Research Story
OpenAI was due to release Lean code for its Yang-Mills proof, per Numerama, but the mass gap is unaddressed
On 10 November 2026, OpenAI was due to publish on GitHub the Lean code of its Yang-Mills proof and a five-physicist assessment, which says it shows existence of pure Yang-Mills theory; according to Numerama, the proof does not address the mass gap.
This story was written with the help of artificial intelligence, checked automatically, then reviewed by an editor before publication, under our AI policy.
What was released and by whom
According to Numerama, OpenAI was due to publish on GitHub on 10 November 2026 the Lean code of its proof on Yang-Mills theory, together with the assessment of the five physicists asked to examine it.
OpenAI announced in late October that one of its internal models had proved the mathematical existence of Yang-Mills theory, Numerama reports. Unlike for two other problems, OpenAI provisionally kept this proof out of public reach.
A GitHub repository named "OpenAI-YM-20261028", under the yang-mills-group account, presents a Wilson construction of Yang-Mills under the Apache-2.0 licence, with Lean modules and verification instructions.
What the committee and the repository say
The committee is chaired by David Gross, 2004 Nobel laureate in physics, and includes Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, according to Numerama. Per the same source, it concludes that the code "démontre l’existence de la théorie de Yang-Mills « pure »", in line with reference statements.
The repository README, which is a statement from the project itself and not an independent finding, says the Lean-verified proof covers 17 auditable theorems. It adds that the Lean kernel replayed 100,168 declarations of the project using only standard axioms.
A readable plain-language version of the proof will appear in neither of the two repositories, according to Numerama.
What is not covered
According to Numerama, the proof does not address the "mass gap", the other component of the Millennium Prize problem, and therefore is not enough to solve it. The README is consistent with this: it states that the public theorem explicitly excludes a positive mass gap and the uniqueness of alternative contacts, both outside the scope of the claim.
The existence result and the committee's conclusion therefore concern a narrower statement than the full prize problem. Whether the physicists' assessment goes beyond the existence claim is not stated in the material available. One assessment file, assessment-1.pdf, sits in a separate "YM" repository of the same group, but its text content is not accessible in the page provided, so its details are not reported here.
Background: custody and a call for release
According to Numerama, on 16 October OpenAI asked David Gross to take custody of the proof and decide on its release. The AGMAI, an advisory group of nine mathematicians hosted by the Institute for Advanced Study and including Edward Witten, had called for publication without delay.
According to Numerama, OpenAI announced the result in late October, and the publication of the code was planned for 10 November 2026. The material available does not address independent verification beyond the five-member committee.
FAQ
Does the proof solve the Yang-Mills Millennium Prize problem?
No. According to Numerama, it does not address the mass gap, the other part of the problem, and so does not suffice to solve it.
Who reviewed the proof?
A committee of five physicists chaired by David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, per Numerama.
What does the Lean verification cover?
The README says 17 auditable theorems, with the Lean kernel replaying 100,168 declarations using only standard axioms.
Sources
- Après des semaines de frustration chez les mathématiciens, la preuve secrète d’OpenAI sur l’un des grands problèmes de physique sort enfin, Numerama,
- https://github.com/yang-mills-group/OpenAI-YM-20261028, github.com, (primary source)
- https://github.com/yang-mills-group/YM/blob/main/docs/assessment-1.pdf, github.com, (primary source)