Levent Alpöge, a mathematician at Anthropic, says he worked with Claude to prove the Riemann Hypothesis, an open problem since 1859. The proof was reportedly formalized in Lean, but no one has yet confirmed its validity: before calling it a solution, outside researchers must check both the code and the statement it proves.
The claim appeared on November 20 on Alpöge’s website. Scientific American reproduces a first-person statement in which he says he discovered the proof while working with Claude. Crypto Briefing also places the announcement there and notes that the formalization is written in Lean. OfficeChai, by contrast, describes an unsigned document posted on the mathematician’s personal website, which Alpöge reportedly did not share on social media.
“I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.”
— Levent Alpöge, mathematician at Anthropic
An extraordinary claim, not an accepted result
The hypothesis was proposed by Bernhard Riemann in 1859 and concerns the zeros of the Riemann zeta function. In its classical form, it states that the nontrivial zeros lie on a specific line, known as the critical line. The distribution of prime numbers is closely connected to this problem, which has become one of the great open questions in number theory.
The scale of the claim matters. The Clay Mathematics Institute offers a one-million-dollar prize for a correct proof, as one of its Millennium Prize Problems. OfficeChai notes that thousands of number theory results have been proved on the assumption that the hypothesis is true. If the proof holds up, those dependencies could then rest on a proven result; for now, there is no basis for taking that step.
Alpöge also claims that the argument proves the generalized Riemann Hypothesis, an extension covering a family of related functions, Dirichlet L-functions, according to OfficeChai. Scientific American also highlights this generalized version and notes that the announcement describes far-reaching mathematical consequences. It is not, however, an immediate practical application: the publication emphasizes that the result would not speed up the internet or directly transform global banking.
Lean helps check steps; it does not decide what has been proved
The proof was not presented as a conventional mathematical paper. Crypto Briefing explains that there was no lengthy preprint or journal submission: the claim is accompanied by a formalization in Lean. OfficeChai describes a document titled “Remark” and 14 compressed code files hosted on Alpöge’s personal website.
Lean is a proof assistant: it allows mathematical arguments to be expressed in a format that software can check step by step. This offers a different way to verify a proof than simply reading and evaluating a prose exposition. But mechanical checking does not eliminate the essential question: whether the final encoded theorem really corresponds to the Riemann Hypothesis and whether the proof relies on acceptable assumptions.
OfficeChai says the document claims that the formalization is complete and uses no additional assumptions, but also notes that outside researchers had not yet confirmed that the files compiled or that the formal result addressed the right problem. Crypto Briefing likewise reports that no independent verification had been reported. That is the dividing line between a potentially verifiable proof and one accepted by the community.
The review would therefore involve two related but different tasks: checking that the Lean project builds and examining exactly what the final theorem asserts. Specialists would also need to scrutinize the definitions on which it depends. As Crypto Briefing summarizes, the checker can confirm the logical validity of a formalized proof; people must determine whether that proof addresses the question it is supposed to resolve.
A lengthy strategy and a contribution attributed to Claude
OfficeChai provides details about the method outlined in the document. According to that publication, the strategy draws on ideas from function fields, a setting in which an analogue of the hypothesis has already been proved. There, the Stepanov method is used to construct an auxiliary function with high multiplicity. The difficulty in transferring this approach to the classical case is that it depends on derivatives, which have no natural equivalent for the integers.
The text credits Claude with a way around this obstacle: a chain of “theta lifts” that would transfer automorphic forms between symmetry groups. Each step would add one unit of multiplicity; the chain would be approximately the square root of the size of the primes under consideration in length. OfficeChai also mentions other number theory tools in the argument, including relative trace formulas, Iwasawa theory, Bessel periods, and a weakened version of Vandiver’s conjecture.
These elements describe what the document claims, not an independent reconstruction of the proof. OfficeChai reports caveats in the text itself: the argument would be very long, some parts contain open questions or intuitions, and the explanation of the sections attributed to Claude may not be clear. It also says that the informal AI-generated material would exceed 250,000 pages. Without a public review of the code and a mathematical exposition examined by specialists, this description cannot be taken as confirmation that the strategy works.
The context of a year of advances attributed to models
The news follows other mathematical results linked to Alpöge and AI models. Scientific American recalls that in July he published a Claude-generated solution related to the Jacobian conjecture. Crypto Briefing describes that result as an explicit counterexample and adds that in August Alpöge presented a positive solution to the Hopf problem. OfficeChai also mentions a refutation of the exponential time hypothesis shared in October.
The closest precedent is the work on the zeros of the zeta function. Crypto Briefing credits an unpublished Claude model with improving, in August, the known proportion of zeros on the critical line from 41.6% to 67.2%. OfficeChai puts the related improvement at a jump from 41% to 67%. In October, that publication adds, OpenAI released a set of 722 AI-generated manuscripts, including a paper on the “Quasi-Riemann Hypothesis” that bounded the region where the zeros could lie, without proving the hypothesis.
This track record helps explain why the claim is attracting attention, but it does not validate it by accumulation. Each advance must be evaluated on its own arguments. The difference between bounding where certain zeros may lie and proving a general statement about all relevant zeros is precisely what makes the new claim so extraordinary.
The cost of a proof that arrives as a black box
Formalization can make checking each logical step more efficient, but there is a cost if the result circulates only as code that is difficult to interpret. OfficeChai reports concerns that AI proofs could become black boxes: demonstrating that a claim is true without helping mathematicians understand why. The document credits Claude with central parts of the reasoning and, according to the same publication, Alpöge acknowledges that the model outperformed him in certain sections.
That is the dilemma for AI-assisted research: a formal verifier could provide a strong guarantee of logical correctness, while a readable explanation remains necessary for other researchers to understand the structure and value of the argument. OfficeChai contrasts the announcement with a call by 25 Fields medalists to publish results more cautiously and with an adequate exposition. The goal is not just to accelerate the production of theorems, but to make their scrutiny and reuse possible.
For companies developing models, the case offers a public test of the capabilities attributed to their systems, but it also raises the bar for evidence. A claim of this magnitude is not established by the author’s prestige, the presence of code, or a model’s involvement: it will depend on whether third parties can review the project and confirm what it proves. The sources do not point to any immediate commercial impact; for now, the impact described is scientific and methodological.
The concrete next step is for outside researchers to obtain the files, check whether the Lean project compiles, and examine the final theorem and its assumptions. Until then, what exists is a claim pending review, not an accepted proof of the Riemann Hypothesis.
Sources: Scientific American, Crypto Briefing, OfficeChai