OpenAI posts 722 AI-written math papers, then pulls three

OpenAI posted 722 math papers written by an unreleased internal model on October 6, 2026. Its research post says the results, grouped into 372 families, came from about 4,000 problems the model was given.

The batch came a month after an OpenAI model proposed a solution to the Navier-Stokes problem. Nature reported that some mathematicians cheered, while others said they'd been scooped. Oxford's Ursula Martin called it "tossing the community a messy first draft and expecting us to clean it up for them."

OpenAI's model claims proofs of long-open problems

The openai/math catalogue claims proofs of Khot's Unique Games Conjecture, Barnette's conjecture in graph theory and the quasi-Riemann hypothesis. Each result took about three hours of ChatGPT Pro thinking compute on average.

OpenAI says two results, on the Riemann zeta function and the Hodge conjecture, didn't follow that standard process. A person also edited one zeta write-up to make it easier to read.

Only about 42% of the top results are checked by computer

About 42% of the top results have proofs written in Lean, a language that lets a computer check each step. The repository warns that "some of the unformalized results could have issues."

On October 7, OpenAI pulled three papers after a sign error broke one and the two that built on it. All three were tied to the Hodge conjecture. It also fixed 14 others, and the catalogue now lists 719 papers.

The group OpenAI consulted had asked labs to stop

OpenAI says it took advice from an advisory group of mathematicians at the Institute for Advanced Study. The group's September 29 advice opens by asking labs to stop testing hard math problems on models the public can't use.

The release meets some of its other asks, such as Lean proofs, compute estimates and a count of the problems the model tried. But it misses others: the group wants papers kept where no AI lab controls them, and OpenAI hasn't named its model. It also shared reasoning summaries for only 10 of 372 families.

The Verge spoke with more than three dozen mathematicians and reported that making sense of the release could take years. The Hacker News discussion drew 1,528 comments.

Recaply read the top 30 comments there on October 11, a snapshot rather than a survey. About half were excited by specific results. Five wrote about lost work or careers, and two asked how the proofs would be checked.

I spent thousands of hours on that problem. I really enjoyed it.

jboggan on Hacker News, October 7, on Barnette's conjecture

The papers, Lean files and reasoning summaries are free on GitHub under the Apache 2.0 license. OpenAI says fixes will be kept as new versions and it's looking at community-hosted homes for the papers. The model isn't available, though OpenAI says it's working to release it and will fund workshops on AI results.