On 9 and 10 November 2026, a committee appointed by OpenAI and chaired by physicist David Gross said an AI-generated proof of the “existence” half of the Yang-Mills Millennium Prize Problem had been formally verified in the Lean proof language, while stating that it does not yet fully understand the proof and that the harder “mass gap” half is untouched. The result is the most closely watched test so far of how research communities should vet AI-produced mathematics, and it is contested on process as well as substance.
What happened and when
OpenAI announced on the evening of Monday 26 October 2026 that an internal model had produced a proof of the existence part of the Yang-Mills problem, one of the Clay Mathematics Institute’s seven Millennium Prize Problems (a $1 million prize). According to Scientific American‘s 27 October report, OpenAI said the model reached the result on 9 September but kept it private until the argument had been encoded in Lean, and that the company did not release the proof but handed it to a group of experts to bring it into the mathematical literature.
That group, described by Scientific American as a committee of “shepherds”, is led by David Gross and includes four colleagues he chose; Michael R. Douglas of Harvard is a named member. On 9 November the committee issued a statement saying the result proves the existence of pure Yang-Mills theory under the standard problem statements. On the night of 10 November it posted a document on GitHub explaining how the proof works and affirming that the proof had been formally verified in Lean. Scientific American published its report on 11 November and updated it on 12 November.
What the committee did and did not verify
The Yang-Mills problem has two parts: showing the theory underlying the Standard Model can be constructed rigorously in four dimensions, and showing that a “mass gap” emerges from its equations. Only the first is claimed. The committee’s own wording is cautious: it says “We do not claim to fully understand the proof yet.” Its explanatory document is described as cursory, and the committee said it used OpenAI’s Astra model to help parse the Lean code.
Two separate questions are easy to blur here. A Lean check confirms that a formal statement follows from its formal assumptions. Whether that formal statement captures what the Clay problem intends is a judgement that people still have to make. Physicist Peter Woit, writing on 11 November, argued that existence in the sense verified is not the standard the Millennium problem requires, citing the Jaffe-Witten problem description on weak existence, and that nothing has been established about the mass gap. That is one critic’s reading, not a finding, but it shows why independent review of the formal statement matters as much as the Lean check itself.
Process disputes: gatekeeping and a rival effort
The shepherding arrangement itself drew objections. Scientific American reports it conflicts with transparency recommendations from an OpenAI-affiliated advisory group, and some mathematicians see it as needless gatekeeping. NYU mathematician Scott Armstrong welcomed the Lean release but said the review “can be done more efficiently and quickly in the open.” Woit wrote that the committee is keeping a roughly 1,147-page explanatory document private and that, within hours of the Lean release, other groups had begun using AI agents to produce public, readable versions of the code.
A parallel effort is also under way. Levent Alpöge, who works at Anthropic, and Northwestern mathematician Ethan Sussman had been posting progress on GitHub since 6 October using Claude-based agents, and the night before OpenAI’s announcement updated the repository, believing they had solved the three-dimensional version of the existence problem. The Clay requirements call for four dimensions, so that result would not meet the prize criteria. NYU’s Tristan Buckmaster disputed OpenAI’s timeline, saying “There is no way it took one month to Lean-verify Yang-Mills.”
What it means for researchers and research offices
- Verification is not publication. Neither the Lean check nor the committee statement is a journal review. The committee itself says it does not yet fully understand the proof, so formal checking does not by itself settle correctness of the intended statement, priority or credit.
- Open versus gated review. Journals, funders and institutions drafting AI policies now have a live example of a developer controlling the release sequence. Teams reusing the Lean code should record exactly which version of the code and statement they checked, in line with computational reproducibility practice.
- Who is credited. Questions about naming the model, the company, and the human reviewers connect to AI research tool disclosure and to how peer review will treat AI-generated arguments; see also AI in peer review.
- Preprint norms. With the Lean code public, outside experts can scrutinise it, possibly with their own AI agents. Expect the community check to run through preprints and public repositories before any journal decision.
What remains open
Still unresolved as of 12 November 2026: whether outside experts confirm the Lean statement matches the Millennium problem’s existence requirement; whether the proof can be rewritten into something humans can follow; the mass gap; and OpenAI’s separate claims about quantum chromodynamics and other Standard Model results, about which, according to Woit, the committee has said nothing.
