OpenAI publishes 722 model-generated mathematics manuscripts
On 6 October 2026 OpenAI released a public GitHub repository containing 722 mathematical manuscripts, organised into 372 result families, produced by an unreleased internal frontier model. The model was given roughly 4,000 open problems across number theory, algebraic geometry, analysis, theoretical computer science and mathematical physics; accepted results consumed on average about three hours of ChatGPT Pro-equivalent compute, and many were accompanied by Lean proof formalizations. OpenAI said it had consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on release practice, and subsequently stated that nearly all results came from a single prompt to a single agent rather than a multi-agent swarm. Three manuscripts were withdrawn the following day over a sign error.
Why It Mattered
This is the point at which machine-generated mathematics stopped being presented as benchmark performance and started being presented as literature. Earlier AI mathematics results arrived one at a time and were framed as demonstrations; this release asserted hundreds of simultaneous contributions to the research record, with formal proof artefacts attached and citation blocks supplied. The scale matters more than any individual claim: a single model, over a period of weeks, generated more purported new results than a large research group produces in years, and the mathematical community had no established process for absorbing or refereeing that volume. The disclosure that the output came mostly from one prompt to one agent — far cheaper than the thousand-agent runs used on earlier problems — indicated that the production of such results was becoming economical rather than heroic. Two consequences follow for the historical record. First, verification became the bottleneck rather than discovery: Lean formalization carried part of the load, but the human-readable manuscripts were dense enough that specialists reported needing AI assistance to read them, and a sign error surfaced within a day. Second, the release forced the question of who sets the norms for releasing machine-produced results, which OpenAI's own advisory group had already tried to answer with guidance that it was accused of disregarding. A history of AI written in 2030 would use this week as the date on which a research discipline first confronted automated production of its primary output at industrial volume.
Who Built It
OpenAI
Applications
- Mathematics
- Automated Theorem Proving
- Scientific Research