AI-generated mathematics manuscripts from OpenAI

OpenAI, Simons FoundationBuilt by: big techStatus: LiveYear: 2026

Last updated October 2026

Overview

On 6 October 2026 OpenAI released a broad range of new mathematical results produced by an internal frontier model, publishing them in a GitHub repository with protocols for paper revisions and citations. Let’s Data Science reports that the release contained 722 manuscripts, grouped into 372 families of results across 17 fields. It also reports that the model was given about 4,000 open problems. The repository includes Lean formalisations of many of the proofs, which OpenAI says it will add to, along with ten summaries of the model’s reasoning, estimates of the compute used and statistics on the number of problems attempted. OpenAI says the average result used the equivalent of about three hours of ChatGPT Pro thinking.

OpenAI says it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study and drew on its public recommendations. Those recommendations had asked AI companies on 29 September to disclose the model, prompts and compute behind such results. Let’s Data Science reports that OpenAI disclosed the counts, the average compute and the reasoning summaries but not the prompts, the compute for each problem or the model’s name. On 7 October OpenAI withdrew three algebraic geometry papers because of a misplaced minus sign.

On 15 October 2026 OpenAI added that it is contributing $5 million to a new Simons Foundation initiative, managed by the foundation, to help mathematicians examine, understand and build on results produced by AI, through workshops, conferences and special programmes. OpenAI will also give researchers supported by the foundation access to its models and computing resources.

Notes and quotes

  • Quanta Magazine reports that OpenAI claimed Lean verification for about 42% of the papers, and that Johan Commelin, director of the Mathlib Initiative, said some formalisations prove weaker statements than the papers claim.
  • Daniel Litt of the University of Toronto suggested that, for now, the papers be treated as “research findings” or “raw data”, according to Quanta Magazine.
  • OpenAI says it is working to release the model that produced the results.

Sources

This entry was written with the help of AI, from the sources listed. Please check the sources before relying on it, and let us know if anything needs correcting.