- OpenAI published 722 mathematical manuscripts organized into 372 related result families. <sup>[\[1\]](https://openai.com/index/sharing-ai-progress-in-mathematics/)</sup>
- The work was generated by an internal frontier model that OpenAI has not released publicly. <sup>[\[1\]](https://openai.com/index/sharing-ai-progress-in-mathematics/)</sup>
-
OpenAI says roughly 4,000 open problems were attempted, with average compute equivalent to about three hours of ChatGPT Pro thinking per result. <sup>[2]</sup>
-
Many papers include Lean formalizations, but OpenAI says the collection is at different stages of verification and that some unformalized results could contain issues. <sup>[1]</sup>
-
An independent mathematics advisory group said the release is a first step, not a judgment or endorsement of the results. <sup>[3]</sup> OpenAI on October 6 published 722 mathematical manuscripts produced by an unreleased internal frontier model. The papers are organized into 372 families covering areas including number theory, geometry, theoretical computer science, mathematical physics and logic. [1]
The collection is housed in a public GitHub repository alongside source files, manuscript-specific citation and build instructions, Lean proof artifacts and an overview catalog. OpenAI said the model was evaluated on roughly 4,000 open research problems after its existing mathematics evaluations had saturated. [2]
The release follows OpenAI’s September announcement that an internal system had produced what the company described as a solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. That earlier claim drew scrutiny over verification and attribution, including questions involving concurrent work by mathematicians at Anthropic and New York University. [4][5]
The Release #
OpenAI’s repository describes the manuscripts as mathematical work produced by an internal model during the company’s evaluation of open research problems. A result family can contain a principal paper, companion arguments, consequences or alternative proofs, so the 722-manuscript count is larger than the number of distinct result families. [1]
The company published 10 abridged summaries of the model’s reasoning and estimates of the compute used to generate the work. OpenAI said the average result consumed compute equivalent to roughly three hours of ChatGPT Pro thinking. The company did not name the internal model in its release. [2]
The repository also establishes procedures for revisions and citations. OpenAI said it plans to add further formalizations and is exploring community-hosted alternatives for the collection. [2]
Verification Gap #
The papers do not all have the same evidentiary status. OpenAI said many, but not all, manuscripts have been formalized in Lean, a programming language used to check mathematical proofs by computer. It also warned that some unformalized results could contain issues. [1]
That distinction matters because the release presents a large body of AI-generated work before the mathematical community has independently assessed every claim. Lean can check whether a formal proof follows from its encoded premises, but experts still need to determine whether the formalized statement accurately captures the intended mathematical result. [6]
OpenAI’s September Navier–Stokes announcement included both a written proof and a Lean formalization, but the company said it was not seeking the associated Millennium Prize. The Clay Mathematics Institute’s prize process is separate from OpenAI’s announcement. [4][7]
The Advisory Group’s Response #
OpenAI said its release process was informed by the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. The group said it appreciated OpenAI’s engagement but stressed that its involvement was not an endorsement of either the results or the company’s process. [3]
“Making this work public is a first step,” the group said in its October 6 statement, adding that the release begins rather than completes the work of incorporating the results into mathematical knowledge. [3]
The group’s broader recommendations call for AI labs to publish model names, prompts, summarized reasoning, compute details and formalization status when releasing substantial mathematical output. They also urge labs to fund community-led efforts that help mathematicians understand complex AI-generated work. [8]
What Comes Next #
OpenAI said it will update the repository as additional proofs are formalized and will support workshops, conferences and other programs focused on understanding major results produced by AI. The company also said it is working toward releasing the model that generated the collection. [2]
For now, the manuscripts remain a research corpus rather than a public model or commercial product. Their significance will depend on how many claims withstand independent checking, how clearly the proofs can be understood by mathematicians and whether the work leads to subsequent results outside OpenAI.
Companies mentioned #
Further sources #
The stories that matter, in one email. Free — unsubscribe anytime.