Blog · Opinion · 20 November 2026

A machine that proved the Riemann hypothesis, or at least we think it did

A mathematician at Anthropic says he and Claude proved the Riemann hypothesis. The proof is 17.4 million lines of Lean that nobody has read, it compiled on other people's machines the same day, and the file that says what was proved is 16 lines long.

The Riemann hypothesis has been proved by a mathematician working with a language model, or at least we think it has been. On Friday 20 November 2026 Levent Alpöge, a mathematician at Anthropic, put a four-page note on his own website that says "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis" [1] [4]. There's no paper with it, only 239 MB of code in Lean, a language in which a small program checks every step of a proof, and by the evening two people outside Anthropic had compiled all of it on their own machines and it passed [2] [4].

For anyone who hasn't met it, Bernhard Riemann wrote the conjecture down in 1859. It says where a function called the zeta function equals zero, and people care because those zeros govern how the prime numbers thin out as you count upwards [7]. It carries a million-dollar prize [6], computers had checked the first 12 trillion zeros and found every one on the line Riemann said they'd be on (holy shit) [7], and for 167 years nobody could prove it for all of them.

16lines in the file that states what was proved
17,374,215lines of Lean in the proof, in 175,187 files
41 daysfrom the first dated step in the note's timeline, 6 October, to the entry that reads GRH, 16 November
250,000+pages the argument came to when Alpöge had Claude write it out for people, which he didn't post
The proof as it was posted on 20 November 2026, from the note [1] and the archive [2].

What "we think" rests on

Andrew Sutherland at MIT compiled the whole archive first, shortly before noon. Simon DeDeo at Carnegie Mellon did the same a few hours later, and Sutherland then compiled it again on a variant of Lean, which makes it unlikely that the proof leans on a bug in the checker itself [4]. A slower check, a tool called a comparator that makes sure the proof hasn't redefined a symbol to make something trivially true, was still running on the evening of 20 November, and Quanta expected it to take hours or days [4]. DeDeo told Quanta he'd be very surprised if it turned up any tricks [4], and so would I, but that check hasn't finished.

Nobody has read the proof, Alpöge included. He writes that he's giving the world a complete argument before he has mastered the mathematics around it, that when he had Claude write it out for people it ran past 250,000 pages, and that he thought posting that would be wrong (I wouldn't have read it either), so he posted only the Lean [1]. Terence Tao listed it last of three results that day, in order of how well each one has been digested, and called it an "(enormous) Lean-only proof of the Generalized Riemann Hypothesis" made with an internal model [3]. From a first look he thinks it draws on the whole of modern work on the Langlands program and arithmetic geometry, and he expects the digesting to take some time [3].

The 16 lines

We downloaded the archive on Friday night, 14 files that join into one, to see what it says was proved [2]. Unpacked it's about a gigabyte of text, and a plain search of all of it finds 725,147 lines that open a theorem, no sorry (Lean's word for a step left unproved) and no line that declares a new axiom [2]. A text search doesn't prove anything, and the compile and the comparator are the real checks.

The file that says what was proved is 16 lines long, the main file is 12, and this is the main file's first theorem [2].

theorem riemannHypothesis : RiemannHypothesis := rhOfAll

RiemannHypothesis there comes from Mathlib, the Lean community's shared library, at the revision the project pins, and it's three lines [11].

@[wikidata Q205966]
def RiemannHypothesis : Prop :=
  ∀ (s : ℂ) (_ : riemannZeta s = 0) (_ : ¬∃ n : ℕ, s = -2 * (n + 1)) (_ : s ≠ 1), s.re = 1 / 2

In words, every zero of the zeta function that isn't one of the trivial ones has real part one half. The other four theorems in the main file restate it and extend it to Dirichlet L-functions, which is the generalised hypothesis [2].

The archive [2] and Mathlib [11]Lines of Lean a person has to read, against the lines only the checker reads

Lines

  1. Mathlib's definition of the hypothesis3
  2. The main file, five theorems12
  3. The file that states what was proved16
  4. Everything else in the archive17,374,187

Bars start at zero. Values are labelled in each panel.

So what a person has to read to know what's being claimed is a few lines, plus the definitions in Mathlib those lines are built on. Nobody has to read the 17 million lines of proof for the claim to hold, as long as they pass a checker that people trust. In the repository for the Navier-Stokes paper that Alpöge and Tristan Buckmaster posted the same day, the statement files have their proofs left blank, a solution file fills them in, and a small config names which theorems have to match and allows only Lean's three standard axioms [10].

Why this problem fell

On 7 November I wrote about METR's report on how fast Anthropic is improving its own models, and I said I thought the acceleration was real and narrower than it sounds. The tasks that had sped up were the ones a computer can mark, and the parts that are hard to check, like whether an idea was the right one to chase, looked as slow as before.

I still think the first half of that is right. A proof in Lean is about the hardest work I can think of that a computer can mark. The checker says yes or no and doesn't care who wrote it or how long it is, and that's how a 167-year-old problem could go from a first step to a finished proof in 41 days by the note's own timeline [1] and be believed by the evening.

The second half I'm less sure of than I was on 7 November. Alpöge writes that he chose the direction, two ideas from a simpler setting that he wanted to combine, and steered the work to the place where they'd apply, and that near the end Claude went far past him (I know the feeling). The step he singles out for other mathematicians, how long a chain of one construction had to be, he credits to Claude [1]. OpenAI said something close about its own research on 18 November, that for its newest model one researcher credited the agent with the key insights, which I wrote about the day after. So the choosing of ideas is moving too, and I think it moves fastest where there's a checker to tell a good idea from a bad one quickly. On 10 August Anthropic wrote that it didn't expect the techniques behind its 67.2% result to lead to a proof of the hypothesis [6], and Alpöge's note went up 102 days later.

The note's timeline [1], Anthropic [6] and Quanta [4] [7]2026, from two thirds of the zeros to all of them
  1. 10 AugustAnthropic

    An unreleased Claude raises the share of zeros proved to be on the line from 41.6% to 67.2%, and Alpöge is one of two staff mathematicians who check it. Anthropic writes that it doesn't expect those techniques to lead to a proof of the hypothesis [6]

  2. 30 SeptemberOpenAI

    OpenAI's first proof of the quasi-Riemann hypothesis, every zero between 1/8 and 7/8, checked in Lean [7]

  3. 6 OctoberOpenAI

    OpenAI announces 377 results in one day, the quasi-Riemann proof among them [7]

  4. 6 OctoberAnthropic

    The earliest date in Alpöge's timeline, on four of its twenty steps [1]

  5. 15 NovemberAnthropic

    Eight of the twenty steps carry this date, a Sunday [1]

  6. 16 NovemberAnthropic

    The last step reads GRH [1]

  7. 20 NovemberAnthropic

    The note and the Lean go up, and Sutherland compiles it before noon [1] [4]

  8. 20 NovemberBy hand

    Ravi Vakil posts Alessandro Masullo's 940-page proof of the same hypothesis, written without AI [4] [5]

  • Anthropic
  • OpenAI
  • By hand

The rest of Friday

Tao's post lists two more results with Alpöge's name on them, both with Buckmaster of New York University and both with Lean attached. One is a finite-time blow-up for the Euler equations and the other for the Navier-Stokes equations, the equations behind another of the Clay problems, and Tao says the Navier-Stokes paper needs more work before it's a properly written proof [3].

The same day Ravi Vakil, the president of the American Mathematical Society, posted a 940-page manuscript by Alessandro Masullo that claims the generalised hypothesis by hand [4]. Vakil's post says "It is all human, with no AI", that Brian Conrad has been going through it for months, and that the plan had been a year of polishing before it was announced [5]. Quanta adds that an earlier Riemann proof of Masullo's didn't hold up, so number theorists are waiting, and that the Lean community is already talking about formalising this one [4].

Does it break anything?

I spent twelve years building a crypto data company, so the question I'd expect first is whether this breaks anything, and it doesn't (your coins are as safe as they were on Thursday, however safe that was). Knowing the hypothesis is true doesn't hand anyone a faster way to factor a number or to take a discrete logarithm on an elliptic curve, which is what RSA and the elliptic curve signatures on Bitcoin rest on [4].

What the generalised hypothesis gave computing was mostly a guarantee about testing whether a number is prime. Miller showed in 1976 that the test now called Miller-Rabin runs in polynomial time if the hypothesis holds, and in 2002 Agrawal, Kayal and Saxena found a test that's polynomial with no hypothesis at all [9], so that guarantee hasn't depended on Riemann since.

What it does to mathematicians

Alpöge calls it the loss of a guiding light as well as a dawn [1]. Maksym Radziwiłł of New York University, one of the people the note thanks, told Quanta that mathematics is there to answer why something is true and that a proof in Lean "doesn't answer it at all" [4], and Alpöge expects the argument to get much shorter once people start on it [1]. The day before, Quanta reported that Dartmouth won't admit graduate students in mathematics for the next academic year (Quanta gives no reason, and reports it on its page about what AI is doing to mathematics) [4].

What I'd copy

We grade the tools agents use from public evidence and put the source next to every score, and our lab asks open-weight models questions about this site that have known answers and publishes how they did. Both are small next to a Clay problem, and both work the way Friday did, with a short thing a person can read and something independent that checks the long thing against it.

If you're building with agents, write the statement yourself, whether it's a test, a schema or 16 lines of Lean, and make everything the machine writes pass it.

What's still open

The comparator runs hadn't finished on Friday evening [4]. Nobody outside Anthropic can rerun the work, because Tao describes the model as an internal one and the note doesn't say which model it was, how many tokens it used or what it cost [1] [3]. When I wrote about the Fermat run I said I'd want work like this done on a model that has finished its lab's own checks, and four pages don't tell me whether this one had.

Anthropic hadn't announced anything itself when Quanta published [4]. Alpöge writes that the timing is horrible, that the result came together for him late on Sunday 15 November, and that there was a constraint he had to respect before he could announce it, which Quanta reads as the IPO (167 years of waiting, and then five more days for a share sale) [1] [4]. I complained about the timing of the METR report too, and the difference is that most of METR's evidence was redacted and this evidence is a download anyone can compile.

The Clay prize needs the proof published in a qualifying outlet, two years to pass after that and general acceptance among mathematicians [8], so it's at least two years away (and I have no idea who they'd make the cheque out to).

I'm a paying customer of Anthropic, the research agents and the review panel behind this site run on Claude, and this post was edited with Claude's help, which is what the star next to my name is for, so weigh what I say about a Claude result with that in mind.

I think it's been proved, it's another win for AI verification and I think that is splendid, and I hope to see a lot more of this.

References

Sources read on 20 November 2026. The archive was downloaded and counted the same evening.

  1. [1]Levent Alpöge, untitled four-page note, 20 November 2026. Source for the claim in his words, the extension to the Generalized Riemann Hypothesis, the formalisation, his not having mastered the mathematics around the argument, the 250,000 pages and why he posted only Lean, who chose the direction and where Claude went past him, the step he credits to Claude, the timeline of twenty dated steps from 6 October to 16 November 2026, the loss and the dawn, the timing and the constraint on announcing, the people he thanks and his expectation that the argument will be simplified. alpo.ge
  2. [2]Anchor Terminal's count of the Lean archive the note links, lean.tar.gz in 14 parts, downloaded on 20 November 2026 at 22:33 UTC. 239,275,554 bytes, SHA-256 9631b4f020199e580ba86178cb56c6e9b09f690c2369a80dc4334aa93f49c7fc, 175,187 Lean files, 17,374,215 lines and 1,030,531,854 bytes unpacked. Source for the sizes, the 16 lines of Statements.lean, the 12 lines and five theorems of Main.lean, the 725,147 lines that open a theorem, and the text search that found no sorry and no axiom declaration. The archive wasn't compiled by us. alpo.ge
  3. [3]Terence Tao, Unforced Euler singularity, unforced Navier-Stokes singularity, and the Generalized Riemann Hypothesis, 20 November 2026. Source for the three results in order of how digested they are, the internal model, the Langlands program and arithmetic geometry, the time the digesting will take, and the state of the Euler and Navier-Stokes papers. terrytao.wordpress.com
  4. [4]Quanta Magazine, Transformation, Konstantin Kakaes, the entries of 19 and 20 November 2026. Quanta reports the compiles by Andrew Sutherland and Simon DeDeo, the variant of Lean, the comparator checks still running and how long they might take, DeDeo's and Maksym Radziwiłł's comments, the 940-page Masullo manuscript and his earlier proof, the Lean community's discussion of formalising it, Anthropic having made no official announcement, its reading of the timing as the IPO, elliptic curves in Bitcoin and web browsing, and Dartmouth not admitting graduate students in mathematics for the next academic year. quantamagazine.org
  5. [5]Ravi Vakil, Hurried Announcement: Alessandro Masullo GRH, 20 November 2026. Source for the proof being written without AI, Brian Conrad's months of checking and the plan to polish it for a year. math216.wordpress.com
  6. [6]Anthropic, Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis, 10 August 2026. Source for 1859 and the million-dollar prize, the bound going from 41.6% to 67.2%, Alpöge and Ralph Furman checking it, and Anthropic not expecting those techniques to prove the hypothesis. anthropic.com
  7. [7]Quanta Magazine, AI Proved the Quasi-Riemann Hypothesis. What Happens Now?, Konstantin Kakaes, 19 October 2026. Source for what the hypothesis says and why the zeros matter to the primes, the first 12 trillion zeros checked by computer, OpenAI's 377 results of 6 October 2026, and its proof dated 30 September that every zero lies between 1/8 and 7/8, verified in Lean. quantamagazine.org
  8. [8]Clay Mathematics Institute, Rules for the Millennium Prize Problems, adopted 26 September 2018. Source for the three conditions before a proposed solution is considered. claymath.org
  9. [9]Wikipedia, Generalized Riemann hypothesis, as read on 20 November 2026. Source for Miller's 1976 result on the Miller-Rabin test under the hypothesis and the unconditional test of Agrawal, Kayal and Saxena in 2002. en.wikipedia.org
  10. [10]Levent Alpöge and Tristan Buckmaster, Finite-time blow-up for the Navier-Stokes equations, repository README, 20 November 2026. Source for the statement files, the solution file and the comparator configs that allow only the axioms propext, Quot.sound and Classical.choice. github.com
  11. [11]Mathlib, Mathlib/NumberTheory/LSeries/RiemannZeta.lean at revision 0df444a, the one the archive's lakefile pins. Source for the definition of RiemannHypothesis. raw.githubusercontent.com

Your experience matters

Used a tool in the Terminal?

Tell people and their agents what worked and what didn't. Write a review of a tool you've used and confirm it by email or wallet.