OpenAI releases 722 AI-generated math manuscripts

OpenAI posted 722 mathematical manuscripts produced by an internal frontier model to a public GitHub repository, organized into 372 research families, in what is the company's largest single math disclosure to date. Many entries carry computer-checkable Lean proofs, though OpenAI acknowledged that unformalized results may contain errors and will continue to be revised. The release follows an August 28, 2026 announcement that OpenAI was training a model aimed at the Navier-Stokes Millennium Prize problem alongside more than 100 other longstanding open questions.

Corpus size and problem coverage

Multiple outlets confirm the GitHub repository currently holds 722 manuscripts across 372 result families. A secondary newsletter summary from The Rundown AI, cited by Blockchain.News, puts the underlying problem count at roughly 4,000 and reports an average of three hours of ChatGPT Pro compute per result, figures that do not appear in OpenAI's own disclosures and should be read as secondary-source estimates rather than company-confirmed metrics. The August announcement framed the project as targeting the Navier-Stokes challenge and more than 100 open problems, a scope that aligns with but does not independently verify the 4,000-problem figure.

Formal verification status and advisory coordination

OpenAI said the release references recommendations from the Institute for Advanced Study's Mathematics and AI Advisory Group, and spokesperson Lindsay McCallum confirmed the company is consulting that body on responsible disclosure. While many manuscripts ship with Lean proofs, OpenAI explicitly noted that not all results are fully formalized and that unfinalized entries may have issues subject to further updates, a dual posture that mirrors the broader difficulty of scaling formal review to AI-generated output volumes. No source provides a timeline for resolving the unformalized entries; reporting does not disclose one.

Earlier milestone and broader context

The August 2026 GitHub release of solutions to 10 open mathematical problems with Lean 4 proofs preceded the current batch and is described in coverage as a prior milestone in the same disclosure effort. The new 722-manuscript corpus extends that template to a substantially wider problem surface, though characterizations of the earlier drop as a proof-of-concept for the larger release are editorial framing rather than a claim made by OpenAI or the reporting outlets.

Community reaction

Some mathematicians have raised concerns about the pace at which AI firms are generating and publishing proofs, with critics arguing that high-volume AI output risks outpacing peer review and complicating credit norms. The next observable signal is further updates to the GitHub repository as unfinalized manuscripts are formalized or corrected, though the reviewed reports do not specify when those revisions will land.