OpenAI has thrown the mathematics world into turmoil. On October 6 and 7, the company published 722 manuscripts on GitHub claiming progress on 372 families of long-standing open problems, spanning theoretical computer science, physics, and mathematics, produced by an unreleased internal AI model whose full capabilities have not been disclosed.
The claimed highlights are extraordinary. OpenAI's post points to a proof of the Unique Games Conjecture, a load-bearing assumption in complexity theory, alongside Barnette's Conjecture and claimed progress on four of the seven Millennium Prize Problems. Just weeks earlier, the company had announced that the same internal system produced a proposed solution to the Navier-Stokes existence and smoothness problem, one of the Millennium Prize Problems carrying a one million dollar bounty, using roughly 10,000 autonomous AI agents over an 88-hour run that generated 130 billion tokens and 2.7 million inter-agent messages. OpenAI said it has no plans to claim the Clay Mathematics Institute prize.
The math community's reaction has been sharp. Fields Medalist Terence Tao of UCLA called the pace of AI-generated results "insane," and shared a statement from the Association for Human Mathematics urging mathematicians to stop collaborating with OpenAI and return to a vision of science that centers human understanding. NYU professor Tristan Buckmaster questioned whether 372 result families could plausibly be reviewed in the available time, and raised concerns about whether data from private Codex sessions had been used to train the model; OpenAI denied accessing private user data while conceding it could not rule out that de-identified usage data had helped.
There is also a genuine verification crisis. Only 10 of the manuscripts reportedly included the model's chain-of-thought reasoning, and just 42 percent underwent Lean code formalization for machine checking. That leaves the mathematical community with hundreds of dense papers to check, explain, and digest, work that moves at human speed no matter how fast the proofs were generated. Cambridge mathematics professor Timothy Gowers said he was "shocked" by how quickly the models improved this year, from mediocre graduate student to top-mathematician level. But solving is only the first of five hurdles: created, checked, explained, accepted, and digested.
Twenty-seven Fields Medalists had already co-signed a critical open letter weeks earlier, and OpenAI responded by creating an independent Mathematics and AI Advisory Group with nine leading mathematicians and physicists. The group advises on publication standards, but it does not control the pace of the research.
