OpenAI posts 722 model-generated math manuscripts

OpenAI released 722 mathematical manuscripts from an internal model, organized into 372 result families, while warning that results sit at different verification stages and not all have Lean formalizations. The public repository lists 10 reasoning summaries and says roughly 4,000 problems fed the collection; its average result used about three hours of ChatGPT Pro thinking compute. Research teams should treat the corpus as inspectable candidates, not 722 independently certified results; our earlier look at Claude’s math work makes the human reproducibility gate concrete.