UniPat AI

UniPat AI, together with Stanford, Harvard, Cornell and over ten other institutions, released Science Prediction Universe (SPU), a benchmark of 116 tasks and 4,207 real wet-lab experiments that tests whether AI can predict experimental outcomes before they are run. Models were given an H20 GPU and an 8-hour workday to autonomously use retrieval and compute resources, yet even Opus 5.5 and GPT-6 Astra remained far from usable predictive accuracy, and a 5x compute increase did not help. The benchmark turns hard-to-quantify scientific judgment into a measurable, comparable score.