OpenAI publishes about 720 maths manuscripts written by an unreleased AI model
The papers sit in a public GitHub repository, many with computer-checked Lean proofs. They have not been peer reviewed, and OpenAI says some unformalised results could have issues.
How this story was checked
- ResearchedWith AI assistance · 4 sources
- AI source checkEach sentence matched to its source
- Editor approvalMayank Sahu · 11 Oct, 20:20
- Published11 Oct, 20:20
In short
On 6 October 2026 OpenAI released maths results produced by an internal model that is not public yet. A public GitHub repository holds about 720 manuscripts in 372 families. Many have Lean proofs that a computer can check; none are described as peer reviewed.
What OpenAI released
OpenAI said on 6 October 2026 that it is releasing “a broad range of new mathematical results” produced by an internal frontier model that is not yet public.1 The results are in a public GitHub repository under the Apache 2.0 licence.24
How many
Coverage of the release describes 722 manuscripts grouped into 372 families of related results.34 When we checked the repository on 11 October, its README listed 719 manuscripts in 372 families.2 The repository can change over time, so we describe the count as about 720.
The model was given about 4,000 problems, and an average result used compute equal to roughly three hours of ChatGPT Pro thinking.12
How the results were checked
Many results come with Lean proofs, which a computer can check step by step.12 The repository says roughly 42% of the top-line results are formalised in Lean and warns that “some of the unformalized results could have issues”.2 A Lean check confirms that each step follows from the stated assumptions; on its own it does not show that every claim in a paper is right.3
The release does not mention formal peer review.1 OpenAI says corrections will be published as new versions, with earlier versions kept available.24
Advice from mathematicians
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 for this release.1 Unite.AI reports that the group recommends disclosing prompts and compute costs, and prefers repositories that no AI lab controls.4 OpenAI says it is exploring community-hosted options.1
Two exceptions
The repository says two results did not follow the standard procedure: a zero-free region for the Riemann zeta function, whose write-up people edited for readability, and a proof of the Hodge Conjecture for CM abelian varieties.2
What is not known yet
How many of these results mathematicians will judge to be new and important. That needs independent review. OpenAI says it will fund workshops and programmes to help people study major AI-produced results.1
Fact ledger
How labels workQuestions readers ask
Are these maths papers peer reviewed?
The release does not say they are. About 42% of top-line results have Lean proofs that a computer can check, and OpenAI warns some unformalised results could have issues.
Can I use the model that wrote them?
No. OpenAI says the internal model is not released yet.
Where can I read them?
In OpenAI’s public GitHub repository, openai/math, under the Apache 2.0 licence.