Anthropic mathematician Levent Alpöge posts a Lean proof of the Riemann Hypothesis found with Claude; the Lean and Mathlib communities certify it
Result confirmed
Importance: historic (5 of 5)The takeaway
No human has yet explained the argument (an informal write-up would exceed 250,000 pages).
Status
- Claim
Result confirmed
- Our reporting
- High confidence
- Verification
- Lean-verified
- Importance
- Historic (5 of 5)
- Last verified
- 22 November 2026
Your AI and this story
- GPT-6.2 Astra173 days after its cutoff
- Claude Opus 5.5143 days after its cutoff
- Gemini 4 Hydrogen234 days after its cutoff
- Grok 4.7173 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 143 days before it.
Key facts
- Statement (20 Nov 2026, Alpöge’s website, quoted by Scientific American and others): ‘I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis’
- Scope claimed: the Generalized Riemann Hypothesis as well as the classical one
- Released as a Lean formalization; per press reports, the proof took shape between early October and mid-November, with the final step dated 16 November, and Alpöge credits Claude with rapid progress near the end
- Status on 20 Nov: unverified. Reports say an independent project had only checked that its copy of the files matched the release and that the target statements can be expressed in Mathlib; whether the formal statement matches the Riemann Hypothesis has not been confirmed
- Our check of the released Lean (20 Nov, archive from alpo.ge, 14 parts, 239 MB; 175,187 .lean files, 1.3 GB; Lean v4.33.1, Mathlib pinned): Main.lean proves
theorem riemannHypothesis : RiemannHypothesis, i.e. Mathlib’s own definition (zeros of riemannZeta other than the trivial zeros and s = 1 have real part 1/2), plus an explicit restatement and the Generalized RH for Dirichlet L-functions (primitive and all characters); no file redefines RiemannHypothesis or riemannZeta; nosorry, noaxiom, nonative_decide,implemented_by,externorunsafeanywhere. We did not compile it, so the proof itself is not yet machine-checked by us - Community replay (Rohan Arun, from 20 Nov 16:01 UTC): github.com/rohanarun/alpoge-rh-replay splits the 175,187 modules into 2,858 packets claimed by volunteers and their agents via a GitHub bot; by 20:27 UTC Arun reported 35 contributors had submitted Lean replays for 32,872 modules (18.76%); its STATUS page says the full source build, axiom audit and fresh-kernel check of Main are still pending
- Hector Pasten (PUC Chile) to Scientific American: ‘If the proof is correct and understood, this completes a dictionary between different branches of math where the connections were only partial … Something like a Rosetta Stone.’
- Context: OpenAI’s October maths release claimed the weaker quasi-Riemann hypothesis; Alpöge co-led the Claude-assisted Jacobian conjecture counterexample (July 2026)
Show 18 more
- Primary source: Alpöge’s four-page note (alpo.ge/remark.pdf): Lean code in 14 archive parts on his site; ‘Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled’; he steered towards combining two ideas from a toy setting, and Claude suggested a chain of theta lifts of length √X; an informal write-up ‘hit 250,000+ pages’, so he posts only the Lean
- Timeline in the note (compiled by Claude): first steps 6 Oct 2026, the decisive iterated-theta-lift steps 15 Nov, GRH 16 Nov; he also expects the Extended Riemann Hypothesis ‘with more time’
- The note does not say which Claude model was used; he cleared the announcement with Dario Amodei and credits OpenAI’s maths releases with ‘good taste’
- Independent build (Scientific American, 20 Nov, second story): MIT’s Andrew Sutherland compiled the Lean code himself and reports the kernel rechecked it using ‘only the three standard axioms’: ‘Honestly the chance that anything is fishy here is approximately zero.’ Kevin Buzzard (Imperial): ‘We still await a successful comparator run’ (the check that the statements match and independent kernels accept the proof)
- Comparator setup: leanprover-community/riemann-hypothesis (created 20 Nov 14:47 UTC) mirrors the release byte-for-byte and states the five target theorems in a file importing only Mathlib; its notes count 17,374,215 lines, 818,755 theorems/lemmas, no comments, and flag custom tactics (kexactB in 67 files, kcopy) that skip the elaborator’s type check, so ‘the kernel is the only check’; no Comparator result posted as of 21 Nov ~12:00 UTC
- Community replay, 21 Nov 11:18 UTC: 80,914 of 175,187 modules (46%) carry build or reuse evidence in rohanarun/alpoge-rh-replay; dozens of credited contributors; full build, axiom audit and fresh-kernel check of Main still pending
- Rival human claim, same day: Stanford’s Ravi Vakil (AMS president) posted Alessandro Masullo’s draft proof of GRH (‘The Music of the Primes’), ‘all human, with no AI’, reviewed for months by Brian Conrad and ‘so far looking good to him’; about eight hours later Vakil added that several people had reported problems in it, ‘identified with the assistance of different AI models’
- Reactions on X: Alex Kontorovich, ‘RH is proved by using everything!’ (587k views) and ‘only 17M lines of Lean’; Thomas Bloom surprised by a coherent proof of 250,000+ pages (407k); Eliezer Yudkowsky called RH the ‘coal mine canary’ for dangerous AI (656k); Kevin Roose asked whether big maths prizes can continue (80k)
- Scott Aaronson (blog, 20 Nov, ‘The stochastic parrots have apparently proved the Riemann Hypothesis’): ‘I’m not 100% sure, but my bet is that the proof will stand’; ‘Mathematicians of course won’t be satisfied with 17 million lines of Lean’; Lean ‘still has issues, from kernel bugs (which have, in the past, allowed proofs of 0=1) to uncertain soundness of the underlying type theory’, which ‘can and should urgently be fixed’, though he sees ‘zero indications’ an AI is exploiting them here
- Certification (Lean community blog, 21 Nov 13:55 UTC, joint statement of the Lean and Mathlib communities endorsed by the Lean FRO and the Mathlib Initiative): ‘We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct’; ‘We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct’; all four criteria of the Xena post of 12 Nov are met
- Basis of the certification: dozens of people report compiling the code and reproducing the same raw proof export; a community file stating RH and importing only Mathlib was matched to Alpöge’s theorem by two independent tools (Comparator and con-leche --compare); the export was accepted by Lean’s kernel, con-leche and con-ron (which carry a formal proof, relative to set theory, that they cannot accept a false proof) and several third-party checkers in other languages, run on an inert export
- Andrew Sutherland’s verification page (MIT, checks run with Claude Code, as of 22 Nov 04:35 UTC): all 175,187 modules compile with zero errors; the five headline theorems use only propext, Classical.choice and Quot.sound; a sandboxed rebuild reproduced the 61.6 GB export byte for byte; leanchecker replayed every module; nanoda, con-leche and lean4lean accepted the proof; con-leche and con-ron --compare confirmed all five theorems; the end-to-end lake comparator run and the Lean-kernel replay of the export were still running
- Scope checked: RH with Mathlib’s definition and GRH for all Dirichlet L-functions (every modulus and character); Dedekind zeta, Hecke and general automorphic L-functions are not among the verified statements. Combined with an independent Lean formalization of Robin’s and Lagarias’s criteria, the proof yields Robin’s inequality for n > 5040 and Lagarias’s inequality for all n, with no mention of zeta
- Community replay (rohanarun/alpoge-rh-replay, ledger of 22 Nov 11:58 UTC): 108,702 of 175,187 modules (62%) carry build, reuse or kernel-check evidence; the replay’s own end-to-end check of Main is still pending
- Terence Tao (blog, 20 Nov): a Lean-only proof generated ‘using an internal AI model’ that ‘from very initial analysis, appears to use the full power of modern developments on the Langlands program and arithmetic geometry’; the Lean Zulip thread moderators limited discussion to the Lean-related aspects
- Opposition (Association for Human Mathematics statement, 20 Nov): the announcement represents ‘a new value system put forward by AI corporations that foregrounds technical validity of logical statements over accountability, clarity, accessibility, and understanding’; ‘no human being can reasonably claim to be responsible for, or even to understand, such a document’; it calls for democratic oversight of AI companies
- Most viewed posts on the rival human draft: science writer Kevin Hartnett, ‘This is massive… He thought he had more time but didn’t’ (1.34M views, 20 Nov); on the claim itself, maths student @sui414 wrote that ‘the world after today is no longer the same world i grew up in’ (632k) and Deedy Das called it possibly ‘the biggest AI math result of all time’ (486k)
- Same day, Alpöge and NYU’s Tristan Buckmaster also released an AI-aided proof of Navier-Stokes blow-up that they say closes a loophole in OpenAI’s earlier solution (Scientific American, TechCrunch)
What happened
On Friday 20 November 2026, Levent Alpöge, a number theorist who works at Anthropic, posted a short first-person statement saying he believes he has, with Claude, found a proof of the Riemann Hypothesis, and that it extends to the Generalized Riemann Hypothesis. He published the proof as Lean code. Same-day coverage (Scientific American, Interesting Engineering, Washington Times, Kit Yates) treats it as a serious but unconfirmed claim.
Why it matters
If correct, it settles the most famous open problem in mathematics, open since 1859 and one of the Clay Millennium Prize problems, and with GRH it would make true a large body of results that were proved only conditionally. A Lean proof shortens verification, but the first check is whether the formal statement says what the Riemann Hypothesis says. Until specialists report, the status is a claim.
Second day (21 November)
The first outside check came from MIT’s Andrew Sutherland, who built the code himself and told Scientific American the kernel accepted it with only Lean’s three standard axioms. Kevin Buzzard said the community was still waiting for a Comparator run, the check that the Lean statements say what the problem says and that independent kernels accept the proof. The Lean community’s mirror reads the five target statements by hand and notes that the code uses custom tactics that skip the elaborator’s type check, so the kernel is the only check there; its Comparator result was not yet posted by midday on 21 November. Rohan Arun’s volunteer replay had covered 46% of the modules. Nobody has yet explained the mathematics: the informal argument would run past 250,000 pages.
The same morning, Stanford’s Ravi Vakil posted a separate, “all human” draft proof of GRH by Alessandro Masullo, which Brian Conrad had been reviewing for months. Within hours, readers using AI models reported problems in it. Vakil said every issue has to be resolved before anyone accepts it.
Certification (21–22 November)
In the early afternoon of 21 November (UTC) the Lean and Mathlib communities published a joint statement, endorsed by the Lean FRO and the Mathlib Initiative, saying they have “full confidence” that the proof is correct and certifying both the statement and the proof. The statement rests on many independent builds that reproduced the same proof export, on two separate tools that matched Alpöge’s theorem to a community file stating the Riemann Hypothesis with only Mathlib, and on several proof checkers written independently of Lean, two of which come with a formal proof that they cannot accept a false proof. Andrew Sutherland’s public log of these checks lists what had passed by 22 November and what was still running. What the certificate settles is that the theorem is true in Lean’s logic; it says nothing yet about why. Tao’s first impression is that the proof draws on the Langlands programme and arithmetic geometry, and the Association for Human Mathematics objects that nobody can take responsibility for, or understand, a document of this size.
Sources
28 sources from 23 sites. Numbers match the chips in the text.
28 sources: 6 primary, 14 press, 8 reactions
Primary
- Levent Alpöge: note on the claimed proof (PDF, 4 pages)alpo.ge, official
- Rohan Arun: community Lean replay of the release (GitHub)github.com, code
- leanprover-community/riemann-hypothesis: Comparator setup and verification notesgithub.com, code
- Ravi Vakil: hurried announcement, Alessandro Masullo’s proposed proof of GRH (all human)math216.wordpress.com, official
- Lean community blog: Lean verification of the Riemann Hypothesis (joint statement, 21 Nov)leanprover-community.github.io, official
- Andrew Sutherland (MIT): RH Lean proof verification files and resultsmath.mit.edu, code
Press
- Scientific American: Anthropic mathematician claims to prove Riemann hypothesisscientificamerican.com, press
- Kit Yates: AI may have helped solve the most famous problem in mathskityates.substack.com, press
- Interesting Engineering: Claude helped prove 167-year-old Riemann hypothesis, researcher claimsinterestingengineering.com, press
- Washington Times: Riemann hypothesis may have been proved by AI mathematicianwashingtontimes.com, press
- Tom’s Hardware: claim remains unconfirmedtomshardware.com, press
- Scientific American: the AI Riemann hypothesis proof is almost certainly true, independent mathematicians attestscientificamerican.com, press
- TechCrunch: AI claimed a solution to the Riemann hypothesis, and a human did tootechcrunch.com, press
- Nature: mathematician says he and AI solved Riemann hypothesisnature.com, press
- New York Times: Anthropic mathematician claims to solve Riemann hypothesis with help from Claudenytimes.com, press
- The Economist: AI may have solved the most important problem in mathematicseconomist.com, press
- Washington Post: using AI, mathematician claims to have solved one of math’s great mysterieswashingtonpost.com, press
- BBC: Anthropic may have solved major maths problembbc.com, press
Show all 14 sources
Reactions
- Rohan Arun on X: 18.76% replayed in under 4 hoursx.com, discussion
- Scott Aaronson: The stochastic parrots have apparently proved the Riemann Hypothesisscottaaronson.blog, discussion
- Xena (Kevin Buzzard): ‘…and the proof is checked in Lean’ (the four criteria, 12 Nov)xenaproject.wordpress.com, discussion
- Terence Tao: unforced Euler singularity, unforced Navier-Stokes singularity, and GRHterrytao.wordpress.com, discussion
- Association for Human Mathematics: statement on the announced proof of the Riemann Hypothesisahmath.org, discussion
- Kevin Hartnett on X: Masullo’s separate human proof (1.34M views)x.com, discussion
- Alex Kontorovich on X: ‘RH is proved by using everything!’x.com, discussion
- Eliezer Yudkowsky on X: ‘The Riemann Hypothesis has fallen’x.com, discussion
Changes
- Lean/Mathlib community certification (statement read in the original), Sutherland’s verification log, replay at 62%, Tao on the Langlands programme, AHM statement, top X posts; status confirmed (Lean-verified), confidence high, title and summary updated
- Second wave: Aaronson post (quotes checked in the original); Sutherland’s independent build, Buzzard on the pending Comparator run, the leanprover-community Comparator repo, 46% replay status, Masullo/Vakil all-human GRH draft, main press and X reactions
- Read the primary note at alpo.ge/remark.pdf (the site address in one report, alpoge.ge, was wrong); added its details
- Inspected the released Lean (statement uses Mathlib’s RiemannHypothesis; no sorry, axioms or unsafe code found; not compiled)
- Added Rohan Arun’s community replay (owner’s tip)
- Filed (owner’s question) from same-day coverage