On Friday 20 November 2026, a mathematician at Anthropic posted a four-page note on his personal website with a sentence that travelled around the world within hours: he believed he had, working with Claude, discovered a proof of the Riemann Hypothesis — the most famous unsolved problem in mathematics, open since 1859. There was no paper — only a 14-part archive of about 17.4 million lines of Lean code, a claim the same method also settles the stronger Generalised Riemann Hypothesis, and an author who says openly that he does not yet fully understand every argument Claude used. This is what was actually claimed, what has been checked so far, and what is still unproven.
What was actually claimed
The claim is bigger than one problem, and it arrives as code, not as a paper. Levent Alpöge, a mathematician on Anthropic’s technical staff and a junior fellow at the Harvard Society of Fellows, wrote in a note titled ‘Remark’: ‘I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.’ The note, published on 20 November 2026, says the method also resolves the Generalised Riemann Hypothesis, which extends the same idea to a whole family of related functions — and that Alpöge hopes to reach the even broader Extended Riemann Hypothesis with more time.
The scale of what was released is itself part of the story. The proof exists only as Lean code: 175,187 files holding roughly 17.4 million lines, split across 14 compressed parts totalling about 239 MB and expanding to more than 61 GB. Alpöge says a written version would run to more than 250,000 pages, so he released the short note and the machine-checkable code instead. The decisive stretch ran from 6 October to 16 November — 41 days — with the key steps dated 15 November and the Generalised Riemann Hypothesis entry dated 16 November. The work was built with a non-public version of Claude over about six weeks, according to the BBC.
Alpöge describes the collaboration plainly: he directed the effort toward a promising approach, and Claude supplied major advances as the argument developed — including, by his account, a central step. “Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line,” he wrote. “No, Claude flew, and danced, and sparkled near that finish line.” He also says he is presenting a complete argument before having mastered the mathematics around it, and that he intends to write it up properly and expects much shorter proofs to follow.
What the Riemann Hypothesis actually says
It is a claim about how prime numbers are distributed — and a proof would tighten, not break, much of mathematics. Prime numbers — 2, 3, 5, 7, 11 and so on, divisible only by 1 and themselves — are the building blocks of all whole numbers, but they thin out irregularly as numbers grow. Mathematicians have known since the prime number theorem of 1896 roughly how many primes to expect below any given number, but only loosely: the theorem does not tightly constrain how far the actual count can deviate from the estimate.
In 1859, Bernhard Riemann found a surprising connection between that deviation and the zeta function, which adds together an infinite series of fractions. The hypothesis says that all of the zeta function’s “non-trivial” zeros — the values at which the function equals zero, other than at negative even integers — have a real part exactly equal to one half. Proving it would pin down those deviations far more tightly, and would turn into theorems in their own right hundreds of results that currently assume the hypothesis is true. It is one of the seven Clay Mathematics Institute Millennium Prize Problems, carrying a $1 million bounty.
Two common misunderstandings are worth clearing up at once. First, computers had already checked the first 10 trillion zeros and found them all on the critical line — but the next zero could still have defied the hypothesis, which is why only a proof, not a check, settles it. Second, contrary to popular accounts, a proof would not reveal the location of every prime number or break encryption systems.
How Claude supposedly did it
The reported route adapts an old technique to new ground, with Claude supplying the missing step. The solution starts with Stepanov’s method, associated with the Russian mathematician Sergey Stepanov, which works in the world of function fields — a simpler setting where the equivalent result is already known. Applying it to ordinary integers had proved difficult because one part of the technique had no obvious equivalent there. Alpöge says Claude found a way through by combining concepts from several branches of mathematics.
The key move, by Alpöge’s account, was Claude’s suggestion: a long chain of transformations called theta lifts — about the square root of the summation length, stacked one on another — to recover an effect the note calls multiplicity, the step that had been missing. Terence Tao, who published a preliminary analysis of the claim, wrote that the proof appears to draw comprehensively on the Langlands Program and arithmetic geometry, and that verification will take some time.
This is also not the first AI-maths headline of the year. In July, Alpöge was involved in an AI-assisted disproof of the Jacobian conjecture; in September, OpenAI said its internal AI models had solved the Navier-Stokes problem, another Clay Millennium problem — an announcement overshadowed by a dispute over unpublished research. In October, OpenAI released 722 AI-generated manuscripts covering 372 groups of results, one claiming a proof of the weaker “quasi-Riemann hypothesis”. The Riemann claim is the third headline AI-maths claim in five months — and the one with the strongest machine-checking behind it so far.
What has been checked so far: claimed vs verified
The code compiles. That is a serious result — and it is not the same as a proof being accepted. Andrew Sutherland, a computational number theorist at MIT, has been running independent checks on high-performance clusters and publishing the results on a public verification dashboard. Here is where each part of the claim stands as of 21 November 2026:
| Claim | Status on 21 Nov 2026 |
|---|---|
| The Lean code is complete and error-free | ✅ Verified by machine: all 175,187 modules compile with zero errors under Lean 4.33.1 |
| The proof relies on no hidden assumptions | ✅ Verified by machine: the five headline theorems use only Lean’s three standard axioms (propext, Classical.choice, Quot.sound) |
| Independent checkers accept the proof | ✅ Verified by machine: nanoda (an independent Rust kernel) accepted the full export in 27 minutes 42 seconds; con-leche accepted it in 3 minutes 15 seconds |
| The theorems state the actual Riemann Hypothesis | ✅ Verified by machine: the statement comparison passed, and mathematician Elliot Glazer confirmed the Lean file states the Generalised Riemann Hypothesis |
| Lean’s own final checks are complete | ⏳ Still running: the official lake comparator run and the full Lean-kernel replay were still underway at the last dashboard update (21 Nov, 10:05 UTC) |
| A human expert has read and understood the proof | ❌ Not done: no independent expert has reviewed the argument; Tao says digestion “will take some time” and experts say understanding it could take years |
| The Clay Mathematics Institute accepts it | ❌ Not started: the problem is still listed as unsolved on the Clay site; nobody has claimed the $1 million prize |
What the machine checks establish is precise and limited: the formalised statements follow from the specified mathematical definitions and Lean’s standard axioms, with no hidden assumptions smuggled in. What they cannot establish is why the argument works — a compiling Lean file cannot explain itself. Sutherland told Scientific American: “Honestly the chance that anything is fishy here is approximately zero.” Fields Medal winner James Maynard told Nature: “If this fully checks out, it is clearly a historic milestone in mathematics.” Both quotes are conditional in exactly the way that matters: if it fully checks out.
The rival claim that arrived the same day
Hours after Alpöge’s note went up, a second, entirely human proof claim appeared — and it is already under heavier fire. Ravi Vakil, president of the American Mathematical Society, posted a draft Generalised Riemann Hypothesis proof by Alessandro Maria Masullo, a Stanford researcher, described as entirely human-developed, without AI assistance. Brian Conrad had reportedly been reviewing Masullo’s draft for months. But where Alpöge’s code has cleared its early machine checks, Masullo’s draft has encountered questions from mathematicians examining it, and it has no machine-checkable formalisation behind it. The two claims are now an accidental experiment: the same problem, attacked the same day, by opposite methods — and so far the machine-checked one is holding up better.
The awkward timing
The announcement landed the day after Anthropic’s stock market debut — and Alpöge knew how that would look. Anthropic began trading on the Nasdaq on 19 November 2026 after pricing its IPO at $83 a share, the largest initial public offering in history. On Friday morning, the day after, Alpöge hit publish. He said he had consulted Anthropic CEO Dario Amodei before announcing, and acknowledged in his note that the timing could be viewed cynically, without elaborating. Anthropic did not respond to a request for comment from the Associated Press.
The timing is a legitimate question and a limited one. It explains why the announcement looks the way it does; it says nothing about whether 17.4 million lines of Lean code compile, which is a question machines have now answered. It is also worth noting the claim did not come from nowhere: in August, Anthropic reported that a previous AI-assisted attempt had shown at least 67.2% of all non-trivial zeros lie on the critical line, up from 41.6% earlier — a result human experts were able to verify. The Riemann claim is the same research programme, pushed to its limit.
What happens next
Three gates stand between this claim and a theorem, and the slowest one is human understanding. First, the remaining machine checks — Lean’s comparator run and the full kernel replay — need to finish cleanly; they were still running on 21 November. Second, mathematicians need to digest the argument itself, a process Tao has already warned will take time, with some experts saying understanding could take years. Third, the Clay Mathematics Institute’s process requires the proof to be published as a paper in a qualifying publication, come to be generally accepted by the mathematics community over two years, and survive scrutiny by a bespoke committee of mathematicians. Only then does the $1 million prize move.
Until those gates are passed, the honest summary is the one the evidence supports: a serious, machine-checked claim that most experts who have looked at it believe is probably right — and not yet an accepted theorem.
Our take
This claim inverts the normal order of mathematics: correctness first, understanding later — and that is the real story. For 167 years, a proof meant a human argument a human could follow. Here, the correctness check may finish before any human understands why the argument works, because the argument exists only as code no person has read end to end. If the remaining checks pass, mathematics will have a result it can trust but cannot yet explain — and a new kind of peer review to invent, one that can audit a 250,000-page argument by machine. That is a bigger shift than any single prize problem. The test to watch is simple: when the comparator finishes, when the first mathematicians publish a human-readable outline, and when the Clay process begins. Until then, call it what it is — claimed, machine-checked, unverified.
FAQ
Did Claude prove the Riemann Hypothesis?
Not yet. Anthropic mathematician Levent Alpöge claims he and Claude found a proof, and the 17.4-million-line Lean formalisation compiles without errors and has passed checks by independent proof-checking systems. But no independent expert has reviewed or understood the argument, final machine checks were still running on 21 November, and the Clay Mathematics Institute still lists the problem as unsolved.
What is Lean, and why does it matter here?
Lean is a proof assistant: a programming language for writing mathematical arguments so that a small program called the kernel can check every step against explicit rules. A proof written in Lean can be verified by computer even when no human has read it — which is exactly why Alpöge released his proof as Lean code instead of a 250,000-page paper.
What is the Generalised Riemann Hypothesis?
It extends the Riemann Hypothesis from the zeta function to a whole family of related functions called Dirichlet L-functions. Alpöge’s note claims his method settles the Generalised Riemann Hypothesis too, and he says he hopes to reach the even broader Extended Riemann Hypothesis with more time.
Would a proof of the Riemann Hypothesis break encryption?
No. A proof would pin down how primes are distributed, but it would not reveal the location of every prime or provide a fast way to factor large numbers, which is what modern encryption relies on.
When will we know for sure?
The remaining machine checks should finish within days. Human understanding will take far longer — Terence Tao has said verification will take some time, and experts say digesting the argument could take years. The Clay Mathematics Institute’s formal process then requires general acceptance over two years before the $1 million prize can be awarded.
Sources and methodology
Claims in this article were verified on 21 November 2026 against: phys.org’s report on the announcement (20 Nov); The Hindu’s explainer on the claim and its checks (21 Nov); CNN’s report on the announcement and the rival claim (21 Nov); Moneycontrol’s comparison of the two claims (21 Nov); the CASRAI summary of the Lean archive release (21 Nov); and a developer’s walkthrough of what the Lean checks actually cover, based on Andrew Sutherland’s public verification dashboard (21 Nov). Direct quotations are from Alpöge’s announcement note and from Sutherland and Maynard as reported by Scientific American and Nature. This article is desk-researched, not hands-on tested: no Lean environment was run for this piece, and every machine-check result is attributed to the mathematician or dashboard that reported it. Where sources conflicted or a check was still running, this article says so rather than picking a side.
Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


