AI Checked Maths' Biggest Proof in 24 Days (2026)

There is a proof in mathematics so large that no single person has ever read all of it. It took roughly 100 mathematicians, more than 500 papers, and decades to write. On October 26, 2026, a Palo Alto startup called Axiom Math announced that swarms of AI agents directed by its engineers and mathematicians had translated the whole thing into a form a computer can check — about 43.1 million lines of code in the Lean proof assistant, produced in 24 days. And in the process, the machines found mistakes.

That last part is the story. Not that AI can do maths at unreadable scale — this week has already shown that twice over — but that a proof nobody could fully read turned out to contain errors nobody had caught, and that fixing them required new mathematics. Here is what Axiom actually did, what it found, and what it does not prove.

What is the "enormous theorem"?

The result is the Classification of Finite Simple Groups, usually shortened to CFSG. A "simple group" is, loosely, an atom of symmetry: a mathematical structure that cannot be broken into smaller symmetric pieces. The classification says that every finite simple group belongs to one of four kinds: cyclic groups of prime order, alternating groups on at least five symbols, sixteen infinite families of groups of Lie type built over finite fields, or one of exactly 26 isolated "sporadic" groups. The largest sporadic group, the Monster, was constructed by Robert Griess in 1980 using an algebra of dimension 196,884. Some authors count a 27th, the Tits group; Axiom's announcement notes the counting difference rather than hiding it.

The proof earned its nickname honestly. Walter Feit and John Thompson's 1963 proof that every non-abelian finite simple group has even order ran to 255 pages and filled an entire journal issue. Daniel Gorenstein laid out a sixteen-step programme to finish the classification in 1972 and declared it complete in 1983 — but one case, the quasithin case, rested on an unfinished manuscript, and it took Michael Aschbacher and Stephen Smith's 1,221-page treatment in 2004 to close what is now called the first-generation proof. A more systematic second-generation proof remains incomplete — which is why Axiom formalised the first-generation proof, while drawing its proof map heavily on the second-generation volumes.

For readers, the plain version is this: modern algebra rests on a list of symmetry atoms that was assembled by a century of human work, and until now nobody could verify the whole list mechanically. That is what changed.

What Axiom's agents actually did

Axiom's system is called AxiomProver. The company opened a project channel on October 1, assembled a proof map from the literature, and split the work into 108 milestones and about 380 smaller tasks. Agents could claim a task, find the relevant arguments, and translate them into Lean; when a task needed mathematics that had not been formalised yet, it could delegate those prerequisites to other agents. Lead agents broke down the hard arguments while others searched for existing results. The timeline Axiom published is unusually granular: the first swarm went live on October 2, a channel message on October 10 read "WE PROVED MONSTER IS SIMPLE," and on October 13 a single milestone closed — the point at which the team realised its milestones were too coarse and replanned the unfinished work into smaller, precisely stated theorems.

The headline numbers, all from Axiom's announcement and project dashboard: approximately 43.1 million lines of Lean, 2.05 million declarations, a corpus of 715 papers and books. Two honest discounts apply. About 1.2 million lines were inherited from an earlier formalisation effort, Qiuzhen-CFSG, and the whole thing builds on Mathlib, the community library that hundreds of contributors built. Even after those discounts, roughly 42 million lines of new Lean appeared in about three and a half weeks.

The holes: 2,000 errors, 34 that mattered, one genuine gap

Formalisation forces every implicit step into the open, and CFSG's literature had plenty of them. Axiom reports correcting over 2,000 errors, most of them minor — typos, table errors, dropped hypotheses. Thirty-four survived filtering as noteworthy deviations from the published record.

The largest sat in the Unbalanced Group Theorem (UGT), a piece of the argument that identifies a whole group from information about its subgroups. Axiom was formalising UGT through John Walter's 1986 memoir when its agents found explicit counterexamples to a closure assertion Walter's proof relied on: the argument assumed a particular operation kept the associated simple groups within a specified family, but the operation could produce a group explicitly excluded from that family. The concrete example is checkable by hand: start with PSL₂(49), which belongs to Walter's family; its field involution has centraliser PGL₂(7), whose simple component PSL₂(7) falls outside the family because 7 is a Mersenne prime.

Two things about this gap matter, and both are easy to get wrong. First, it does not overturn the classification. Axiom states explicitly that the counterexamples did not refute the goal of UGT, or even the conclusion of Walter's Proposition II.5.9(ii) — they left one inference unjustified, which is a different and smaller problem. Second, Axiom did not paper over it. With group theorists Richard Lyons and Ronald Solomon consulted, its prover produced and formalised an alternative argument built on Solomon's 1976 theory of maximal 2-components. The new result, Theorem 2.3 ("Preservation of simple tops"), now stands as a piece of group theory in its own right — and after the formalisation, Axiom's mathematicians found a shorter first-generation route that skips it entirely, recorded as Remark 9.5 in the accompanying paper. A smaller finding points the same way: a lemma in Aschbacher and Smith's quasithin volumes asserted two quantities were equal, and the agents found a counterexample in SL₂(5) where one quantity is 1 and the equality requires 2.

The caveat that matters more than the headline

A machine check proves that the encoded deductions follow from the encoded assumptions. It does not, by itself, prove that the formal statement is the theorem mathematicians meant. Axiom says this itself: humans inspect the top-level statement and its definitions to confirm it expresses the intended classification, while Lean checks the formal deductions. That human confirmation is the part no agent fleet can outsource, and it is the honest limit of every "AI verified X" headline this week.

The project's failure modes are the most transferable part of the story — and the most useful warning for anyone running AI agents on real work. Some milestones concealed enormous prerequisites, which agents had assumed rather than proved. Axiom deleted tens of thousands of lines built on those assumptions and restarted parts of the work on a sounder plan, noting that "a recurring difficulty was determining whether adding an assumption to a theorem deferred work or changed its meaning." That is the multi-agent version of a test suite passing against a mock: locally valid, globally lethal. Milestone counts lied for the same reason — the formalisation was complete only when the pieces formed a single proof with no unjustified assumptions remaining.

The other bottleneck was verification itself. Agents produced proofs faster than Lean could recheck them, and changing a shared result could force Lean to recheck all its dependents. Axiom's answer was infrastructure, not prompting: distributed caching of build results, parallel independent checks, and expensive finite computations externalised into separately checkable certificates. In any self-checking system, verification throughput — not generation speed — sets the ceiling.

Why this matters beyond mathematics

Every paper that invokes CFSG in one sentence now rests on an argument that can be inspected step by step, with the unresolved literature documented rather than inherited. That is a durable artefact, and it arrived in under a month of wall-clock time. But the wider lesson is about trust in machine-checked work generally: a passing verifier certifies that the machine did what the specification said, not that the specification was what you wanted. In mathematics that gap is benign and openly managed. In an agent fleet handling money, infrastructure or safety-critical code, the same gap is where incidents live.

It also reframes what "AI does maths" means. The scarce human skill did not disappear; it moved upstream, to writing the specification, reviewing the formal statement, and deciding when an assumption changes a theorem's meaning. That is the same shift visible across this week's AI news, from Google's Gemini 4 Argon handling long-horizon engineering work to open-weight models like Kimi K3.1 Pro closing the gap with closed frontier systems — the models generate, humans specify and verify.

The same week in AI maths — and today's security thread

Axiom's announcement landed in the same week as OpenAI's October 26 science claims, and the two should not be conflated. OpenAI says one of its internal models produced a claimed refutation of the Exponential Time Hypothesis on October 16 and formalised it in Lean the next day; Anthropic published a separate, differently-argued ETH claim on October 27. But OpenAI's own Lean notes say the machine-checked time bound is weaker than the sharper bound in its manuscript, the outside reviewers — Russell Impagliazzo and Ramamohan Paturi of UC San Diego and Ryan Williams of MIT — wrote that their involvement "should not be taken as certification of the claim," and the physics write-up (a claimed existence result for quantum field theories) is still withheld by its review group. Axiom formalised a settled theorem with a public artefact; the ETH claims are new results still awaiting independent scrutiny. Keep the two in separate boxes.

One more thread from today, briefly: AI security had its own split day. Google unveiled Gemini 4 Carbon, a cybersecurity-focused frontier model, but it is being made available first only to a subset of defenders through Google's Fairwind Program while safeguards are tested — it is not publicly available, and its benchmark claims are Google's own, not independent verification. Meanwhile Reuters reported that Z.AI's open-weight GLM-5.3 completed 3 of 13 multi-stage cyberattack scenarios in controlled simulations run by security firm Irregular — the strongest result Irregular has recorded from an openly available model — while also repeatedly reproducing code from memory despite instructions not to. Those were controlled simulations, not real attacks, and Reuters notes the tests were not identical to the separate assessment in which OpenAI's GPT-6 Astra completed 9 of 10 scenarios, so the two scores are not directly comparable. Defenders get a gated model; open-weight models show the dual-use risk. Both stories, like Axiom's, come down to the same question: who checks the checker.

The bottom line

A proof too big for any human to read has now been checked by machines, and the machines found real mistakes — including one gap serious enough to need new mathematics to close. That is a genuine milestone, and the public Lean artefact means it can be inspected rather than merely believed. But the honest version of the story has three tiers, and they should not be collapsed: the classification itself is established mathematics; the Lean formalisation is machine-checked; and whether the formal statement captures exactly the theorem mathematicians meant remains a human judgement, confirmed by reviewers rather than by the machine. Anyone who tells you AI "proved" something new this week, or that the classification was wrong, is selling you a simpler story than the one that actually happened.

Sources: Axiom — Formalizing the Enormous Theorem · The Agent Report — Axiom's agent swarm · Unite.AI — Axiom's 43.1M-line Lean proof · CryptoBriefing — Axiom formalizes the enormous theorem · OpenAI's ETH claim — what has actually been verified · AI Weekly — OpenAI hands ETH, Kakeya, Yang-Mills proofs to reviewers

Recommended Reading

If this story has you curious about the mathematics behind it, you can browse books on group theory on Amazon.

As an Amazon Associate, AI Trends Daily earns from qualifying purchases.