Evaluating AI's ability to perform scientific research tasks (openai.com)

Recent advancements in AI have prompted a deeper exploration of models' capabilities in scientific research, focusing on their reasoning skills. Notably, the release of GPT-5 has shown significant promise, as it achieved a gold-medal performance in prestigious competitions like the International Math Olympiad. Research indicates that GPT-5 can meaningfully expedite scientific workflows, cutting research time from days to hours by assisting in literature searches and complex mathematical proofs. This is documented in the paper "Early Science Acceleration Experiments with GPT-5," which highlights the potential of AI to accelerate scientific progress. To better measure AI's performance in scientific reasoning, the new FrontierScience benchmark has been introduced, targeting expert-level capabilities across physics, chemistry, and biology with over 700 expertly crafted questions. Initial evaluations reveal that GPT-5.2 outperforms its predecessors in both structured reasoning (scoring 77% on Olympiad questions) and expert research tasks (25% on Research questions). Although these advancements illustrate significant progress, they also underscore the limitations in AI's ability to perform open-ended scientific tasks. The development of FrontierScience not only sets a standard for evaluating AI in scientific contexts but also highlights the continued need for nuanced benchmarks that capture the breadth of scientific research capabilities and the importance of human intervention in validating findings.