OpenAI shares mathematical results and Lean formalizations from an internal frontier model

OpenAI has announced a collection of mathematical results produced by an internal frontier model, with supporting material in a GitHub repository. The release includes formalizations of many proofs in Lean, ten summaries of model reasoning and information about compute use and attempted problems.

The October 6 announcement is a disclosure of research results and supporting artifacts. OpenAI says it is working to responsibly release the model that produced them. It does not announce that the internal model is already available to researchers or identify it as a model people can select in ChatGPT.

Proof artifacts accompany the results

Lean is a programming language that allows a computer to check mathematical proofs. OpenAI says its repository contains formalizations of many of the proofs and that it will add further formalizations as it obtains them. The wording matters: the announcement does not say that every result in the collection already has a Lean formalization.

For a reader examining a particular result, the useful starting point is its own proof and any accompanying formalization. A release-wide statement about many proofs should not be substituted for checking which artifacts accompany the specific mathematical claim being discussed. The available formalization and the paper's exposition answer related but different inspection needs.

OpenAI also describes protocols for paper revisions and citations. Those provide a way to follow changes to an individual paper rather than assuming its initial version will remain the final account. The company says it wants to improve citations, mathematical exposition and presentation in future releases.

Compute estimates provide context for the work

The disclosure includes statistics on the number of attempted problems and ten summaries of the model's reasoning. OpenAI says the average result used compute equivalent to roughly three hours of ChatGPT Pro thinking. It presents that as an estimate expressed through a familiar usage reference, not a promise that a ChatGPT Pro session can reproduce a selected result in three hours.

The average also does not describe the distribution of work across individual results. Readers evaluating efficiency should look at the repository's result-specific information where available, alongside the attempts that preceded a successful result. Counting only a finished proof would leave out part of the search process that the company says it is disclosing.

Reasoning summaries can help explain how OpenAI obtained the work, but a summary is not itself a substitute for reviewing a proof. The announcement provides ten such summaries; it does not claim to publish a complete trace for every attempt or result.

OpenAI consulted an independent advisory group

OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study while developing its approach to sharing the results. It drew on the group's advice and public recommendations for the release.

Consultation should remain distinct from a claim that the advisory group verified or endorsed every mathematical result. The announcement describes advice about disclosure practices, and it says OpenAI continues to examine community-hosted alternatives that meet the committee's guidelines. It does not describe a completed independent review of the entire collection.

The model release and further programs are still ahead

OpenAI plans to fund workshops, conferences and special programs focused on understanding major AI-produced results, with more details to come. It also says it will continue evaluating internal frontier models in mathematics and other sciences while working on responsible access.

The present release gives the mathematical community artifacts to inspect and a stated process for revisions. Assessments of originality, correctness and usefulness should refer to the relevant paper and proof rather than infer those properties from the model's frontier label. Franklin has not independently checked the published proofs or reproduced the model's work.