OpenAI Published 722 Math Manuscripts From an Unreleased Model. Verifying Them Is the Hard Part

OpenAI published 722 mathematical manuscripts produced by an unreleased internal frontier model, grouped into 372 families of results. Only a fraction arrive with a Lean-verified main proof, and with no model or prompts disclosed, reproducibility — not discovery — is the hard problem.

Key Takeaways

OpenAI published 722 mathematical manuscripts produced by an unreleased internal frontier model, grouped into 372 families of results.
Only a fraction arrive with a Lean-verified main proof, and with no model or prompts disclosed, reproducibility — not discovery — is the hard problem.

AI synthesis5 sources3

OpenAI released 722 mathematical manuscripts on GitHub this week — the output of an internal frontier model the company has not released — grouped into 372 "families" of results across open and long-standing problems. The repository is public under an Apache-2.0 licence, and it mixes finished write-ups with machine-checkable Lean formalizations, the same proof assistant OpenAI leaned on for its disputed Navier–Stokes claim in September.

The scale is the headline; the verification is the story. Only a fraction of the manuscripts ship with a Lean-verified main proof — one tally circulating among mathematicians puts it at 162 of 722, with around 560 lacking a machine-checked central result. OpenAI's own README warns that unformalized results "could have issues," and the release does not include the model, its weights, or the prompts that produced the work, leaving mathematicians to audit the text by hand without any way to reproduce the derivations. New Scientist's verdict: OpenAI has "dumped 722 maths papers — now it must clean up the mess."

Coverage highlighted at least one result described as approaching the Riemann hypothesis, which one mathematician quoted in trade coverage said would be Fields Medal-worthy work if a human had produced it. That claim, like the rest, remains unverified. The Wall Street Journal noted that the release follows the company's September announcement of a Navier–Stokes Millennium Prize solution, produced by its GPT-6 Astra system in roughly 17 hours, and quoted an observer calling the new material potentially "the most significant moment in mathematical history."

The context matters. OpenAI's September claim triggered a public fight over credit, with mathematicians alleging their work was used without attribution — a dispute this wire covered at the time. An external advisory group had asked OpenAI to publish papers rather than bulk-upload machine output, and Lean verification, while rigorous where it is applied, only checks that a proof is correct, not that a theorem is interesting or that the result is new.

What happens next is a slow, human grind: reading, checking, and either confirming or rejecting hundreds of claims, most of which currently have no computer-checkable proof attached. OpenAI has framed the release as an open invitation to scrutinize — which is precisely what the mathematics community is now doing, one manuscript at a time.