News 5 min read

Lean community certifies Riemann Hypothesis proof made with Claude

Last updated on


Levent Alpöge, a mathematician who works at Anthropic, says he believes he has proved the Riemann Hypothesis while working with Claude, and that the two have formalized the proof in the Lean proof assistant. He announced it on November 20, 2026 in a short essay that publishes the Lean code rather than a written paper. The same day, Stanford’s Ravi Vakil posted a separate, human-only proof of the Generalized Riemann Hypothesis by postdoc Alessandro Masullo. On November 21, the Lean and Mathlib communities certified Alpöge’s Lean proof as correct, so one of mathematics’ most famous open problems now has a machine-checked proof made with an AI model’s help. Vakil has since added that several people have reported problems in Masullo’s human proof, which is not accepted until they are resolved.

Update: the Lean community certifies the proof

In a joint statement endorsed by the Lean FRO and the Mathlib Initiative, the Lean and Mathlib communities say they have “full confidence” that Alpöge’s proof is correct. The statement says the proof is a Lean library of 17 million lines of code that depends directly only on Mathlib, and that the community spent the day after the release checking it. It lists four checks, following the criteria in a Xena Project post from November 12:

  • The statement is right. The theorem being proved is Mathlib’s own RiemannHypothesis, which Lean and number theory experts have vetted.
  • The library compiles. Dozens of people built it without errors and reproduced the same raw proof export file.
  • The code proves that statement. The community wrote a separate file that states the hypothesis using only Mathlib, and two independent tools (comparator and con-leche) confirmed that the theorem in Alpöge’s library is exactly that statement.
  • The proof is valid. The exported proof was checked by Lean’s own kernel, by the con-leche and con-ron checkers, and by several third-party checkers written in other programming languages. The statement says con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, which answers the worry that a flaw in Lean’s own logic could be exploited.

The checks ran on an inert export of the proof, so any malicious code hidden in the Lean sources could not affect them. A longer write-up of the verification is hosted at MIT. What is still missing is a human-readable proof: the statement says it looks forward to mathematicians explaining the new ideas in it.

What’s new

The claim. The Riemann Hypothesis, posed in 1859 and one of the Clay Mathematics Institute’s Millennium Prize problems, is about how evenly the prime numbers are spread out. Alpöge writes that he believes he has, “working with Claude, discovered a proof”, and that the method extends to the Generalized Riemann Hypothesis (GRH). He expects the Extended Riemann Hypothesis to follow with more time.

Lean, not a paper. Alpöge says the formalization “is finished” and that he and Claude formalized the arguments without assumptions. The Lean project is published as a 14-part archive linked from the essay’s footnote. He chose to release only the Lean code because the argument is currently very long, and he did not want to post a huge, unreadable document. He says he will now work on understanding it, simplifying it and writing it up properly, and expects shorter proofs to follow.

How Claude helped. The essay describes months of joint work, with Alpöge steering the overall strategy (combining two ideas from a simpler “toy” setting) and Claude supplying many of the key technical steps near the end, in areas where Alpöge says he has little mastery himself. A rough timeline Claude put together starts on October 6, 2026. The essay does not say which Claude model was used.

Timing. Alpöge writes that the breakthrough happened late on a Sunday night and that he messaged Anthropic CEO Dario Amodei about whether he could announce it. He calls the timing “horrible” and says he alone chose not to wait longer.

A second, human proof. Vakil’s post shares a draft by Masullo titled “The Music of the Primes”, with a proposed proof of GRH. Vakil writes that “it is all human, with no AI”, that Brian Conrad has been reading it for months, and that it was released early as a draft so others can check and formalize it. In an update about eight hours later, Vakil wrote that several people had reported problems in the proposed proof, found with the help of AI models, and that all issues must be resolved to the community’s satisfaction before it would be accepted. TechCrunch reports that Masullo spent 12 years on it.

More AI-assisted results. TechCrunch also reports that Alpöge and NYU mathematician Tristan Buckmaster released a separate result on the Navier-Stokes equations in a public GitHub repository.

What it means for developers

Lean is how this got checked. Because the proof is machine-checked Lean code, anyone could compile it, and the Lean community’s Zulip thread worked through it within a day. The usual open questions for a formal proof, whether the Lean statement says exactly what the hypothesis says and what it depends on, are the ones the community’s statement answers: Mathlib’s own statement, and it depends directly only on Mathlib. Mathematicians still want a human-readable argument, which Alpöge says does not exist yet.

AI is moving from benchmarks to open problems. This follows Anthropic’s report this week that Claude agents solved two open problems in black hole physics. The pattern in both is the same: a long-running model does much of the technical work, and the result is checked by something other than the model, here a proof assistant.

Try Lean yourself. If you have never used it, Lean 4 is a programming language as well as a proof assistant, and the formal proof here is ordinary source code you can build and read.

How it compares

AI models have been contributing to research mathematics for a while, including OpenAI work TechCrunch covered in September. Alpöge himself credits OpenAI’s mathematical results for showing him that Navier-Stokes was within reach. The Riemann Hypothesis is a far older and more famous target, and two independent proofs posted on one day, one with AI and one without, is unusual. Alpöge’s Lean proof has now passed the Lean community’s checks; people have reported problems in Masullo’s draft, and it is not accepted until they are resolved.

Sources