OpenAI Publishes 722 Math Manuscripts From an Internal Model

OpenAI has published math results generated by an unreleased frontier model, packaging preprints, Lean formalizations, abridged reasoning summaries, and compute estimates in a public GitHub repository. The artifacts give mathematicians and developers material they can inspect independently. No API, model weights, or inference access accompanies the release.

The corpus extends OpenAI’s September Navier-Stokes claim, which proposed that smooth solutions to the three-dimensional fluid equations can develop a singularity in finite time. That problem is tied to one of the Clay Mathematics Institute’s Millennium Prize Problems. OpenAI described the internal model behind the proof as significantly more capable than GPT-6 Astra and says the same system produced the broader catalogue.

A Catalogue Built Around Paper Families

At publication, the catalogue contains 722 manuscripts grouped into 372 families. A family can include a principal result, companion arguments, consequences, or alternative proofs. Discipline tags and an overview PDF provide routes from subject-level indexes to individual papers and their supporting files.

The repository’s three main directories separate papers, formal proofs, and selected reasoning summaries, allowing reviewers to move from a manuscript to any linked verification artifact:

• preprints/ contains PDFs, LaTeX sources, and citation metadata for each manuscript.
• lean/ contains machine-checkable proofs and a formalization.yaml catalogue that maps formalizations to papers.
• reasoning_traces/ contains abridged accounts of the model’s approach to 10 results.

Formal coverage remains incomplete, and OpenAI says it will add Lean proofs as they become available. Results without formalizations still require expert review of the mathematical argument.

One Pipeline, Two Exceptions

OpenAI says it gave the model roughly 4,000 problems and used a largely fixed generation process. Related outputs were grouped into families and filtered for mathematical significance. The company reports an average allocation equivalent to about three hours of ChatGPT Pro “thinking” compute per result. That estimate omits the run-level prompts, hardware, sampling settings, and timing data needed to reproduce the generation process.

OpenAI identifies two exceptions to the fixed pipeline: a zero-free region for the Riemann zeta function and a proof of the Hodge conjecture for CM, or complex multiplication, abelian varieties. Humans lightly edited the zeta-function manuscript for readability. OpenAI says the remaining manuscripts retain the model-generated text.

The 10 reasoning summaries span the irrationality exponent of pi, Kaplansky’s direct-finiteness conjecture in characteristic two, the Mézard-Parisi formula for diluted spin glasses, and quasipolynomial bounds for arithmetic progressions. These are specialist research problems commonly addressed in journal papers. Because the summaries are abridged, they reveal selected strategies without providing complete execution traces or sufficient data to replay a run.

What Lean Actually Verifies

Lean checks whether a formal proof term has the claimed type under the project’s imported definitions, axioms, and dependencies. A successful build establishes that the encoded theorem follows within that formal environment, as checked by Lean’s kernel.

Mechanical verification leaves several questions for human reviewers. They must confirm that the formal statement matches the theorem claimed in the paper, that the definitions capture the intended concepts, that the assumptions are acceptable, and that the result is novel and significant.

OpenAI’s earlier Navier-Stokes announcement drew questions from Tristan Buckmaster, a mathematics professor at New York University, after the company said 10,000 agents produced the result in 88 hours. For the formalized portion of this release, researchers can download the Lean files, rebuild them, inspect their assumptions, and compare the encoded statements with the manuscripts.

Governance After Navier-Stokes

OpenAI is routing review and communication through an advisory group at the Institute for Advanced Study, formed after criticism surrounding the Navier-Stokes announcement. Its remit covers significance assessments, review practices, dissemination, and the use of AI tools in mathematical research and education. Internal decisions about the pace of capability development sit outside its remit.

The repository preserves its public release history and uses versioned citation procedures, so corrections can appear as new versions instead of silent edits. Each manuscript directory includes BibTeX metadata for version-specific citation, and the repository uses the Apache 2.0 license.

A Practical Review Path

1. Identify the paper and version. Use the discipline tags, family structure, and manuscript metadata to locate the precise claim and its related papers.
2. Check for a formalization. Follow the mapping in formalization.yaml to determine whether the manuscript has a corresponding Lean artifact.
3. Rebuild the proof. Use the repository’s documented Lean environment and run any comparator checks included with the formalization.
4. Inspect the formal statement. Compare the Lean theorem, assumptions, definitions, imported axioms, and dependencies with the prose claim in the preprint.
5. Review unformalized work conventionally. Examine each proof step, cited result, edge case, and novelty claim through ordinary expert review.
6. Use summaries within their limits. The reasoning files can inform agent design and evaluation work, but their abridged form cannot reproduce the model’s complete trajectory.
7. Cite a fixed version. Use the manuscript’s BibTeX block and record the repository version because later corrections may change the text or formal artifacts.

Evidence Will Set the Value

OpenAI says the internal system has resolved more than 100 long-standing open problems across many areas of mathematics. Those claims now face two validation routes: mechanical checking for the encoded subset and expert peer review for the remaining manuscripts.

Hundreds of papers create a substantial review workload, even when some proofs compile in Lean. If a meaningful share survives formal inspection and subject-matter review, model-generated work could increase the rate at which candidate results enter mathematical literature. Error rates, novelty assessments, correction history, and the durability of the proofs will determine the release’s scientific value.

The repository also provides a concrete publication protocol for AI-generated research: versioned preprints, explicit citations, partial formal verification, limited reasoning summaries, and external advice on review and communication. Its record of verification and correction will show whether that protocol scales beyond this release.