← Back to all posts
News

Anthropic's Levent Alpöge Claims a Claude-Assisted Proof of the Riemann Hypothesis

November 21, 2026 · 05:20 UTC · News
On this page 9 sections
Anthropic's Levent Alpöge Claims a Claude-Assisted Proof of the Riemann Hypothesis

TL;DR

On November 20, Levent Alpöge, a mathematician at Anthropic, posted a four-page note claiming that he and Claude have discovered a proof of the Riemann Hypothesis. In his words: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis... the method easily extends much further, resolving the so-called Generalized Riemann Hypothesis." He expects the Extended version to follow, and says the arguments are formalized in Lean "without assumptions." There is no paper. The deliverable is 175,187 Lean files, about 17.4 million lines, mirrored on GitHub with a build recipe and an axiom check, and no one has independently compiled it yet. Terence Tao relayed the claim alongside two more results from the same group, Scott Aaronson is betting the proof stands, and an all-human GRH draft rushed out the same evening already has reported gaps, found with AI assistance.


At a glance
A claimed proof shipped as a buildable artifact
17.4M
lines of Lean behind five theorems
175,187
lean files, 1.3 GB checked out
0
sorry, admit, axiom, native_decide in a text scan
167y
since Riemann's 1859 memoir
3
claimed results from one group in a day
0
independent builds published so far

What the note actually says

The note is four pages, half confession. It opens with the claim, then assigns credit: "Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line." Alpöge says he had earmarked this problem for himself decades out, spent recent months touring tools with Claude, and guided the search to the right analogy before the model pulled away from him.

The problem itself is 167 years old and one of the Clay Mathematics Institute's million-dollar Millennium Problems. In its usable form it says the count of primes up to n stays within a square-root-sized error of n/ln(n), and a long shelf of number theory is proven only conditionally on it.

Then the timing. The note went up one day after Anthropic's record IPO, in the week the Nasdaq gave back all of its 2026 gains. Alpöge says the finish arrived late on Sunday night, November 15, that he messaged Anthropic chief executive Dario Amodei right away ("Hypothetically, if I were to have..."), and that "there was a constraint to respect" on announcing. His preemption is verbatim: "There is an obvious cynical view, to do with the timing, but it is false." The sequencing you can judge; the artifact you can check.

There is no paper for a reason. Alpöge says he generated the readable version anyway and "hit 250,000+ pages," and that its Theorem 1 opened by observing that because the reals are totally ordered, a zero off the critical line must sit on one side of it or the other. He calls that document "slop" and writes that he "decided to only present Lean."

oct 6 oct 20 nov 15 nov 16 nov 19 nov 20 stepanov determinant idea switch to the von mangoldt sum theta-lift chain machinery lands grh, per the note's timeline anthropic ipo (context) note posted, lean published timeline as listed in alpo.ge/remark.pdf; ipo date shown for context

The artifact: five theorems, one pinned definition

The proof landed as fourteen tarball parts on Alpöge's site and is now mirrored at CocoSgt/riemann-hypothesis-lean: 175,187 Lean files, about 17.4 million lines, 1.3 GB checked out, SHA-256 published. The mirror's status line is the one to keep pinned: "As of mirroring (2026-11-20), no independent build or review of this code has been published. Treat it as a claimed proof until it has been compiled and checked."

What gets proven is short. Five theorems in src/Main.lean: the Riemann Hypothesis, zeros on the line for zeta, and three nonvanishing and zero-on-line statements for Dirichlet L-functions, which together are GRH for that class. The headline statement is Mathlib's own RiemannHypothesis definition, pinned to one commit, and the other four statements live in a 16-line file that imports nothing but Mathlib. The exam question comes from the textbook, not from the student, which shuts the classic loophole where a formalization proves something adjacent to the famous claim.

where the 17.4M lines live
One directory holds 70 percent of the proof
lean librariesdevelopment
development
12.2M
formalism
3.2M
ground theory
1.1M
reciprocity
563K
companion
194K
positivity
55K
Lines per directory of the mirrored src tree; statements and build aid are rounding error. Source: mirror README.

The checking procedure is concrete enough to run this weekend. Clone the tree or reassemble Alpöge's parts and verify the hash. Fetch the pinned Mathlib cache, then lake build Main; the lakefile sets maxHeartbeats = 4000000 and --tstack=400000, so bring cores and RAM, and expect it to run a while. Then run #print axioms on the five theorems: a clean result lists only Lean's three standard axioms, and anything else means the proof rests on something unproved. For the paranoid, lean4checker replays the compiled environment at the kernel level, so the verdict does not depend on the elaborator. A static text scan at mirroring time found zero sorry, zero admit, zero axiom declarations, zero native_decide, and zero unsafe or opaque across all 175,187 files.

how to check a 17-million-line proof
The mirror's procedure, in order
Fetch
Clone or reassemble the parts
verify the SHA-256, 1.3 GB tree
Build
lake build Main
pinned Mathlib, 4M heartbeats
Axioms
print axioms on five theorems
clean lists three standard axioms
Replay
lean4checker at kernel level
verdict skips the elaborator
Report
Publish hash, toolchain, output
first public build wins
The mirror asks anyone building it to log wall time and peak memory. No such figures exist yet.

The README ends with a section titled "For agents: suggested procedure," a numbered checklist telling an AI how to verify the proof and report the commit hash. The build instructions are written for machines. The ideas are still waiting for mammals.

Tao's post: the claim came in a set of three

Terence Tao relayed the news the same day, listing three results "in decreasing order of how 'digested' the results currently are." Most digested: Alpöge, Tristan Buckmaster, and Matei Coiculescu, "with heavy AI assistance," have a human-readable proof that the unforced incompressible Euler equations blow up in finite time in 3D, formalized in Lean. Next: Alpöge and Buckmaster claim finite-time blowup for the unforced 3D Navier-Stokes equations, the question behind one direction of a Clay Millennium Problem, again with a Lean formalization. Tao says that strategy follows an expected path, computing an approximate discretely self-similar solution and then converting it with rigorous perturbation theory, but "more work needs to be done to create a proper, professionally written proof."

The GRH proof he ranks least digested: "an (enormous) Lean-only proof," generated, as he puts it, "using an internal AI model." He points to the Lean Zulip thread as the place to follow verification, says early analysis suggests the argument "appears to use the full power of modern developments on the Langlands program and arithmetic geometry," and forecasts that "the digestion process will take some time, though in this case at least we have a human author to communicate with."

If the cast looks familiar, it should. Alpöge and Buckmaster were the other side of September's Navier-Stokes standoff, when OpenAI published its own 10,000-agent blowup claim and Buckmaster published the dispute over priority. OpenAI's math dump, which AI Bacon covered in October, included a zero-free region Re(s) > 7/8: a fixed strip to the right of the critical line, cleared forever. Alpöge's claim is that every zero sits on the line itself.

Aaronson's ledger, and the kernel worry

Scott Aaronson wrote the skeptic's reference take. His worry is the checker, not the author: Lean's kernel has historically had bugs that "allowed proofs of 0=1," and the soundness of the underlying type theory is not fully settled, so "a sufficiently motivated AI could almost certainly exploit them to prove a false statement." Against that he sets a ledger: "so far, not one of the AI results with Lean certificates has needed to be retracted (while several without Lean certificates have needed to be)." Verdict: "I'm not 100% sure, but my bet is that the proof will stand."

That ledger has a data point of our own. OpenAI's math repository withdrew three manuscripts in October over a sign error, and all three were unformalized drafts, which is the exact line Aaronson draws. On the kernel worry, Lean contributor Joachim Breitner answered in Aaronson's comments that external checkers with their own consistency proofs, he names con-leche, now cover both the bug and the theory questions for anyone who wants a second opinion.

Acceptance is a separate matter from correctness. Mathematicians "won't be satisfied with 17 million lines of Lean," Aaronson writes; they will be satisfied "when, and only when, they're able to explain and distill and reconstruct the ideas for themselves. And that will probably take years, since human mathematicians move only at biological speeds." His nine-year-old, informed that RH had fallen, "replied that he's not even surprised at this point."

The same evening, the all-human counterclaim

Hours after Alpöge's note, Ravi Vakil posted from an AMS trustees meeting: mathematician Alessandro Maria Masullo had a proposed GRH proof of his own, checked closely for months by Brian Conrad and "so far looking good to him." Vakil's distinguishing property, in his own words: "It is all human, with no AI." The plan had been a year of polishing with Conrad before announcing; the AI news pulled the schedule forward.

About eight hours later came the update: "Several people have contacted me to report problems in the proposed proof, identified with the assistance of different AI models." Vakil adds the fair caveat that every issue must be resolved to the community's satisfaction either way. Still, on Friday the machine-assisted proof shipped with a machine-checkable certificate, and the human one shipped with machine-found gaps.

What to steal if you ship AI-generated work

  • Ship the checkable artifact, not the prose. Alpöge had a 250,000-page document and posted the Lean instead. If your agent's output reduces to something a machine can verify, that reduction is the product.
  • Pin the statement outside the generator's control. Mathlib's definition at a pinned commit closes the "you proved something else" loophole. The same hygiene belongs in agent evals: the grader comes from a repo the agent cannot touch.
  • Make verification a protocol. Hash, build command, axiom check, kernel replay, report format. The mirror's checklist is a working template for publishing anything AI-made under audit, and it still reads fine when the auditor is another agent.
  • Answer "why now" with the artifact. A result like this lands inside a comms calendar whether you like it or not. The timing question never fully dies; the compile command is the only reply that matters.

Read it with these caveats

  • No independent build or review has been published. Until someone runs lake build Main and posts the axioms output with a commit hash, this is a claim, and the mirror says so itself.
  • The author hedges first. He writes that he is "presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics," that he "may have baked lots of strange suboptimal choices into the argument," and that he expects it to be "massively simplified, likely immediately."
  • Scope discipline: RH and GRH for Dirichlet L-functions are what the five theorems cover. The Extended RH is an expectation, and the Euler and Navier-Stokes items are younger still, with the Navier-Stokes writeup explicitly unfinished.
  • A Lean certificate rules out gaps in the proof, not a mis-stated theorem. That is why the 16 lines of statements, and Mathlib's pinned definition, will get the hardest reading of all.
  • "Stands" is Aaronson's bet, not a verdict. The Clay Institute has a process, and the community's acceptance runs on its own clock.

Key Takeaways

  • On November 20, Anthropic mathematician Levent Alpöge posted a four-page note claiming a proof of the Riemann Hypothesis with Claude, extending to the Generalized Riemann Hypothesis, with the Extended version expected; the deliverable is 175,187 Lean files, about 17.4 million lines, mirrored on GitHub with build and axiom-check instructions.
  • The headline statement is Mathlib's own RiemannHypothesis definition at a pinned commit; a static scan found zero sorry, admit, axiom, native_decide, unsafe or opaque across the tree; and the mirror's own status is "treat it as a claimed proof until it has been compiled and checked."
  • Terence Tao relayed the claim in a set of three: a human-readable-plus-Lean blowup for unforced 3D Euler (with Buckmaster and Coiculescu), a Lean-formalized blowup for unforced 3D Navier-Stokes (with Buckmaster, writeup unfinished), and the Lean-only GRH proof, early-read as full-strength Langlands and arithmetic geometry.
  • Scott Aaronson's bet is that the proof stands; his ledger notes no Lean-certificated AI result has been retracted while several uncertificated ones have. Mathematician acceptance, he adds, will take years at biological speed.
  • The same day, Ravi Vakil rushed out Alessandro Maria Masullo's all-human GRH draft, vetted for months by Brian Conrad; within about eight hours, several people had reported problems in it, identified with AI assistance.
  • Alpöge announced one day after Anthropic's record IPO. He says he told Dario Amodei the night of November 15, that a constraint applied, and that "the obvious cynical view" of the timing "is false."

Sources: Alpöge's announcement note, Lean proof mirror, Terence Tao, Scott Aaronson, Ravi Vakil on Masullo's GRH draft, The New York Times, Quanta Magazine, Euler blowup Lean repo, Navier-Stokes Lean repo

AIAnthropicRiemann HypothesisLeanClaudeFormal VerificationMathematicsTerence Tao
CONSOLE
$