OpenAI Published 722 AI-Written Math Papers on GitHub — and Admits Some May Have Errors
October 8, 2026
OpenAI has released 722 mathematical manuscripts generated by an unreleased internal AI model, posted publicly to GitHub on October 6. The company says the collection reports solutions to hundreds of open problems — but it warns that some of the results may contain errors.
What happened
The papers live in a public repository called openai/math, organized into 372 families of related results and classified by mathematical discipline, released under an Apache 2.0 license. According to OpenAI, an unnamed internal frontier model was posed roughly 4,000 open problems during evaluation, and each surviving result consumed, on average, about three hours of ChatGPT Pro thinking compute.
Many of the manuscripts come with formalizations in Lean, a proof-assistant language that lets a computer verify a proof step by step, independent of human review. But not all of them do — and OpenAI’s own README is blunt about it: some unformalized results “could have issues.” The company says it will fix errors quickly and keep a public revision history so earlier versions stay accessible.
The details
The release also includes abridged reasoning summaries for ten of the result families, covering topics like the irrationality exponent of pi, the Mahler conjectures, NP-hardness, spin glasses, and the three-dimensional relativistic Vlasov-Maxwell system. Two results sat outside the standard procedure: work on a zero-free region for the Riemann zeta function (for Re(s) > 11/12, with the write-up human-edited for readability) and a proof of the Hodge Conjecture for CM abelian varieties.
OpenAI frames the project as a byproduct of model evaluation — it expanded the effort after performance on existing mathematical benchmarks saturated. Before publishing, the company consulted the Institute for Advanced Study’s Advisory Group on Mathematics and Artificial Intelligence, which published responsible-release recommendations on September 29 after collecting more than 600 responses from the mathematical community.
This is the third major math release from the lab in about nine weeks. On August 1, OpenAI unveiled an internal model that produced ten results on long-open problems, including a sphere-packing improvement that hadn’t moved since 1978. Then on September 8 came a claim of finite-time blow-up solutions to the Navier-Stokes equations — one of the Clay Mathematics Institute’s seven Millennium Prize Problems — backed by a 166-page manuscript.
Why it matters
The honest headline here is scale plus candor: 722 manuscripts is an astonishing research output, but the interesting part is OpenAI shipping them with an explicit “some of this might be wrong” label and a revision history. That’s a healthier model than quiet confidence — mathematical proof is exactly the kind of thing computers can check, and Lean formalizations give skeptics something concrete to audit. The model itself remains unreleased, which will frustrate anyone who wants to reproduce the pipeline. Reaction among mathematicians has been mixed — genuine awe at the volume, alongside real questions about whether machine-generated proofs without full formalization actually advance the field. We’ll be watching which of these results survive peer scrutiny.
FAQ
How many math papers did OpenAI publish?
722 manuscripts, organized into 372 families of related results, posted to GitHub on October 6, 2026 under an Apache 2.0 license.
Can I trust the proofs?
Many include Lean formalizations that a computer can verify. OpenAI warns that unformalized results could contain errors, so check the verification status before building on any result.
Did OpenAI release the model that wrote them?
No. The model remains unreleased and unnamed.
What’s the most significant claim?
Among the highlights are a proof of the Hodge Conjecture for CM abelian varieties and a new zero-free region for the Riemann zeta function — both still awaiting independent verification.
Sources: AI Weekly, BigGo Finance, Startup Fortune, AutonAI News