A research release with inspectable artifacts
OpenAI says an internal frontier model produced a broad set of new mathematical results and has released them in a GitHub repository. The package includes papers, protocols for revisions and citations, statistics on attempted problems, and ten summaries of the model’s reasoning. The company says the average result used compute equivalent to roughly three hours of ChatGPT Pro thinking.
Many of the proofs have Lean formalizations. Lean is a proof assistant that lets a computer check whether a formal proof follows from its definitions and rules. “Many” matters: OpenAI says additional formalizations will be added as they are completed, so the repository is not yet a claim that every result has been mechanically verified.
What Lean does and does not establish
A formalization can check the proof as encoded, making a mathematical argument more inspectable than a prose-only claim. It does not automatically settle whether the formal statement captures the intended theorem, whether the exposition is clear, or whether the result is important. Those judgments still benefit from mathematicians and independent review.
OpenAI says it is consulting an independent advisory group at the Institute for Advanced Study and is publishing the repository with revision and citation protocols. The company also acknowledges it wants to improve citation quality, exposition and presentation. That transparency is useful because readers can distinguish a proposed result from a result already widely vetted.
A different release format for AI science
The practical change is access to the work itself. Researchers can examine individual manuscripts, follow revisions, inspect available Lean files and see how much problem-solving effort OpenAI reports. The repository gives the community concrete objects to review rather than asking it to accept a single headline about AI solving mathematics.
The model remains internal. OpenAI says it is working toward a responsible release, but this post does not make the model weights or an API available. The contribution today is the public research set and supporting evidence, with broader mathematical evaluation still to come.
Source published October 6, 2026. Coverage is based on the maker’s announcement and demonstration.