OpenAI's 722-manuscript maths release leaves mathematicians racing to verify it

OpenAI published 722 mathematics manuscripts produced by an unreleased internal frontier model to its public "openai/math" GitHub repository on 6 October 2026, and more than three dozen mathematicians told The Verge, in a report published 9 October, that making sense of the release could take years.

What was released

• OpenAI said in its announcement that the repository "currently contains 722 manuscripts organized into 372 related result families," including papers, proof artifacts, Lean formal-verification code for many results, and 10 abridged summaries of the model's reasoning.
• The results came from an evaluation involving "approximately 4,000 problems," with an average compute cost per result equivalent to roughly three hours of ChatGPT Pro "thinking" time, according to OpenAI's post.
• The Verge reported that OpenAI "said only 300 top-line results out of 719 manuscripts had been formalized, around 42 percent," and that the company "will update the repository with more formalizations as we obtain them."
• More than three dozen mathematicians described the release to The Verge using words including "staggering," "overwhelming," "unprecedented" and "pure insanity."
• "Just going over the entire list of abstracts is overwhelming," Álvaro Lozano-Robledo, a mathematics professor at the University of Connecticut, told The Verge.
• Imperial College London mathematician Kevin Buzzard said he found only around six theorems in his field, algebraic number theory, that "stood out," and that few of those had been formally verified, leaving him to choose between reading what he called "possibly-not-correct slop" or waiting for others to check it, he told The Verge.

Why it matters

The release shows a frontier AI lab now producing scientific output at a volume that outpaces the scientific community's existing tools for peer review and verification, with OpenAI itself acknowledging on GitHub that the manuscripts are "at different stages of verification" and that "many, but not all" have been formalised, The Verge reported. For universities and research institutes across the Caspian region exploring AI research tools, the episode is a reminder that an AI system's output volume is not evidence of its reliability, and that claims should be treated as unverified until independently formalised or peer-reviewed, regardless of how the underlying model was marketed.