretations of raw data, such as gel images or control problems, can lead to incorrect conclusions and wasted resources in follow-up experiments. By focusing on these challenges, BioBasicsBench provides a critical tool for assessing and improving the reliability of AI in scientific research.