OpenAI Launches 100-Task BioBasicsBench as GPT-6 Astra Tops GPT-5.6 Sol 66.7% to 43.8%
Updated
Updated · OpenAI · Nov 6
OpenAI Launches 100-Task BioBasicsBench as GPT-6 Astra Tops GPT-5.6 Sol 66.7% to 43.8%
1 articles · Updated · OpenAI · Nov 6
Summary
GPT-6 Astra scored 66.7% on OpenAI’s new BioBasicsBench at xhigh reasoning effort, versus 43.8% for GPT-5.6 Sol, in a benchmark aimed at routine wet-lab analysis.
BioBasicsBench measures whether AI agents can interpret raw lab evidence, catch errors and recognize when data cannot support a conclusion—using 100 tasks built with biotech company Ligo.
OpenAI said Astra’s gains were strongest in multimodal handling, control-checking and keeping experimental context in mind, such as spotting ambiguous dye signals or mislabeled control plates.
The benchmark still exposed major weaknesses: Astra sometimes overrode controls with assumptions, discounted contradictory observations and overcounted colonies by about 12%.
OpenAI framed the release as a step toward more reliable AI lab assistance and a way to track biosafety and biosecurity implications as models improve.