OpenAI publishes hundreds of AI-generated math manuscripts for scrutiny

OpenAI has published 722 mathematical manuscripts, grouped into 372 related result families, that it says were produced by an unreleased internal model while it was tested on roughly 4,000 open research problems. The release matters less as a stack of claimed breakthroughs than as an unusually detailed attempt to make AI-generated research inspectable: it includes papers, revision rules, attempted-problem statistics, and brief reasoning summaries for 10 results. Many proofs have also been translated into Lean, software that can mechanically check whether a formal proof follows its stated rules, but OpenAI says not every manuscript has that protection and unformalized work may contain errors. Its estimate of about three hours of ChatGPT Pro-level reasoning compute per result offers a rare sense of the resources involved, while leaving mathematicians with the harder task: deciding which results are important and withstand human review.