OpenAI published 722 AI-generated mathematical manuscripts on October 6, 2026, in a public GitHub repository. The collection groups them into 372 families of related results. But the number is not a count of independently verified proofs. OpenAI says many, but not all, of the manuscripts have Lean formalizations, and results without them may contain issues.

Different stages of checking

Lean is a programming language used to express mathematical proofs in a form a computer can check. A Lean formalization checks a formal proof against its stated assumptions. It is not the same as independent expert review of every manuscript.

The repository describes the work as being at different stages of verification. OpenAI says it will add formalizations as they become available, and that some results without them may have problems. The company plans to record corrections and revisions in new versions while keeping earlier versions accessible.

From about 4,000 problems to 722 manuscripts

OpenAI says it gave an unreleased internal model approximately 4,000 problems during an evaluation. It then grouped the outputs into related result families and manuscripts, selecting work it considered significant enough to include. The 722 manuscripts do not represent 722 separate problems or 722 confirmed breakthroughs.

OpenAI estimates that an average result used computing equivalent to about three hours of ChatGPT Pro thinking. That is an estimate of computing effort, not a claim that each manuscript took three hours to write. The model remains unreleased, so researchers can inspect the collection but cannot run that model themselves.

What happens next

OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study about sharing the work. The GitHub repository includes the manuscripts and supporting materials, along with procedures for citations, corrections and revisions. OpenAI also says it plans to fund workshops and other programs to help researchers examine major AI-produced results.

Mathematicians now have to determine which arguments hold up, which need revision, and what the formalized proofs establish. The public release makes the work available for scrutiny, but it does not finish that scrutiny.

There are 722 manuscripts to examine, grouped into 372 result families. The repository’s own caveat is central: verification is still incomplete.