AI Labs Claimed a Record Day in Math. Here's What Holds Up and What Doesn't

AI Labs Claimed a Record Day in Math. Here's What Holds Up and What Doesn't

7 min readOctober 29, 2026

Quick verdict

For one day the AI math news was loud enough that Claude Opus 5.5, asked to assess it, called it "the biggest single day in the history of math and theoretical computer science." That line is doing a lot of work, and the honest read is that some of what landed is real and checkable while some is a screenshot and a confident claim. Axiom's formalization of a famous theorem in Lean is the part you can actually verify. The dueling claims from OpenAI and Anthropic that they refuted a bedrock complexity conjecture are the part to treat as "interesting if it holds." Both are worth understanding, for different reasons.

The verifiable one: a 15,000-page proof, machine-checked

Axiom said it completed a full Lean auto-formalization of the Classification of Finite Simple Groups. If that name means nothing to you, the scale will: the original proof runs about 15,000 pages across roughly 500 papers by more than 100 mathematicians over decades. Very few humans understand all of it. Axiom's system turned it into code a proof checker can verify line by line.

The method is the interesting part. An agent swarm produced about 50 million lines of Lean in 24 days. Agents claimed tasks, reviewed each other's work, and merged through GitHub under human guidance, and in the process reportedly caught hundreds of errata in the published proof. Researchers note it displaces the four-color theorem as the largest computer-verified proof on record.

This is the kind of result that does not need you to take anyone's word for it. A Lean proof either checks or it does not, and the code is the artifact. It is also the least flashy of the day's claims, which is usually a good sign.

The loud ones: dueling claims to disprove the Exponential Time Hypothesis

OpenAI announced a batch of new results in harmonic analysis, complexity theory, and quantum field theory, including a claimed refutation of the Exponential Time Hypothesis, with materials posted to its openai/math repo. Hours later, Anthropic's camp said Claude independently disproved the same conjecture, finding the key algorithm on October 16 with "14 million output tokens and no human input," then refining it with researchers into a paper. Anthropic's formal-math repo has an eth directory to match.

The Exponential Time Hypothesis is a working assumption that problems like 3-SAT need roughly exponential time. A lot of fine-grained complexity theory is built on it, so refuting it would reshape a pile of conditional lower bounds and hardness assumptions. What it would not do is break your bank's encryption tomorrow; the fallout is theoretical, not an immediate security event. That distinction got lost in some of the hotter takes.

Both claims arrived mostly as announcements, repos, and screenshots rather than peer-reviewed papers, and the timing (two labs, the same conjecture, the same week) is exactly the kind of thing that deserves patience. The Lean formalizations give reviewers something concrete to check, which is better than a PDF alone. The right posture is to wait for mathematicians to work through them before calling it settled.

Yang-Mills, and the rest of the dump

There was more. OpenAI also claimed a natural-language proof related to Yang-Mills existence, dated weeks earlier and held back until the Lean formalization finished, plus a result on the Kakeya conjecture in higher dimensions. A Yang-Mills Working Group at KITP published a statement interpreting the AI result. One caution worth repeating from the people who actually work in the field: the Yang-Mills item looks like part of the Millennium Prize problem, not a complete solution of existence and the mass gap. Big if true, but not the whole prize.

Smaller results rounded out the day. Ten Claude agents proved in Lean that seven electrons on a sphere settle into a pentagonal bipyramid, a case of the Thomson problem. And a Stanford study had 29 experts rate research problems that an AI proposed; they judged about a third "interesting." That last one is a useful gut check. The models are generating real mathematical work, and also a lot of output that domain experts shrug at.

Video: the race to solve open math problems

If you want the backdrop on how frontier labs ended up pointing their biggest models at decades-old math problems, this walks through the race and the stakes.

Why it matters if you pay for AI

You are not going to prove a theorem with your ChatGPT or Claude subscription, so the direct effect on your week is zero. The signal is what these announcements say about where the top models are now capable, and about how the labs compete. When two rivals race to claim the same complexity result within hours, that is a marketing fight as much as a math one, and it tells you the frontier is close enough that bragging rights matter. Our OpenAI vs Anthropic vs Google rundown tracks how tight that race has gotten.

There is a quieter lesson too. The result you could actually verify, Axiom's Lean formalization, was the least hyped. The ones that came as screenshots got the biggest reaction. That pattern is worth carrying into how you read any model's claims about its own output, including the everyday ones. If you have followed AI's run at research math before, our look at Claude's Riemann hypothesis bound is the same story in miniature: genuine progress, oversold framing, and a gap that only checking closes. And because capability and value keep shifting between Claude Opus 5.5 and its rivals week to week, the sane way to pay for all of it is to not marry one vendor, which is what an AI aggregator is for.

FAQ

Did AI actually disprove the Exponential Time Hypothesis?

Two labs claim so, and both posted Lean formalizations, but the results had not gone through peer review when they were announced. Treat it as a strong claim with artifacts to check, not a settled fact, until mathematicians have worked through the proofs.

Would refuting ETH break encryption?

No. The Exponential Time Hypothesis underpins theoretical complexity results, like conditional lower bounds for exact algorithms. Refuting it reshapes theory; it does not hand anyone a practical attack on deployed cryptography.

What is the Classification of Finite Simple Groups, and why does the formalization matter?

It is a landmark theorem whose original proof spans about 15,000 pages across roughly 500 papers. Formalizing it in Lean means a proof checker can verify the whole thing, which is reassuring for a result almost no single person fully holds in their head.

So how much of this day is confirmed?

The Axiom formalization is a machine-checkable artifact. The ETH and Yang-Mills claims are announcements with repos attached and still need independent review. The biggest-day-ever framing came from an AI model reacting in the moment, not from the math community.

Sources

Further reading

Try all the models mentioned in this article

Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Try it for 7 days for $1.99.

Try Admix for $1.99

Related articles