Riemann Hypothesis: Anthropic’s Levent Alpöge Claims a Proof With Claude

Anthropic mathematician Levent Alpöge says he and Claude have proven the Riemann Hypothesis, math’s most famous open problem, and has posted a Lean proof that computers can check. His four-page announcement went up on Friday, November 20, 2026, the day after Anthropic’s IPO, and Scientific American broke the story the same day. As of publication, no independent expert has confirmed the proof, and the Clay Mathematics Institute still lists the problem as unsolved. This post explains what was claimed, what Claude reportedly contributed, and what has to happen before anyone can call the problem closed. We will update it when the verification lands. New to the problem? Start with this plain-English explainer of the Riemann Hypothesis and the claimed proof.

TLDR

Levent Alpöge, a number theorist who works at Anthropic, claims that he and Claude proved the Riemann Hypothesis and the much broader Generalized Riemann Hypothesis between early October and November 16, 2026. Instead of a readable paper, he released a formal proof written in the Lean proof assistant, saying the human-readable version currently runs past 250,000 pages. The key idea he credits to Claude is a chain of “theta lifts” whose length grows like the square root of the range of numbers being studied, a way to recover the “multiplicity” trick that solved the problem’s easier cousin over function fields. If the Lean files check out and the statement is correctly encoded, this would be the biggest result in the history of mathematics produced with AI, and possibly the biggest result in number theory, full stop.

Thoughts

The whole claim now rests on roughly one line of Lean code, not on the proof itself. A proof assistant like Lean is very good at answering one question: does every step follow from the last? It is useless at answering a different one: is the final theorem actually the Riemann Hypothesis? If the statement at the end of the archive is subtly weaker than the real conjecture, or if somewhere a definition of the zeta function or an axiom quietly assumes what it should prove, Lean will happily certify the wrong thing. That is why Scientific American’s framing, that experts should quickly be able to tell “whether the hypothesis is properly coded,” is exactly the right one. Alpöge says the arguments were “formalized without assumptions,” which is the claim that matters most. The audit is unusually tractable: read the final statement, list the axioms the proof depends on, rebuild it from scratch. Reviewers do not need to understand 250,000 pages of mathematics to do that, and that is the genuinely new thing here. For the first time, a claimed solution to a problem of this size can be checked before anyone understands it.

The timeline at the end of Alpöge’s note is the most revealing document in this story, more than the announcement itself. It covers about six weeks, from a “Stepanov determinant” on October 6 to “GRH” on November 16, and almost every hard step is dated November 15: the iterated theta lifts, the norm computations, the divisibility bounds, the seesaw identity, the gluing across Hecke characters, the fight with a weakened form of Vandiver’s conjecture. That is a single day in which a dozen steps, each of which could be a serious paper on its own, fell into place. Alpöge is also unusually candid about the division of labor. He chose the analogy and the toy setting; Claude supplied the technical mastery of automorphic forms on higher-rank groups, an area he admits he never learned. In a footnote he says that for parts where “Claude is the expert and not me” he is paraphrasing a Claude-written summary that “might make zero sense.” Taken at face value, that is a new model of what a mathematician does: supply the taste and the direction, then supervise a collaborator who out-calculates you.

The calendar is impossible to ignore: the announcement landed the day after Anthropic went public. Alpöge clearly knows it. On page two he writes that “the precise timing is horrible,” that the result came together late on a Sunday night, that he messaged Dario Amodei to ask whether he could announce, and that “there was a constraint to respect.” He does not name the constraint, but the likeliest reading is the pre-IPO quiet period, which limits what a company and its people can say in the final days before listing. He then says he “could not live if I waited even longer,” and adds that “there is an obvious cynical view, to do with the timing, but it is false.” The cynical view writes itself: one day after the listing, an employee of a newly public AI company announces that its model solved the most famous problem in mathematics. His defense is structurally sound, though. A stock-pumping stunt would want a glossy paper and a press tour; he released raw Lean files that anyone can run and that would humiliate him and the company within days if the statement were wrong. This is also not his first such claim. Scientific American notes that he posted a Claude-generated solution to the Jacobian conjecture in July. Whatever you think of the timing, the proof will be judged on the code.

The scale of the jump is hard to overstate. In August, Anthropic reported that Claude had helped raise the proven share of Riemann zeta zeros sitting on the critical line from 41.6% to 67.2%. That was a big result, and it was still worthless as evidence for the hypothesis, because even 99.999% leaves room for infinitely many exceptions. In October, OpenAI’s model produced a proof of the weaker “quasi-Riemann hypothesis,” which Quanta reported one Stanford mathematician called the most significant AI-for-math achievement so far. Now, a month later, the claim is not a better percentage or a weaker cousin but the full conjecture, plus the Generalized version that covers every Dirichlet L-function. Scientific American is right that no superlative fits. A huge body of published number theory begins with “assume GRH.” If this holds, all of it becomes unconditional at once.

Even if it all checks out, something is still missing, and Alpöge knows it. A proof nobody understands settles whether Riemann was right but teaches almost nothing about why. Why should a chain of roughly a thousand theta lifts tell you something about the primes up to a million? Mathematicians care about proofs because the ideas inside them get reused on the next problem, and right now those ideas live in a quarter million pages that Alpöge himself describes as unreadable. His note is full of confidence that simplification will come fast, that some of his choices (he singles out an appeal to Behrend’s construction) are “obviously going to be unnecessary,” and that the same toolkit has already shown him Navier-Stokes “was going to fall.” Those are extraordinary predictions on top of an extraordinary claim. The fair position for now is the one his own note models: astonishment, mixed with a clear admission that the understanding has not caught up with the result.

Key Takeaways

  • Levent Alpöge, a mathematician employed by Anthropic, announced on November 20, 2026 that he believes he has, working with Claude, proven the Riemann Hypothesis.
  • He says the method also proves the Generalized Riemann Hypothesis and expects it to reach the Extended Riemann Hypothesis with more time.
  • The proof was released only as Lean code, split into fourteen archive files, because the written version Claude produced runs to more than 250,000 pages.
  • No independent mathematician had confirmed the result at the time of the announcement, and the Clay Mathematics Institute still lists it as an open Millennium Prize Problem.
  • The central idea credited to Claude: chain theta lifts together, each one mimicking a layer of multiplicity, with the chain length set to the square root of X, the length of the prime sum being studied.
  • The argument descends from Stepanov’s method, which proved the analogue of the Riemann Hypothesis over function fields using multiplicity, a tool that had no obvious counterpart over the ordinary integers.
  • Alpöge’s timeline runs from October 6 to November 16, 2026, with most of the hardest steps dated November 15.
  • He openly says Claude was the expert on automorphic forms on higher-rank groups, an area he admits he never mastered.
  • The decisive check is whether the Lean statement really encodes the Riemann Hypothesis and whether the proof relies on hidden assumptions, not whether the individual steps compile.
  • In August, Anthropic reported Claude helped raise the proven share of zeta zeros on the critical line from 41.6% to 67.2%, a big gain that still said nothing about whether every zero lies on it.
  • The announcement came the day after Anthropic’s IPO. Alpöge says he consulted CEO Dario Amodei first, had “a constraint to respect,” and calls the cynical reading of the timing “false.”
  • He credits OpenAI’s recent math results for showing “great taste” and says one of them convinced him Navier-Stokes would fall too.
  • He promises to simplify the argument, write it up properly and lecture on it, and expects much shorter proofs to follow quickly.

Detailed Summary

What the Riemann Hypothesis Says, in Plain Terms

Count upward from one and climb a step every time you meet a prime. You reach 25 steps by 100 and 78,498 by one million. From far away the staircase looks like a smooth curve, and mathematicians have a formula for that curve, which predicts about 78,628 primes below a million. Bernhard Riemann’s 1859 question is how far the real staircase can stray from the smooth curve as you keep counting forever. He tied that error to the zeros of the zeta function, and his hypothesis says every nontrivial zero sits on a single vertical line with real part exactly one-half. Ten trillion zeros have been checked by computer and every one obeys. One exception further out would still break it.

Alpöge’s Announcement and the Generalized Riemann Hypothesis

Alpöge posted a short note on his personal site stating that he believes he has, working with Claude, found a proof. He goes further: the method “easily extends” to the Generalized Riemann Hypothesis, which makes the same prediction for the entire family of Dirichlet L-functions and underpins a huge body of conditional results across number theory. He says he and Claude formalized the arguments with no assumptions. Scientific American, in a breaking story by Joseph Howlett, called the conjecture the North Star of number theory and noted the claim would cap a year in which AI went from unreliable student to superhuman force in mathematics.

Multiplicity, Stepanov’s Method and Theta Lifts

Over function fields, the analogue of the Riemann Hypothesis was proven decades ago, and Stepanov found an elementary route that uses multiplicity: building an auxiliary object that vanishes to high order, the way (x − 3)² vanishes twice at 3. Over the ordinary integers there is no obvious notion of multiplicity, which Alpöge compares to differentiating whole numbers. His toy-setting work suggested a “natively automorphic” substitute: each theta lift between unitary groups wins one layer of multiplicity. Claude’s suggestion was to chain about √X of them for a prime sum of length X, so a range of one million would need about a thousand lifts. Each lift carries a cost, and the hard part was showing the gains survive the whole chain.

How Claude and Alpöge Split the Work

Alpöge says he had planned since his student days to spend decades on this problem and tried to learn every tool he might need. In the end, he writes, he steered them to the right analogue of a toy setting combining two very different ideas, and then Claude “far outpaced” him near the finish line. He freely admits he has little mastery of automorphic forms on higher-rank groups and that his descriptions of those parts are paraphrases of summaries Claude wrote. He also acknowledges he may have baked “strange suboptimal choices” into the argument.

Why He Released Only Lean Code

The full written argument is enormous. A footnote says it passed 250,000 pages, and Claude jokingly titled it the GRH “book,” with a Theorem 1 that begins by noting a real number is either greater than, less than or equal to one-half. Rather than post what he calls an unreadable travesty of “slop,” Alpöge chose to present only the Lean formalization, split across fourteen downloadable archive parts. Lean is a proof assistant in which every deduction is checked by software, so the files can be verified mechanically even before any human understands them.

The Six-Week Timeline From Determinant to GRH

A rough timeline that Alpöge says Claude compiled lists twenty steps. It starts on October 6, 2026, with the Stepanov determinant, an Iwasawa tower idea and the Bost slope inequality. Mid-October brings determinants for GGP periods and a relative trace formula. The Guinand-Weil explicit formula enters on October 20. Then on November 15 come iterated theta lifts between totally definite unitary groups, norm and divisibility computations using Chowla-Selberg, Weyl and the Siegel mass formula, the Rallis seesaw, a cost analysis of each lift, and a weakened Vandiver conjecture. The final entry, November 16, reads simply “GRH.”

The IPO Timing Question and Anthropic’s Role

The note was posted the day after Anthropic’s IPO. In a second page titled as a remark he felt obliged to make, Alpöge says the breakthrough happened late on a Sunday night and he messaged Dario Amodei to ask, hypothetically, about his ability to announce. He says there was “a constraint to respect,” that Anthropic’s leadership and the mathematicians he consulted held very similar views, and that he alone is responsible for refusing to wait any longer. He thanks a long list of people, including mathematicians Manjul Bhargava, Peter Sarnak and Maksym Radziwiłł; those are acknowledgments, not endorsements of the proof.

What Would Settle the Riemann Hypothesis Claim

Independent reviewers need to confirm three things: that the final Lean theorem states the Riemann Hypothesis in full, that the proof depends on no hidden axioms or unproven placeholders, and that the archive compiles. Scientific American expects that to take experts little time. After that comes the slower and more human work Alpöge has promised: simplifying the argument, writing it properly and explaining why a chain of theta lifts controls the primes. The Clay Mathematics Institute’s million-dollar prize has its own rules, which require publication and a waiting period, so any formal recognition is further off.

Notable Quotes

“I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.”

Levent Alpöge, opening his announcement

“No, Claude flew, and danced, and sparkled near that finish line.”

Levent Alpöge, on Claude’s role in the final stretch

“It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.”

Levent Alpöge, on what solving the problem means to him

“There is an obvious cynical view, to do with the timing, but it is false, and I am pleased that I am in circumstances where it is obviously false.”

Levent Alpöge, addressing the timing of the announcement

“Issues can arise in Lean, but it shouldn’t take long for experts to verify whether the hypothesis is properly coded, and if the proof is correct.”

Joseph Howlett, Scientific American, on how the claim will be checked

“There is no superlative strong enough for how this would transform mathematics, instantly unlocking a deluge of truths contingent upon the ambitious conjecture.”

Joseph Howlett, Scientific American, on the Generalized Riemann Hypothesis

“I hope the argument provides a glimpse towards the miracles that lie ahead, and shows us the joy and wonder we are about to find in the simplest things.”

Levent Alpöge, closing his announcement

Read Levent Alpöge’s full announcement here, and the Scientific American report here.

Related Reading