Inter-office memo
- To
- Everyone following what AI can prove
- From
- Mo, Chief Executive
- Date
- November 20, 2026, 3:16 PM CST
- Re
- The Riemann Hypothesis
An Anthropic mathematician says Claude helped prove the Riemann Hypothesis. Is it right?
On November 20, 2026, Anthropic mathematician Levent Alpöge said that “working with Claude” he proved the Generalized Riemann Hypothesis, and with it the Riemann Hypothesis, in 17.4 million lines of Lean. MIT’s Andrew Sutherland says he compiled it; checks of its axioms hadn’t finished, and Anthropic hadn’t commented. Hours later, Ravi Vakil posted a separate, “all human” proposed proof by Alessandro Masullo.
What Alpöge announced
On November 20, Levent Alpöge, a mathematician who works at Anthropic, by his home page, posted a four-page note. It opens with the claim: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis. In fact, the method easily extends much further, resolving the so-called Generalized Riemann Hypothesis, and I expect, with more time, the Extended Riemann Hypothesis too. Claude and I have also formalized the arguments without assumptions.”
The proof is released only in Lean, a language in which a computer checks every step, as an archive in 14 parts on his site. Alpöge writes that a prose write-up would be “a huge, unreadable document”; his own ran to “250,000+ pages,” and “it being obviously inappropriate to attempt to post some massive unreadable travesty of ‘slop’, I decided to only present Lean.” A mirror on GitHub counts 175,187 Lean files and about 17.4 million lines.
His account of who did what is specific. He chose the approach: “I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.” Near the end, “Claude far outpaced me.” A timeline he says Claude assembled starts on October 6, 2026 and ends on November 16 with one word: “GRH.” He says he hasn’t caught up with it yet. He is “presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics,” and plans “to write it properly.”
The note doesn’t say which Claude. Terence Tao, summarizing the day, says Alpöge used “an internal AI model.” Anthropic hasn’t named the model or commented on the claim; its newsroom had no post about it on the afternoon of November 20.
What the Riemann Hypothesis says
The Riemann zeta function is zero at every negative even number. Those are the trivial zeros. The Riemann Hypothesis, from 1859, says every other zero has real part exactly 1/2, so they all sit on one vertical line. That line controls how evenly the primes are spread: if it holds, the count of primes up to any number stays within a standard error term of what a simple random model predicts, as Scott Aaronson puts it. The first 12 trillion zeros have been checked numerically, Quanta notes, but no one had proved it for all of them. It is one of the Clay Mathematics Institute’s seven Millennium Prize Problems.
The Generalized Riemann Hypothesis says the same of a much larger family of functions. In Alpöge’s Lean files, that family is the L-functions of Dirichlet characters. The headline theorem in Main.lean proves Mathlib’s own statement of the original hypothesis: a definition from Mathlib, Lean’s human-curated math library, not one the project wrote. Mathlib’s comment above it says “constructing a term of this type is worth a million dollars”:
def RiemannHypothesis : Prop :=
∀ (s : ℂ) (_ : riemannZeta s = 0) (_ : ¬∃ n : ℕ, s = -2 * (n + 1)) (_ : s ≠ 1), s.re = 1 / 2
The other four theorems, one restating the hypothesis for zeta and three about Dirichlet L-functions, use definitions in a 16-line Statements.lean that imports only Mathlib, per the mirror.
How sure anyone is
Lean is the reason to take this seriously, and it isn’t the end of the checking. Kevin Buzzard’s guide to AI-written Lean, from November 12, lists what has to happen before “it was checked in Lean” means much: confirm the statement says what it claims, compile the code, confirm nothing in the proof files redefines terms or adds axioms (Lean’s comparator tool does this), and rule out bugs in Lean’s kernel by replaying the proof in independent checkers. Where this one stands on the afternoon of November 20:
- The statement. The Riemann Hypothesis part uses Mathlib’s definition, and the rest sit in a file that imports only Mathlib, which Buzzard says is how “we know we can trust” a statement. The Dirichlet statements are short and readable.
- Compiling. “This is indeed an historic moment,” MIT’s Andrew Sutherland told Scientific American “after he had independently compiled Alpöge’s result in Lean.”
- Axioms and hidden tricks. A text search of all 175,187 files found no
sorry, noaxiomdeclarations and nonative_decide, both the mirror and an independent check report, but that is no substitute for asking Lean. That check’s full build started at 16:12 UTC on 256 cores and expects to need “roughly 20 hours”; its axiom check is pending and its comparator and independent-kernel runs are planned. - Understanding. No one has a readable argument. Tao calls the Lean Zulip thread “probably the best place to follow the efforts to verify the proof and extract the strategy of proof,” which, “from very initial analysis, appears to use the full power of modern developments on the Langlands program and arithmetic geometry.”
Aaronson’s bet is “that the proof will stand,” and he adds that mathematicians “won’t be satisfied with 17 million lines of Lean” until they can reconstruct the ideas themselves, which “will probably take years.”
A human proof, the same afternoon
At 18:18 UTC, Ravi Vakil posted a separate proposed proof of the Generalized Riemann Hypothesis, by Alessandro Masullo. “It is all human, with no AI,” he wrote. Brian Conrad “has been looking closely at it for months and gave extensive feedback,” and “the hope was to bring Alessandro to Stanford for a year to polish the writing and then announce it.” Instead, “Now it seems best to put out the current draft. I hope people quickly check it, formalize, fill any small gaps if they occur.”
The draft, The Music of the Primes, is 940 pages. It has no Lean proof yet; a thread on formalizing it opened on the Lean Zulip the same day. “There is a long story to tell,” Vakil wrote, without telling it yet, and no one has said whether the two proofs share any ideas.
Navier–Stokes and Euler, the same morning
Alpöge also has his name on two other results posted on November 20, both about whether fluid equations can “blow up” in finite time. Tao lists them as more digested than the Riemann claim:
- Euler, with no external force, by Alpöge, Tristan Buckmaster and Matei Coiculescu, “with heavy AI assistance,” in Tao’s words. The repository has a paper submitted to arXiv and a Lean proof the authors say contains no
sorryand uses only Lean’s three standard axioms. - Navier–Stokes, with no external force, by Alpöge and Buckmaster, “again with heavy AI usage,” Tao writes. OpenAI’s September proof of blowup “technically satisfied the conditions” for Clay’s prize, Scientific American says, but it “imagined an unrealistic, finely tuned force,” which many considered a loophole; this one has no force at all. Buckmaster wrote that “After last time, we couldn’t risk waiting, so the Navier–Stokes write-up is unpolished and heavily LLM-assisted.” The repository says the certificate the paper relies on “is not in this repository”; the paper says it will be released, and the interval-arithmetic computation that checks it isn’t part of the Lean formalization.
The day’s results, side by side
| GRH, Alpöge | GRH, Masullo | Navier–Stokes | Euler | |
|---|---|---|---|---|
| Authors | Levent Alpöge | Alessandro Masullo | Alpöge, Buckmaster | Alpöge, Buckmaster, Coiculescu |
| AI’s part | “Working with Claude”; “an internal AI model,” per Tao | “All human, with no AI,” per Vakil | Write-up “heavily LLM-assisted” | “Heavy AI assistance,” per Tao |
| Readable proof | None; Lean only | 940-page draft | 79-page paper, “unpolished” | Paper submitted to arXiv |
| Lean | 17.4 million lines | Not yet | About 600,000 lines; certificate not released | Complete, by the authors’ account |
| Checked by others | Compiled by Sutherland; axiom checks pending | Brian Conrad, for months | Not yet reported | Not yet reported |
Does anyone win $1 million?
Not soon. The Clay Mathematics Institute’s rules say it “does not accept direct submission of proposed solutions” and considers one only when three conditions are met: it is “published in a Qualifying Outlet,” “at least two years must have passed since publication,” and it “must have received general acceptance in the global mathematics community.” Alpöge’s proof is a Lean archive on his website; Masullo’s is a draft on a blog.
What it changes, and what it doesn’t
Mathematics, a great deal, if it holds. “If the proof is correct and understood, this completes a dictionary between different branches of math where the connections were only partial,” Hector Pasten told Scientific American. “Something like a Rosetta Stone.” Many theorems have been proved on the assumption that the Generalized Riemann Hypothesis is true; a proof makes them unconditional. It also caps a fast run. In August, Anthropic reported that an unreleased Claude, asked to “take a real stab” at the hypothesis in Claude Code, raised the share of zeros known to be on the line from 41.6% to 67.2%, adding, “We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis.” In October, OpenAI announced a proof of a weaker “quasi-Riemann hypothesis,” putting the real part of every nontrivial zero between 1/8 and 7/8, Quanta reported.
Not your encryption. RSA depends on how hard it is to find a huge number’s prime factors. The Generalized Riemann Hypothesis helps decide whether a number is prime, a different task, so it “wouldn’t provide a known shortcut,” Scientific American explains.
Not the Claude you use. Anthropic announced no model, plan or price change. Its most capable public model is still Claude Fable 5.2, at $8 per million input tokens and $40 per million output, and it recommends Claude Opus 5.5 for most work.
The timing. The note came a day after Anthropic’s stock began trading, a point Alpöge raises himself: “The precise timing is horrible.” He writes that it came together “late Sunday night,” that he “messaged Dario,” eventually “asking about my ability to announce,” and that “there was a constraint to respect.” “There is an obvious cynical view, to do with the timing, but it is false.”
Checking it yourself
Building it takes a large machine and a sandbox. Buzzard warns, “do not run random Lean code on your local machine unless you sandbox it first,” and when he tried Anthropic’s 13-million-line formalization of Fermat’s Last Theorem, he ran out of memory until he had “a machine with 96 cores and 500 gigs of ram.” With Lean’s installer, elan, the mirror’s README gives the steps, starting from Alpöge’s original parts:
for i in $(seq -w 1 14); do curl -LO https://alpo.ge/lean.tar.gz.part$i; done
cat lean.tar.gz.part{01..14} > lean.tar.gz
shasum -a 256 lean.tar.gz # 9631b4f020199e580ba86178cb56c6e9b09f690c2369a80dc4334aa93f49c7fc
tar -xzf lean.tar.gz && cd lean
lake exe cache get
lake build Statements # quick: the statements only
lake build Main # the whole proof
Then #print axioms riemannHypothesis in a file that imports Main should list only propext, Classical.choice and Quot.sound. The README ends with a procedure written for coding agents, so you can hand the job to Claude Code or Codex, inside a sandbox, and have it log time and peak memory. Reading Statements.lean and Main.lean takes a minute and needs no build.
What hasn’t been said
- Which model did the work, how long it ran, and whether it will be released.
- Whether Anthropic stands behind the claim. The company hasn’t commented.
- When a readable proof comes. Alpöge says he intends “to speak about it, to write it properly,” with no date.
- How Masullo’s and Alpöge’s proofs relate, if at all, and why Masullo’s draft came out the same day.
- When the Navier–Stokes certificate is released.
We’ll update this memo as the checks finish.
11.Questions.
Did AI prove the Riemann Hypothesis?
- Possibly, and no one has confirmed it yet. On November 20, 2026, Anthropic mathematician Levent Alpöge said he found a proof “working with Claude” and released it as 17.4 million lines of Lean, a language a computer checks step by step. MIT’s Andrew Sutherland says he compiled it; the checks of its axioms and statement, and any human understanding of the argument, are still to come.
Which Claude model proved the Riemann Hypothesis, and can I use it?
- Nobody has said. Alpöge’s note says only “Claude,” Terence Tao says Alpöge used “an internal AI model,” and Anthropic hasn’t named it or said whether it will be released. The most capable Claude anyone can use is Claude Fable 5.2, released on November 9.
Is the Riemann Hypothesis proof verified?
- Not yet. Andrew Sutherland told Scientific American he compiled it in Lean, and its headline theorem uses Mathlib’s own definition of the Riemann Hypothesis. But on the afternoon of November 20, a public independent build was still running, its axiom check pending, and no one had published a review of the argument itself.
Will Alpöge win the $1 million Millennium Prize?
- Not soon. The Clay Mathematics Institute considers a solution only after it’s published in a qualifying outlet, at least two years have passed, and it has “general acceptance in the global mathematics community.” Alpöge’s proof is a Lean archive on his website, and Masullo’s is a draft.
Does proving the Riemann Hypothesis break encryption?
- No. Scientific American explains that the Generalized Riemann Hypothesis helps decide whether a number is prime, which is different from finding a huge number’s prime factors, so it “wouldn’t provide a known shortcut” against RSA.
Who is Alessandro Masullo?
- The author of a second, separate proposed proof of the Generalized Riemann Hypothesis, a 940-page draft called “The Music of the Primes.” Ravi Vakil posted it on November 20, hours after Alpöge’s note, writing that “It is all human, with no AI” and that Brian Conrad had been checking it for months.