OpenAI Posts 722 AI Math Proofs on GitHub
An unreleased model produced 722 manuscripts covering 372 result families. Formal checking is automated; judging whether the results matter is not.
In short
OpenAI has published 722 machine-generated mathematics manuscripts on GitHub, grouped into 372 result families, which are formally checkable in Lean but have not been confirmed by any independent expert review.
At a glance
- 722 manuscripts are public on GitHub, grouped into 372 result families.
- Fields covered: number theory, algebraic geometry, theoretical computer science.
- Formal checking runs through the Lean proof language, with revision logs in the repository.
- Per t3n, about three hours of compute per result on standard resources.
- The model that wrote them is unreleased and is not identified by a version name.
OpenAI has opened a public repository holding 722 machine-generated mathematics manuscripts, grouped into 372 result families. The subject areas are number theory, algebraic geometry and theoretical computer science. The author was a language model the company has not released. What exists now is a corpus large enough that no single reviewer can clear it.
What the repository actually contains
The proofs are written in Lean, a language built so that a machine can check each step of an argument instead of a reader trusting it. Revision logs accompany the files and trace how individual results came together. According to t3n, a single result took roughly three hours of compute on the company's standard resources.
Machine-checkable and mathematically interesting are different tests
Lean settles one question: does the argument hold together. It says nothing about whether a proven statement is new, whether anyone was looking for it, or whether it lands on a problem the field cares about. That second judgment is human work, and it has not happened yet.
Some results are said to touch a Millennium Prize Problem. That framing comes from the release itself rather than from a referee report. Reading 722 files as 722 solved open problems converts a file count into a claim about knowledge.
Who gets the credit
The Advisory Group on Mathematics and Artificial Intelligence has drafted guidance on how AI-assisted proofs should be published. The practical worry it addresses: once researchers paste unfinished work into a chatbot, separating a model's contribution from a person's becomes hard after the fact. Anthropic is pushing into the same area, so attribution is not a single-vendor question.
What is still unverified
Three gaps matter. The generating model is neither named nor available, so nobody outside OpenAI can reproduce the output. None of the 372 result families has completed peer review. And this report rests on one reachable trade-press account, because a second outlet we queried would not load — we could not independently check the Millennium Prize claim.
FAQ
How many AI math proofs did OpenAI publish on GitHub?
722 manuscripts, grouped into 372 result families. The file count is not the number of open problems solved.
What is Lean and why does it matter here?
Lean is a proof language whose steps a computer can verify. It confirms internal correctness, not whether a result is new or significant.
Has an AI solved a Millennium Prize Problem?
Unresolved. One result is said to touch such a problem, but that assessment comes from the release itself, not from peer review.
