Did AI crack the first half of a million-dollar physics puzzle? Expert panel says OpenAI proved Yang-Mills theory 'exists,' but the hardest part, the mass gap, is still unsolved

The backstory: physics' favorite theory has shaky math foundations

Everything around us is made of fundamental particles, which interact through a handful of basic forces. Physicists describe those forces with the Standard Model, one of the most successful theories ever devised; its predictions match experiments with stunning accuracy.

Yang-Mills theory is part of the mathematical skeleton underneath it. The strong force, which glues protons and neutrons together inside atomic nuclei, is described by this kind of theory.

Here's the awkward part. Physicists have calculated with it, extremely accurately, for decades, yet mathematicians have never rigorously proven that the theory can be built in mathematics completely and without contradiction. Imagine a building people have lived in happily for decades whose structural calculations were never formally checked.

In 2000, the Clay Mathematics Institute named this one of its seven Millennium Prize Problems, with a $1 million prize. It has two parts: first, prove the theory exists mathematically; second, and harder, prove it has a "mass gap."

In plain terms, the mass gap means showing that even the lightest particle in the theory can't be massless: there has to be a minimum mass above zero. That's closely tied to why the strong force only acts over extremely short distances. Physicists believe it's true; no one has proved it mathematically.

What happened

A working group chaired by theoretical physicist and Nobel laureate David Gross has published its assessment of a proof submitted by OpenAI, together with the underlying Lean code in a public repository. The assessment is dated November 10, 2026; the repository went online at 03:38 UTC on November 11, which was the night of November 10 in U.S. Eastern Time. It covers OpenAI's first submission, dated October 28, not a later submission the group is still reviewing.

In the group's judgment, the proof shows that, for any simple compact gauge group (roughly, any of the standard versions of this kind of theory), you can chop four-dimensional space-time into a grid of tiny cells (what physicists call Wilson lattice gauge theory), make the grid finer and finer until it approaches smooth, continuous space-time, and end up with a quantum field theory that satisfies the "Wightman axioms." Think of those axioms as a certification standard: the basic rules any respectable quantum field theory has to obey.

Translated: for the first part of the Millennium Problem, whether pure Yang-Mills theory can be rigorously built in mathematics, the group believes there is now a proof that a computer can check step by step.

What's a Lean proof? Lean is a computer language designed to check mathematical proofs. Every step written in it must pass the program's checks, with no skipped steps or hand-waving. It's like an exam grader who never lets anything slide.

Where the claim stops

The group draws clear lines:

First, the mass gap is not proved, and OpenAI's proof doesn't claim it. So under the full Jaffe-Witten statement of the Millennium Problem, the problem is not solved.

Second, the proof doesn't show that the "infinite-volume limit" is unique. Loosely put, it proves the theory can be built, not that this is the only way it can come out.

Third, the group openly says it used OpenAI's Astra model to help digest the Lean code, and that "we do not claim to fully understand the proof yet." Passing a computer check is not the same as humans fully understanding every step of the argument.

The group says a clean rebuild, an axiom audit, and a Comparator check all passed, and the repository includes reproduction instructions and dated verification records. It adds that its only present connection to OpenAI, besides accepting submissions, is the use of Astra, and that it has no present connection to the Clay Mathematics Institute.

Scientific American reported on November 11 that OpenAI had announced in late October that its model completed the existence part. What's new is that the assessment and the Lean code are now public for wider inspection.

There's controversy, too. Routing the release through a designated committee first drew criticism from some mathematicians who wanted the materials opened sooner. New York University mathematician Scott Armstrong told Scientific American he welcomed the Lean proof's release and hoped future results would be shared more widely from the start.

What it could change

If the proof holds up to independent reproduction and wider human reading, it could matter in two big ways.

For mathematics, half of a foundational question that has stumped mathematical physicists for decades would have a machine-checkable answer.

For how science gets done, AI would be doing more than looking things up or writing code; it would be contributing to some of the most rigorous frontier mathematics, in a form where every step can be checked. That raises a new question: when a proof is so long that experts need AI help just to read it, how should we think about the gap between "proved" and "understood"?

The hardest part, the mass gap, remains open, and independent reproduction is just beginning.

What it means for you

In the near term, this won't change your phone, your power bill, or your daily life. Physicists already compute with Yang-Mills theory; this is about firming up the foundation, not inventing a new machine.

It's still worth watching as a signal: AI is moving into the hardest, most rigor-demanding corners of human knowledge. The same step-by-step machine checking could one day be used to verify engineering, software, or financial models that touch our safety and our wallets. And whether AI truly "understands" what it proves may be debated as long as the proof itself.

Original source published: Yang-Mills Working Group, Nov 11, 2026, 11:38 Beijing time (UTC+8); Scientific American, Nov 11, 2026, 20:30 Beijing time (UTC+8)