Claude and the Riemann Hypothesis Proof: What's Verified So Far
Anthropic's Levent Alpöge says he proved the Riemann Hypothesis with Claude and posted it as Lean code. What the checks show, what is still running, and why it matters.

Anthropic's Levent Alpöge says he proved the Riemann Hypothesis with Claude and posted it as Lean code. What the checks show, what is still running, and why it matters.

On Friday, November 20, 2026, Anthropic mathematician Levent Alpöge posted a short note on his website, remark.pdf ↗, that opens with a sentence few people expected to read this year: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He says the same method resolves the Generalized Riemann Hypothesis, and that he and Claude "have also formalized the arguments without assumptions."
There is no paper. Alpöge released the proof only as Lean ↗ code, split into 14 archive parts on his site. In a footnote he says a written version "hit 250,000+ pages," and that posting "some massive unreadable travesty of 'slop'" would have been wrong, so he "decided to only present Lean." He also writes that he messaged Anthropic CEO Dario Amodei "late Sunday night" before announcing, and calls the timing, the day after Anthropic's Nasdaq debut, "horrible."
Scientific American ↗ and TechCrunch ↗ both reported the claim the same day. The hypothesis dates to 1859 and is one of the Clay Mathematics Institute's $1 million Millennium Prize problems. Alpöge's note does not name which Claude model he used.
The image above is an original graphic made by aitoolranked. It is an illustration, not part of the proof.
The useful thing about a Lean proof is that anyone with enough hardware can check it. MIT mathematician Andrew Sutherland is publishing results on a verification page ↗ (checks run with Claude Code). As of its 09:30 UTC update on November 21:
| Check | Status |
|---|---|
| All 175,187 Lean modules compile | Pass, zero errors |
| Only Lean's three standard axioms used | Pass |
| Two independent kernels (nanoda, con-leche) | Both accept |
| Statement matches the challenge (con-leche compare) | Accepts all five theorems |
Official lake comparator run | Still running |
| Lean default and paranoid kernels on the export | Still running |
The page lists the formal statement: Mathlib's own RiemannHypothesis definition, plus the analogous result for Dirichlet L-functions. It adds that "GRH" here "means Dirichlet L-functions only." Sutherland also combined the proof with an independent formalization of Robin's criterion, which gives two inequalities about sums of divisors with no mention of the zeta function. Those compile too.
Sutherland told Scientific American, in a follow-up piece ↗: "Honestly the chance that anything is fishy here is approximately zero." Imperial College's Kevin Buzzard was more careful in the same article. The first question, he wrote, is whether "the statement of the theorem in the Lean code" matches the real theorem, and the comparator run is still needed to rule out code that redefines symbols or exploits an unknown kernel bug.
Hours later, Stanford's Ravi Vakil, president of the American Mathematical Society, posted a human-written draft by Stanford postdoc Alessandro Masullo, saying "it is all human, with no AI." TechCrunch says Masullo describes it as 12 years of work. That draft has no Lean version and has not been verified. Brian Conrad, a Stanford mathematician who has been giving Masullo feedback, told Scientific American ↗ it is "undergoing major revision."
"Oh my god: Anthropic mathematician Levent Alpöge (@alpoge) says he has proven the Riemann Hypothesis, open for 167 years." — Chubby ↗ (@kimmonismus), in a post with over 6,000 likes
"My Claude is downloading the Lean source code for the Riemann Hypothesis proof right now, and we're going to compile it and check the proof ourselves." — lastmjs ↗ (@lastmjs)
"Nobody is going to read it. So the review is two questions: does the Lean code compile, and is the theorem at the top really the Riemann Hypothesis." — Çağrı, KAM Studio ↗ (@KamStudioLabs)
"Claim of Riemann Hypothesis proof via lean slop" — Meryem Arik ↗ (@MeryemArik9)
Reddit was not reachable from our tools for this story.
This is the biggest AI capability claim of the year, and the checks so far favor it. A model claiming a famous result means little. A 175,187-module Lean proof that passes several independent kernel checks using only standard axioms is much harder to dismiss. Until the comparator finishes and number theorists can explain the argument, the accurate wording is "very likely verified," not "solved."
Expect the explanation to come later than the proof. Alpöge says himself that he has not "mastered" much of the mathematics involved and expects the argument to be "massively simplified." The proof could stand while human understanding takes months to catch up, which is the new problem this creates for mathematicians.
What it means for people who pick AI tools. This lands after OpenAI's own math push, including the GPT-6 Astra science results, and DeepMind's AlphaProof Code. For everyday work none of this changes which assistant writes your code or emails. Our ChatGPT vs Claude vs Gemini comparison and Claude Code review are still the better guides for that. It does show that Claude plus a formal checker can do research-grade reasoning that people can verify without taking it on trust.
Verdict: take it seriously, check back for the comparator result, and be wary of anyone who says "AI solved math." One mathematician, one model and one proof checker produced a result that other mathematicians are now verifying in public.