New benchmark tests AI agents on complex, real-world wet-lab data analysis

A new evaluation tool, BioBasicsBench, has been introduced to assess how effectively artificial intelligence agents can interpret the raw, often ambiguous data encountered in routine biological laboratory work. Developed in partnership with Ligo, a biotech firm specializing in enzyme design, the benchmark comprises 100 tasks covering five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. The benchmark tests agents on interpreting raw evidence—such as gel images, microscopy stacks, and plate reader exports—and recognizing when data is insufficient to draw a conclusion. Testing seven models, including GPT-6 Astra and GPT-5.6 Sol, the evaluation involved 18,500 results across 37 model and reasoning-effort settings. At the highest reasoning setting, GPT-6 Astra achieved an average score of 66.7%, significantly surpassing GPT-5.6 Sol's 43.8%. The results highlight that advanced models like Astra show improved capabilities in handling multimodal data and discerning experimental context. For instance, Astra correctly identified an error in a DNA cloning control plate due to an incorrect antibiotic label, whereas GPT-5.6 Sol missed this detail. Furthermore, Astra demonstrated an improved ability to recognize data limitations, such as when image contrast can be reversed by changing focus, preventing unwarranted conclusions.