OpenAI announced on October 26, 2026 that its internal AI models had produced a complete mathematical proof that four-dimensional quantum Yang-Mills theory, with quarks included, exists as a rigorous object. The proof runs 1,147 pages and, according to OpenAI, has been formally checked by the Lean proof assistant. Yet almost nobody has read it. The manuscript went to a five-member committee of senior physicists chaired by Nobel laureate David Gross, and that committee has chosen to keep it out of public view until it finishes its own line-by-line review. The result is a strange situation for science: a claimed answer to half of a Millennium Prize Problem, announced to the world, verified by a machine, and locked in a drawer.

At the same time, a Northwestern University mathematician working with a member of Anthropic's technical staff and many parallel instances of Claude has been attacking the same problem in public, posting preliminary results in October and predicting completion within weeks. Two frontier labs, two very different ways of doing science, and one unresolved question for everyone else: how should we treat a proof that we cannot yet read?

The problem: why the "glue" needed a proof

Almost none of the mass of ordinary matter comes from the particles you might expect. A proton is built from three quarks held together by gluons, the carriers of the strong force. The quarks themselves account for only about 1% of the proton's mass. The other 99% is the kinetic and potential energy of the quarks and gluons bound inside it. In a real sense, matter is mostly glue.

The mathematics of that glue is Yang-Mills theory, a non-Abelian gauge theory formulated by Chen Ning Yang and Robert Mills in 1954. When Yang presented it at Princeton, Wolfgang Pauli repeatedly pressed him on what mass the theory's field would have, and Yang had no answer. In 1973, David Gross and colleagues discovered asymptotic freedom: the strong force behaves like a rubber band, weak when quarks are close together and dramatically stronger as they pull apart. That work earned Gross the 2004 Nobel Prize in Physics, and the resulting theory, quantum chromodynamics (QCD), is confirmed routinely at particle accelerators such as CERN's Large Hadron Collider.

The physics worked. The mathematics underneath it did not. Calculations performed in continuous space produce infinities, which physicists removed through renormalization procedures that Richard Feynman himself described as dubious tricks. In practice, physicists compute on a finite lattice, a grid that chops space into discrete cells, then assume the results remain well-behaved as the cell size shrinks to zero. Think of a pixelated screen: the picture must resolve cleanly as the pixels become infinitely small. Nobody had proven that this continuum limit actually exists for Yang-Mills theory in four dimensions.

Translucent proton sphere with three quarks joined by glowing elastic bands above a grid that grows finer left to right

▲ Quarks bound by gluons and the lattice continuum limit

In 2000, the Clay Mathematics Institute named "Yang-Mills and the mass gap" one of its seven Millennium Prize Problems, with a $1 million bounty. The prize requires two things: a proof that the quantum field theory rigorously exists, and a proof that it has a positive mass gap, meaning its lightest excitation has strictly positive mass. OpenAI's claim covers only the first.

Two labs, two approaches

The strongest prior attempt came from mathematician Tadeusz Bałaban, who wrote 14 dense papers in the 1980s building a rigorous framework for gauge theories before stopping. Gross has noted in lectures that serious progress on this constructive program largely ceased afterward. The ideas were there; the sheer volume of technical work was not something any individual could finish.

That is exactly the bottleneck that AI changed. In early September 2026, a mathematician at Northwestern University set out to complete Bałaban's program with automated help. On September 8 he contacted a mathematician on Anthropic's technical staff, and together they orchestrated multiple distributed instances of Claude to generate thousands of pages of detailed proofs and cross-check each other's steps. The Northwestern researcher told Scientific American that the obstacle in constructive field theory had never been a shortage of concepts, only the enormous manual workload. The team posted its results openly on October 6, and on October 25 the lead researcher publicly estimated that full completion was weeks away.

One day later, OpenAI announced that its own models had already finished. The timelines side by side:

Date (2026) OpenAI (private) Northwestern and Anthropic team (public)
Sept 8 Collaboration begins using Claude
Sept 9 AI generates core Yang-Mills existence proof
Sept 18 Proof extended to include quarks
Oct 6 Preliminary results posted openly
Oct 16 Lean formal verification completed
Oct 25 Lead researcher projects completion within weeks
Oct 26 Public announcement; manuscript sent to five-person committee

The compression is striking either way. Work that took Bałaban roughly a decade took the Claude-based team about seven weeks, and OpenAI's system reportedly moved from pure gauge theory to a version with quarks in nine days.

The drawer: why only five people have the proof

OpenAI did not publish on September 9. It spent four weeks pushing the proof through Lean, an interactive theorem prover that checks each logical step of a formal argument the way a calculator checks arithmetic, without human slips. Lean verification finished on October 16. OpenAI then delivered the 1,147-page manuscript to a "shepherding committee" of five prominent physicists: Michael R. Douglas, David J. Gross as chair, Arthur Jaffe, Igor R. Klebanov and Nikita Nekrasov. The committee's report, titled "Proof by OpenAI of the Existence of Quantum Yang-Mills theory" and dated October 2026, exists, but the proof itself does not circulate beyond those five.

Mechanical checking machine stamping a long scroll beside a sealed case holding a thick volume and five reading desks

▲ Formal verification versus closed human review

The committee's stated reason is that it wants to complete a thorough manual verification before the paper is disseminated. That decision has drawn pushback from several directions. OpenAI's own external Advisory Group on Mathematics and AI, whose members include Fields Medalist Edward Witten, said there was no valid scientific justification for delay. The Anthropic staff mathematician on the rival project wrote on X that he was disappointed OpenAI had declined to collaborate openly with the ongoing academic effort. The objection is simple: mathematics has historically advanced through open public verification, and a proof that cannot be inspected is, to the wider community, still a claim.

There is a plausible case for the committee's caution as well. A machine-generated argument of more than a thousand pages is not something a reviewer can digest in a weekend. The episode appears to expose a genuine new difficulty: human experts now have to audit, and truly understand, output that a machine produced in days, and nobody yet knows how long that should take.

How to read an AI proof you cannot see

Two separate questions are easy to blur here. First: is the proof correct? Second: is it accepted?

On correctness, the Lean verification carries real weight. A proof that Lean accepts has been checked step by step, so in the formal sense its reasoning contains no gaps or errors. The committee nonetheless wants to work through the arguments by hand before anyone else sees them. Until that review concludes and the manuscript is public, the honest description is a machine-verified argument that human experts are still evaluating.

On acceptance, nothing has been settled. The Clay prize is not awarded, because the mass gap half remains unproven. Pauli's 1954 question about the mass of the field is, 72 years later, technically still open. And no independent group outside the five reviewers has reproduced or even read the result.

It also helps to be clear about what does not change. QCD already predicted accelerator data accurately for 50 years, and textbook formulas remain the same. The achievement, if it holds, is foundational rather than empirical: it would show that a cornerstone of theoretical physics rests on sound mathematics instead of unvalidated approximations. Physicist Sabine Hossenfelder has pointed out that some dismissive reactions ring hollow, since a human author producing the same result would be celebrated widely.

Why this matters beyond physics

Whatever the committee concludes, the past two months mark a shift in how fundamental research happens. The chess comparison is tempting but incomplete: chess engines were built for one game, while the systems at work here reason across domains and are being pointed at the hardest open problems in several fields at once. Both OpenAI and Anthropic are investing in proprietary laboratory initiatives aimed at biology, pharmacology and materials science, and mathematics and theoretical physics sit underneath all of them.

The cultural change may be as consequential as the technical one. Discoveries that once emerged through public preprints and open peer review are now arriving via corporate announcements, with access controlled by the company and a committee it selected. One team chose to publish its partial progress in the open; the other chose secrecy followed by a closed review. Which model the field adopts will shape who gets to check, build on and trust the next decade of results. On an optimistic reading, the remaining Millennium Prize Problems could fall to AI within one to two years, though that is a prediction, not a result.

What to take away

  • OpenAI claims an AI-generated, Lean-verified, 1,147-page proof that quantum Yang-Mills theory with quarks exists. The claim is credible on its face but remains unpublished and under review by five physicists chaired by David Gross.
  • The mass gap problem is still open. No Millennium Prize has been awarded, and no physical prediction has changed.
  • A separate, openly posted effort using Claude was weeks from completing the same existence proof when OpenAI announced.
  • The controversy is less about the mathematics than about process: whether a corporate lab should hold a scientific result behind a private panel.

For readers following this, the practical stance is patience with precision. Treat the existence proof as a strong, formally checked claim rather than an established theorem until the manuscript is released and independent experts have examined it. Watch for the committee's full report and for the public team's completed work, which would offer an independent route to the same result. And if you work in a field that depends on heavy, mechanical mathematical labor, note the method that both projects share: break a decades-long program into modular sub-proofs, run them in parallel across many model instances, and verify the output with a formal theorem prover rather than trusting it on sight.