OpenAI's collection includes Lean proof artifacts and reasoning summaries. OpenAI says verification varies across the papers and that citations and exposition need further work.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [X](https://x.com/OpenAI/status/2107596713791767021)
Why it matters #
OpenAI is moving AI mathematics from benchmark claims to a public research corpus. The papers and some formal proofs are available for inspection, but verification varies by paper and the model remains proprietary. Mathematicians still need to assess and explain the results before they can be used as research.
OpenAI published 722 mathematical manuscripts produced by an unreleased internal model on October 6th, making public a large batch of results while warning that some proofs have not been formally checked. The papers, organized into 372 families, are available in a GitHub repository alongside selected reasoning summaries and supporting materials.
The collection spans subjects including number theory, combinatorics, theoretical computer science and mathematical physics. The repository says OpenAI posed approximately 4,000 problems to the model, then grouped its output into results it judged significant enough to include. That selection was made by OpenAI; the announcement does not describe an independent assessment of the full collection's significance or correctness.
Verification is uneven. The repository includes Lean formalizations for many results, allowing a computer to check the formal proof, but OpenAI says some manuscripts do not have them and warns that unformalized results could contain issues. Lean artifacts enable a check where available, but they do not verify every manuscript in the release. OpenAI says it will add formalizations as they become available.
OpenAI also released 10 summaries of the model's reasoning, estimates of compute use and statistics on attempted problems. The company says the average result used compute equivalent to about three hours of ChatGPT Pro reasoning. That estimate does not measure the human effort required to understand, verify or build on a result. Most work followed the same procedure, according to the repository, with exceptions including work on the Riemann zeta function and the Hodge Conjecture for CM abelian varieties. One write-up about a zero-free region for the Riemann zeta function was edited by a human for readability.
The release follows a public debate over how AI-generated mathematics should enter a discipline built around authors who understand and take responsibility for their proofs. An independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, published recommendations on September 29th after receiving more than 600 survey responses. The group says it operates independently, its members do not accept payment for the work, and it has no decision-making power over OpenAI.
The group described its recommendations as "general guidelines for the responsible release of AI-generated mathematics by AI labs," informed by more than 600 survey responses.
OpenAI's release includes selected reasoning summaries, attempted-problem statistics, compute estimates and a revision policy that preserves earlier versions. The GitHub repository is controlled by OpenAI, and OpenAI says it is still exploring community-hosted alternatives. The public materials pair summaries with the repository's model-produced papers, but do not name the unreleased model or publish prompts for each result. OpenAI says future releases should improve citations, mathematical exposition and presentation.
Machine-checkable proofs and human understanding answer different questions. Formalization can establish that a proof follows within a specified system; it does not by itself explain why a result matters, how it relates to prior work or what new lines of inquiry it opens. OpenAI says it will fund workshops, conferences and special programs around major AI-generated results, but gave no funding amount or timetable in its announcement.
OpenAI has not released the generating system. OpenAI says it is working toward a responsible release of the model, without providing a date or access terms. For now, mathematicians can inspect the papers and available proof artifacts, but the model and the process that generated most individual results remain unavailable for independent replication.