OpenAI Publishes 372 Math Results, Saying One Prompt to One Agent Made Them
OpenAI published 372 mathematical result families to a public GitHub repository on 6 October, claiming nearly all came from a single prompt given to one unreleased internal model. Headlines include work on the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis. Many proofs are formally checked in Lean, but whether the work is genuinely novel will take mathematicians months to assess.
On this page
A repository where a press release used to be
OpenAI published 372 mathematical result families to a public GitHub repository on 6 October, one month after a claimed Navier-Stokes solution that was itself generated by a 10,000-agent swarm costing millions in compute. The company says nearly all of the new results came from a single prompt given to a single AI agent running on an internal, unreleased frontier model, a claim that landed with a spokesperson's own qualifier that some results "might have taken multiple attempts."
The headline claims are large: a solution to the four-dimensional Kakeya conjecture, improvements to major computer algorithms, and progress toward the Riemann hypothesis, with each result family resolving or substantially advancing a major open question in mathematics or theoretical computer science. Coverage from The Verge confirms the release and the framing as a batch of long-standing problems addressed at once.
What is verified, and what is not
The verification picture splits cleanly in half. Many of the proofs have been formally checked in Lean, the proof-validation language, which makes them almost certainly correct as mathematics. What nobody can confirm yet is novelty: whether the work contains genuinely new ideas or recombinations of existing techniques is a question that will take mathematicians months to answer, and many results are not yet understood even by OpenAI's own mathematicians, per the company.
The process fight is as loud as the math. After the Navier-Stokes controversy, OpenAI formed an independent advisory group whose recommendations call for releasing the model, the exact prompts, and compute time per result; OpenAI disclosed only average compute statistics, released no prompts, and says it is not bound by the guidelines. MIT mathematician Andrew Sutherland's response to the single-agent claim was to ask for receipts, arguing it should be treated as unverified until the model itself is released.
Why this matters beyond mathematics
Two things are easy to miss under the headlines. First, the release itself is a behavior change: whatever one thinks of the claims, the work is public on GitHub for the field to check, which University of Toronto mathematician Daniel Litt welcomed on the grounds that secrecy serves no one. Second, the economics changed. The previous milestone took a 10,000-agent swarm; this one, by the company's account, took one prompt, which if accurate says more about the model's reliability at long-horizon research tasks than any benchmark number published this year.
The receipts question will resolve the way these questions always do: slowly, in public, on the repository. Anyone can read the proofs, run the Lean checks, and look for recombination versus invention, and the mathematics community will do exactly that over the coming months. Treat the announcement as a claim with unusual instrumentation attached, not as settled fact, and note that the instrumentation is now finally in everyone's hands.