OpenAI Releases a Batch of AI-Produced Math and Asks How Labs Should Disclose It

OpenAI has published a large batch of mathematics produced by an unreleased AI model, along with the compute it used and summaries of the model's reasoning. The release matters as much for how it was disclosed as for what it claims: it is one of the first attempts to set norms for how an AI lab announces scientific results, and many of the claims have not yet been independently checked.

What Was Released

In an Oct. 6, 2026 post, OpenAI says it is releasing "a broad range of new mathematical results produced by an internal frontier model." The results live in a public GitHub repository, openai/math. According to Unite.AI and Interesting Engineering, the repository holds 722 manuscripts grouped into 372 research families across pure mathematics, theoretical computer science, and mathematical physics, published under the Apache-2.0 license with BibTeX citations for each.

Most results were produced with a single fixed procedure on an unreleased model, per Unite.AI, which was given roughly 4,000 problems. OpenAI says the average result used the equivalent of "roughly three hours of ChatGPT Pro thinking." The repository includes statistics on attempted problems and 10 summaries of the model's reasoning; OpenAI's own page does not give the number of attempts or successes.

How the Disclosure Was Handled

OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study (AGMAI) to shape the release, and that it will "update our standards for future disclosures of major scientific advancements." Unite.AI reports the group's Sept. 29, 2026 recommendations, which followed more than 600 replies from the mathematical community. They call for results to be deposited in repositories no lab controls, for public disclosure of the model, prompts, compute, and cost, and for funding of efforts to help humans understand the results, managed by nonprofit institutions. OpenAI says it is still exploring community-hosted alternatives that meet those guidelines. It also says it will fund workshops, conferences, and programs on understanding major AI-produced results.

The company is "working to responsibly release the model that produced these results," with no date given.

What Has and Hasn't Been Checked

OpenAI says many proofs are formalized in Lean, a language that lets a computer check that each step of a proof follows from its assumptions, and that it will add more formalizations as they become available. The caveats are important, and OpenAI's page states few of them. The repository's README says some unformalized results could have issues, according to Unite.AI, and corrections will be logged as new versions with earlier ones kept accessible. An analysis by ExplainX that read the repository found:

• No evidence that any result has been refereed.
• Lean entries for most of the eight headline claims it examined, but some formalize a narrower statement than the headline, and at least one claim (on integer multiplication) had only a preprint.
• A mismatch for two claims between the repository's contents list and its machine-readable catalogue that it could not reconcile.
• Uncertainty over whether the "three hours" figure includes failed attempts.

Even a complete Lean proof only confirms the stated theorem is proved; it does not by itself show that the theorem as formalized is the one the headline describes. Sources also describe the most prominent results inconsistently, so we are not restating specific claims here. Kevin Buzzard of the Xena Project wrote on Oct. 1, before the release, that "you cannot fool Lean" and urged OpenAI to publish its theorems, ExplainX reports; we found no reaction from outside mathematicians to the repository itself that we could verify.

Why It Matters

• Disclosure norms are being written now. OpenAI published reasoning summaries and compute estimates, and the IAS advisory group is pushing for more.
• Verification is the bottleneck. Lean helps, but review capacity and what each formalization actually states are open questions.
• Treat headline claims as provisional. No result has been confirmed as refereed, and some are not formalized.
• The model is still private. No release date, so outsiders cannot reproduce the results.

For how AI labs are being pushed on safety and disclosure more broadly, see our report on the new AI czar and our power-user playbook on building your own evaluations.