OpenAI's GPT-5 Models Tested on BioBasicsBench, Showing Varied Performance in Scientific Data Interpretation
OpenAI has introduced BioBasicsBench, a new benchmark designed to evaluate how well AI models interpret raw experimental data from 100 routine wet-lab tasks. This benchmark aims to assess AI agents' ability to read results, identify flawed controls, and recognize when evidence is insufficient for conclusions, mirroring the complexities of real-world scientific work. OpenAI tested seven of its own models across two generations: GPT-5.2, GPT-5.4, GPT-5.5, and GPT-5.6 Sol from the GPT-5 line, and GPT-6.1 Sol, GPT-6 Luna, and GPT-6 Astra from the GPT-6 line. Models were run at various reasoning-effort settings, with 37 model-and-setting combinations and five attempts per task, totaling 18,500 graded answers. At the 'xhigh' setting, GPT-5.6 Sol achieved a score of 43.8%, while the newer GPT-6 Astra performed significantly better at 66.7%. The tests revealed that while models can make correct observations, they sometimes pair them with incorrect conclusions, highlighting current limitations in reliable judgment ...
