Explainer20 min read22 sources
Has the Riemann hypothesis been solved? Alpöge's Claude and Lean claim, and what is checked
On 20 November 2026 Levent Alpöge said that, working with Claude, he found a proof of the Riemann hypothesis, released as Lean code. What is checked and what is open.
Not by any accepted standard yet. On 20 November 2026 Levent Alpöge, a mathematician who works at Anthropic, posted a four-page note on his own website. In it he writes: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis."1 He says the method extends to the generalised Riemann hypothesis, and he released the argument only as Lean code, a very large archive of computer-checked mathematics. Several people have since compiled that code and re-checked it with independent programs. When we last looked, at 11:15 UTC on 21 November, none of the finished checks on Andrew Sutherland's public page at MIT had rejected the code, three runs were still in progress, and the Clay Mathematics Institute still listed the Riemann hypothesis as "Unsolved".215
What is not established matters as much. There is no paper that a person can read from start to end: Alpöge writes that he "decided to only present Lean", and that he intends to write the argument properly. A passing Lean check shows that the code proves the statement as written in Lean, here Mathlib's own definition of the Riemann hypothesis, provided Lean's kernel, three standard axioms and Mathlib's definitions are sound. It does not explain the argument, and it is not peer review. Specialists have only begun to work out the strategy, and Clay's prize rules ask for publication in a qualifying outlet, normally a refereed journal, then a two-year wait and general acceptance.
Below: what was announced, what has been checked and by whom, what stays open, what Clay requires, a second claim of the same day, and what a proof would change. Earlier posts cover the quasi-Riemann hypothesis, all seven Millennium Prize Problems and how to check a Lean proof that comes with a paper. Every source was checked on 21 November 2026, between 11:00 and 11:20 UTC. Some checks were still running, so the state described here may change after publication.
What did Alpöge announce on 20 November 2026?#
The note, at alpo.ge/remark.pdf1, is short and personal. Its central claims: "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." A footnote links fourteen files, lean.tar.gz.part01 to part14, which join into one compressed archive of a Lean 4 project.
On the division of work, he writes that "this time Claude far outpaced me near the finish line", and: "I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe." The note names the model only as Claude, with no version. Terence Tao describes the release as follows: "Levent Alpöge, using an internal AI model, has generated an (enormous) Lean-only proof of the Generalized Riemann Hypothesis."6
"The argument is incredibly long at the moment", the note says, and "I may have baked lots of strange suboptimal choices into the argument". In a footnote Alpöge says that a generated write-up ran to more than 250,000 pages, and that rather than post it, "I decided to only present Lean." His stated plan: "I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly." The last pages give "a rough timeline Claude provided after a thorough search", dated 6 October to 16 November 2026. The Proofs and Prompts newsfeed22 summarised the note the same day.
This is Alpöge's personal announcement on his own site. At 11:03 UTC on 21 November, Anthropic's newsroom21 and research pages carried no item about it, and we found no company statement. In August 2026 Anthropic did publish a separate, smaller result, examined by Alpöge among others, that raised a lower bound for the share of zeta zeros on the critical line "from 41.6% to 67.2%". That post said: "We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis."20
Tao's post of 20 November also lists two claims about singularities in the Euler and Navier–Stokes equations, by Alpöge with Tristan Buckmaster and, for Euler, Matei Coiculescu.6 They are a different subject and are not covered here.
What is the Riemann hypothesis, and how does the generalised one differ?#
The Riemann zeta function is ζ(s) = 1 + 1/2^s + 1/3^s + …, extended to all complex numbers s except s = 1. It is zero at −2, −4, −6 and so on, the trivial zeros. Riemann's hypothesis of 1859 says that every other zero has real part exactly 1/2. In Clay's words, it "asserts that all interesting solutions of the equation ζ(s) = 0 lie on a certain vertical straight line."15
The link to primes: Enrico Bombieri's official problem description for Clay explains that the hypothesis is equivalent to a sharp bound on how far the count of primes up to x can stray from its expected value, roughly √x times a logarithm. He adds: "The failure of the Riemann hypothesis would create havoc in the distribution of prime numbers."18
The numerical evidence was strong but finite. Clay's page says the hypothesis "has been checked for the first 10,000,000,000,000 solutions." Dave Platt and Tim Trudgian showed by a rigorous computation, published in 2021, that every zero with imaginary part between 0 and 3·10^12 lies on the line: "We verify numerically, in a rigorous way using interval arithmetic, that the Riemann hypothesis is true up to height" 3·10^12.19
The generalised Riemann hypothesis (GRH) asks the same of related functions. Dirichlet L-functions are the versions of ζ that count primes in arithmetic progressions, such as primes ending in 7. Wider families include the Dedekind zeta functions of number fields and Hecke L-functions. Which family a claim covers matters, because many applications need more than the Dirichlet case.
| Statement | What it asserts | Alpöge's Lean code | Masullo's draft |
|---|---|---|---|
| Riemann hypothesis | Every non-trivial zero of ζ(s) has real part 1/2 | Stated and claimed, using Mathlib's own definition | Not listed on its own; by our reading it falls within the scope below |
| GRH for Dirichlet L-functions | The same for every Dirichlet L-function | Stated and claimed, for every modulus and character | Not listed on its own; by our reading it falls within the scope below |
| GRH for wider families | The same for Dedekind zeta functions, Hecke and other L-functions | Not in the code; Alpöge expects the "Extended Riemann Hypothesis" with more time | Claimed, in the draft's words, for "all number fields, all unitary Hecke characters" |
| Form of the argument | Lean code only, no readable paper | A 940-page human-written draft, no Lean |
The Lean project states five theorems. One is theorem riemannHypothesis : RiemannHypothesis, where RiemannHypothesis is the definition in Mathlib, Lean's community library, at the official release the project pins.5 The other four restate the claim for ζ and cover Dirichlet L-functions, for all characters and for primitive ones. Sutherland's page notes that GRH there means Dirichlet L-functions only, and the notes of the community checking repository described below say: "Dedekind zeta functions and other L-functions are not covered."3
Key terms#
- Lean and its kernel: Lean is a language in which proofs are written as code; its kernel is the small program that checks each step.
- Mathlib and its axioms: Mathlib is the large library of mathematics written in Lean by its users. It relies on three standard axioms,
propext,Classical.choiceandQuot.sound. - Export: a file listing every definition and proof behind a theorem, so that programs other than Lean can re-check them.
- Challenge and Comparator: a challenge file states the theorems on their own, importing only Mathlib. Comparator checks that a submitted proof proves exactly those statements, with permitted axioms only, and replays it through one or more kernels.
- Independent kernels: checkers written separately from Lean's own, such as nanoda, con-leche and lean4lean. A bug would have to fool several of them at once.
Is the Riemann hypothesis proof verified? What has been checked so far#
The main public record is Andrew Sutherland's page, RH Lean proof: verification files2. It describes itself as "Andrew Sutherland (MIT), checks run with Claude Code." We read it at 11:00 and again at 11:15 UTC on 21 November; it was last updated at 09:30 UTC and says it "will be updated as the remaining checks finish."
| Check | Who ran it | What it covers | State, as we last read it |
|---|---|---|---|
| Compile all 175,187 files with Lean 4.33.1 | Sutherland | Lean accepts each file; the kernel of that version checks each declaration as it is added | Passed, zero errors, finished 16:43 UTC on 20 November |
#print axioms on the five theorems | Sutherland | Which axioms the theorems rest on | Passed: only the three standard axioms |
| Mathlib files against the official cache | Sutherland | That the Mathlib underneath is the official release | Passed: 8,322 Mathlib files identical byte for byte |
leanchecker replay | Sutherland | Lean's own kernel re-checks every module from the compiled files | Passed: 175,187 of 175,187 |
| nanoda and con-leche on the export | Sutherland | Two independently written kernels re-check the full exported proof | Both accept |
con-leche --compare | Sutherland | The five theorems match the separately stated challenge | Accepts all five |
| lean4lean, module by module | Sutherland | A further kernel design re-checks the compiled files | Passed: 175,187 of 175,187, five of them after a step limit was raised |
| Lean's default and paranoid kernels on the export | Sutherland | Lean's kernel, plain and hardened, on the export | Running: "single-threaded; no verdict yet" |
lake comparator with Lean 4.36.0 | Sutherland | Statement match, axioms and kernel replay in one run | Running: statement and axiom stages passed, kernels in progress |
A plain lake build on another machine | Hertaspace repository | An independent rebuild by a different method | In progress: job 144,445 of 183,903 at 11:06 UTC |
| Comparator with a sandboxed build | leanprover-community repository | The same judging, with the build isolated | Build in progress, no results posted |
Sutherland calls con-leche's compare mode "an independent implementation of Comparator's statement matching". Two checkers bundled with Lean 4.36.0 could not process one kind of definition in the proof and stopped, which his page records as "exit 2, not a rejection"; the newer con-leche handles it and accepts.2
He also ran a test that does not depend on how Mathlib defines ζ. He combined the archive with an independent formalization of Robin's criterion, a known equivalent of the Riemann hypothesis that speaks only of sums of divisors. The result is a Lean theorem saying that for every n above 5040, the sum of the divisors of n is less than e^γ · n · log log n, where γ is Euler's constant, with "no hypotheses and no mention of zeta". It still rests on Mathlib's definitions of the divisor sum, exp and log, and on that formalization, which went through the same checks. Two kernels accept the combined proof; its lake comparator run was still going.
The other two efforts. The Hertaspace repository4 rebuilds the project with a plain lake build on 256 cores. Its status file, refreshed at 11:06 UTC, showed two failed jobs, both errors reading a compiled file, and no axiom result yet. Its record explains the earlier one: "This is an error reading a compiled Mathlib file, not an error in the project", and a second pass is planned; it did not yet comment on the later one. Its checks "were set up and run by a large language model agent (Claude) under human direction." The leanprover-community repository mirrors the archive unmodified and adds a challenge file, a Comparator setup and a sandboxed build.3 Its public CI reproduces the challenge export with the same hash that Sutherland's page reports. At 11:15 UTC its notes still said: "The build is in progress."
What named mathematicians have said. Sutherland told Scientific American: "The Lean proof has been compiled and independently rechecked by the Lean kernel with only the three standard axioms," and "Honestly the chance that anything is fishy here is approximately zero."9 That article's headline says the proof "is almost certainly true, independent mathematicians attest", but the only checker it names and quotes is Sutherland. It also quotes Kevin Buzzard, writing on the Lean community's chat server: "We still await a successful comparator run, which will check that (a) the code does not do anything nefarious like redefine the meaning of any of the symbols in the statement and (b) (assuming it’s run in paranoid mode) the code compiles with several independent kernels". That thread, which Tao calls "probably the best place to follow the efforts to verify the proof and extract the strategy of proof", requires signing in, so we report only what readable sources quote from it.6
What does “checked in Lean” leave open?#
Buzzard's post of 12 November, "…and the proof is checked in Lean"7, sets out four questions to ask of a Lean claim: does the Lean statement match the theorem, does the code compile, does it prove that statement without hidden changes, and does it exploit a bug in Lean's kernel. For this archive the answers so far look like this.
The statement. This part is unusually simple here. The headline theorem uses Mathlib's definition, written before the claim, whose comment says that "constructing a term of this type is worth a million dollars." People checking the code read it by hand: it excludes the trivial zeros and the point s = 1, where Mathlib's ζ has an arbitrary value.3 Sutherland states the limit: "A passing check shows that the formal statements, as written in Lean against Mathlib's definitions, follow from the three standard axioms according to the kernels used. It relies on those kernels and on Mathlib's definitions being what they are intended to be."2
The kernels. The project is pinned to Lean 4.33.1, which predates some later runtime fixes. The community notes conclude that "a pass by the 4.33.1 kernel alone, which is what a successful build is, counts as limited evidence here." They also note tactics in the proof for which "the elaborator has not checked the term against the goal, so the kernel is the only check". This is why the pending kernel and Comparator results matter.
The build conditions. Sutherland's page says that "The build ran on cloud machines without a sandbox", and the export the independent kernels read came from that build. The community notes add that Comparator "cannot detect a modified Mathlib", which is why both efforts compare Mathlib's files with the official copy.
Understanding. No check tells anyone why the argument works. Tao writes that, from very initial analysis, the strategy "appears to use the full power of modern developments on the Langlands program and arithmetic geometry", and: "I imagine the digestion process will take some time, though in this case at least we have a human author to communicate with to assist with this task."6 TechCrunch quotes the number theorist Chris Birkbeck in the chat: "The cohomology bound is proved by what looks like an automorphy-lifting argument."12 Readers have begun posting unofficial maps of the proof as issues in the community repository; one, written with Claude's help, says of itself: "It is not a verification."3 Buzzard's general warning applies: even code that passes every test "does not mean that any corresponding pdf also contains a correct proof!" Here there is no such paper yet.
How long is the proof, and can you check it yourself?#
By our count of the downloaded archive, the fourteen parts join into 239,275,554 bytes, which unpack into 175,187 Lean files and 17,374,394 lines. The lean-toolchain file names Lean 4.33.1, and Mathlib is pinned at commit 0df444a3, its official v4.33.1 release. Sutherland's page and the community notes give the same bytes, files, versions and archive hash, which begins 9631b4f0. The notes count 179 fewer lines, 17,374,215, and 818,755 theorem or lemma declarations, and say the source has no comments.3 For comparison, Anthropic's machine-written formalization of Fermat's Last Theorem was about 13 million lines.
The exported proof is 61,570,595,788 bytes unpacked, and con-leche accepted 1,468,418 declarations in it.2 Alpöge's footnote puts a generated prose version at more than 250,000 pages.
Checking it yourself takes serious hardware. Sutherland offers his export for download: "With Lean v4.36.0 installed, nothing needs to be built. You need roughly 100 GB of RAM and 70 GB of disk." The quickest independent check, con-leche, needs about 50 GB of memory and took under four minutes on 64 threads. Rebuilding from source is far heavier: the Hertaspace build had run for almost 19 hours on 256 cores and 2 TB of memory when we looked, and the community notes expected 7 to 10 hours on 48 cores. Building a project runs its code, so Buzzard's rule holds: "do not run random Lean code on your local machine unless you sandbox it first."7
What does the Clay prize require?#
The Riemann hypothesis is one of Clay's seven Millennium Prize Problems, each carrying $1 million. Clay's rules page16 says that "CMI does not accept direct submission of proposed solutions", and that three conditions must all hold before it will consider one: "(i) the proposed solution must be published in a Qualifying Outlet (see §6), and (ii) at least two years must have passed since publication, and (iii) the proposed solution must have received general acceptance in the global mathematics community".
The full rules of 26 September 2018 define a Qualifying Outlet chiefly as "a refereed mathematics publication of worldwide repute", and provide for a special advisory committee after the two years. They add that Clay "will pay special attention to the question of whether a Prize solution depends crucially on insights published prior to the solution under consideration."17
Neither claim of 20 November has been published in a refereed journal, so far as our sources show. At 11:15 UTC on 21 November, Clay's Riemann hypothesis page15 still carried the label "Unsolved".
What about the second claim made the same day?#
On 20 November Ravi Vakil, a Stanford mathematician whom Scientific American identifies as president of the American Mathematical Society, posted "Hurried Announcement: Alessandro Masullo GRH"13 on his course blog. He wrote that "Alessandro Maria Masullo has a proposed proof of GRH", that Brian Conrad of Stanford had been looking at it closely for months, and: "It is all human, with no AI." The plan had been a year of polishing. 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 draft14, "The Music of the Primes", is a 940-page PDF. Its acknowledgements say the work began when Masullo was a graduate student at Stanford under Conrad and that "It took nearly 12 years to assemble." They give the scope as "generality: all number fields, all unitary Hecke characters" and are frank about its state: "The math is complete, and only filling in some writing remains on several parts across this long paper." Also: "The introduction itself is unfinished", and "Brian Conrad and I are aware of several glitches here and there".
About eight hours after posting, Vakil added an update: "Several people have contacted me to report problems in the proposed proof, identified with the assistance of different AI models." He also wrote: "As is usual, all issues need to be resolved to the satisfaction of the community before this proof would be accepted as true. (I make no claim as to the correctness of the proof strategy.)" Conrad told Scientific American: "To judge the success of the method requires the paper to reach a stage of more final revision," and "It’s undergoing major revision."10
The two claims differ in form as much as in scope: Alpöge's is machine-checkable code with no readable paper, Masullo's a readable draft with no machine check, aimed at a wider family of L-functions.
What would a proof of the Riemann hypothesis change?#
Little in daily life, and a great deal inside mathematics. Clay's page says a proof "would shed light on many of the mysteries surrounding the distribution of prime numbers."15
Results that already assume it. Many theorems in number theory are proved on the assumption that RH or GRH holds. Bombieri lists consequences of GRH-type statements, including a strong effective form of Chebotarev's density theorem, a result of Hooley's on cubic forms, and "Miller's deterministic polynomial time primality test".18 Each needs the hypothesis for a particular family of L-functions, so what a proof unlocks depends on which family it covers. Bombieri also notes that "many deep results in number theory which are consequences of a general Riemann hypothesis can be shown to hold independently of it".
Statements about whole numbers. Sutherland's Robin test shows one kind of consequence: a plain inequality about sums of divisors, for every n above 5040.2
Cryptography. Scientific American's Lee Billings addresses the worry about encryption that relies on factoring. Testing whether a number is prime is different from factoring it, and "GRH would improve mathematicians’ understanding of the former but wouldn’t provide a known shortcut for the latter."11
New methods. If the argument holds, its ideas may matter more than the statement. Hector Pasten of the Pontifical Catholic University of Chile told Scientific American: "If the proof is correct and understood, this completes a dictionary between different branches of math where the connections were only partial".8 That depends on someone turning 17 million lines into an argument people can follow.
What this means for a research record#
Much of a claim's history can be pinned down while its correctness is open. From public sources a reader can already note, with dates: who claims what, and where (Alpöge, in a note on his own site, 20 November); which model, as he reports it (Claude, no version given); which exact files (the note's SHA-256 begins 3fc97784, the hash the community notes recorded on 20 November and our own download gave on 21 November; the joined archive's begins 9631b4f0); who ran which check, with which tool versions, and when each finished; and that no readable proof exists yet. Sutherland's page and both repositories already keep most of this.
Papex lists Alpöge's claim on its Riemann hypothesis page, quoted as its source states it, alongside earlier partial results. At 11:15 UTC on 21 November the problem's status there stayed Open, following Clay's page. A listing points to a claim; it does not check it.
A note on interests: this post reports on a claim made with Anthropic's Claude models by an Anthropic employee, and it was drafted with Claude. Two of the checks it describes were also run with Claude, by their authors' account. Readers should weigh it with that in mind and check the sources below.
Sources#
- Levent Alpöge, remark.pdf, four-page note announcing the claim, with links to the Lean archive (alpo.ge, 20 November 2026; SHA-256 3fc97784…), checked 21 November 2026 a b
- Andrew Sutherland, RH Lean proof: verification files (MIT; page last updated 21 November 2026, 09:30 UTC), checked 21 November 2026 at 11:00 and 11:15 UTC a b c d e f
- leanprover-community/riemann-hypothesis: verification notes, README, challenge file and issues (GitHub), checked 21 November 2026 at 11:02 and 11:15 UTC a b c d e
- Hertaspace/alpoge-rh-lean-verification: README, VERIFICATION.md and build status (GitHub; status refreshed 21 November 2026, 11:06 UTC), checked 21 November 2026 at 11:15 UTC ↩
- Mathlib, Mathlib/NumberTheory/LSeries/RiemannZeta.lean at commit 0df444a3, the definition of RiemannHypothesis (GitHub), checked 21 November 2026 ↩
- Terence Tao, Unforced Euler singularity, unforced Navier-Stokes singularity, and the Generalized Riemann Hypothesis (What's new, 20 November 2026), checked 21 November 2026 a b c d
- Kevin Buzzard, “…and the proof is checked in Lean” (Xena, 12 November 2026), checked 21 November 2026 a b
- Joseph Howlett, Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math (Scientific American, 20 November 2026), checked 21 November 2026 ↩
- Clara Moskowitz, The AI Riemann hypothesis proof is almost certainly true, independent mathematicians attest (Scientific American, 20 November 2026), checked 21 November 2026 ↩
- Joseph Howlett, Dueling claims of Riemann hypothesis proofs emerge (Scientific American, 20 November 2026), checked 21 November 2026 ↩
- Lee Billings, What does a solution to the Riemann hypothesis mean for you? (Scientific American, 20 November 2026), checked 21 November 2026 ↩
- Tim Fernholz, AI claimed a solution to the 167-year-old Riemann hypothesis, and a human did, too (TechCrunch, 20 November 2026), checked 21 November 2026 ↩
- Ravi Vakil, Hurried Announcement: Alessandro Masullo GRH, with an update about eight hours later (Math 216 blog, 20 November 2026), checked 21 November 2026 ↩
- Alessandro Maria Masullo, The Music of the Primes, draft, 940 pages (posted on the Math 216 blog, 20 November 2026), checked 21 November 2026 ↩
- Riemann Hypothesis, status label “Unsolved” (Clay Mathematics Institute), checked 21 November 2026 at 11:00 and 11:15 UTC a b c d
- Rules for the Millennium Prize Problems (Clay Mathematics Institute), checked 21 November 2026 ↩
- Millennium Prize Description and Rules (Clay Mathematics Institute, 26 September 2018), checked 21 November 2026 ↩
- Enrico Bombieri, Problems of the Millennium: the Riemann Hypothesis, official problem description (Clay Mathematics Institute), checked 21 November 2026 a b
- Dave Platt and Tim Trudgian, The Riemann hypothesis is true up to 3·10^12, arXiv:2004.09765; Bulletin of the London Mathematical Society 53 (2021) 792–797, checked 21 November 2026 ↩
- Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis (Anthropic, 10 August 2026, updated 13 August 2026), checked 21 November 2026 ↩
- Anthropic News (Anthropic), no item about the claim, checked 21 November 2026 at 11:03 UTC ↩
- Newsfeed, item of 20 November 2026 on Alpöge's claim (Proofs and Prompts), checked 21 November 2026 ↩