Levent Alpöge, a mathematician at Anthropic, claims that he and the AI model Claude have found a proof of the Riemann hypothesis. He made the claim in a statement posted on his personal website. Published on November 20, 2026, the release consists of a four-page explanatory text and Lean proof-assistant code that a third party's tally puts at about 17.4 million lines. A group working on independent verification has confirmed what the theorems in the code state, and has checked that the published files match the original. As of November 20, however, machine verification of the full proof was still in progress, and the Riemann hypothesis has not been accepted as resolved.
Zeros That Govern the Distribution of Primes
The Riemann hypothesis conjectures that all non-trivial zeros of the Riemann zeta function lie on the "critical line", where the real part equals 1/2. A zero is an input at which a function takes the value zero; the zeta function takes complex numbers as inputs. The real component of a complex number is called its real part, and its value indicates where a zero sits. The heart of the problem is whether every zero other than the "trivial zeros" at the negative even integers lies on the critical line.
Zeta zeros are tied to the distribution of primes because the function has two representations. According to an official explanation by E. Bombieri published by the Clay Mathematics Institute, in the region where the real part of the complex input exceeds 1, the zeta function can be written both as a series summing over all natural numbers and as the "Euler product" over primes. When the function is extended to a wider region by analytic continuation, the positions of its zeros turn out to determine how accurately the distribution of primes can be described.
Expressed in terms of the number of primes, the relationship is the following formula.
This error estimate holds if and only if the Riemann hypothesis is true. Here, is the number of primes up to , and is the logarithmic integral, which approximates the average distribution of primes. The trailing means that, for sufficiently large , the difference between the actual prime count and the approximation stays within a constant multiple of .
If the Riemann hypothesis were proved, it would place a strong limit on how far the number of primes can deviate from their average distribution. The notation, however, does not specify a concrete constant, so this formula alone does not let one directly calculate the number of primes up to an arbitrary number. Proposed in 1859, the Riemann hypothesis is not a vague question about whether primes follow some pattern; it is a problem of rigorously bounding the error in their distribution.
From the Idea of Multiplicity to Chaining Theta Lifts
Alpöge's statement names as the starting point of the proof Stepanov's argument, which solves the analogue of the Riemann hypothesis for function fields over finite fields. In that argument, "multiplicity", which counts how many times the same object appears, plays an important role. Alpöge says he wanted to apply this idea to the world of integers. He says he "believes" he found the proof jointly with Claude, and explains that he formalized the argument in Lean without adding any extra assumptions.
According to the statement, the research began with an idea from Fourier analysis in a simple setting and developed into a method that repeats an operation called the "theta lift". A theta lift is an operation that constructs one mathematical object from another; in this approach, the aim is to chain multiple lifts together to obtain an effect corresponding to multiplicity. For the von Mangoldt sum of length , which adds up primes and their powers with weights, Claude reportedly suggested chaining the theta lifts about times. The research timeline included in the statement records the idea of repeating lifts between different unitary groups to reproduce the role multiplicity plays in function fields.
That said, the four-page statement is a document explaining the ideas and research history behind the proof; it is not a paper in a form that lets mathematicians follow the entire argument. The timeline it contains is also an outline Claude produced by searching the research records, and Alpöge himself cautions that some of the explanations may need correction. He acknowledges that the proof contains a great deal of advanced mathematics, and that there are areas he does not fully understand himself. He says he intends to re-learn and simplify the argument, then explain it through talks and papers.
To understand this claim, it also matters how it differs from a result announced in August 2026. In a preprint (arXiv:2608.13637) released by Alpöge and Ralph Furman, they proved, without extra assumptions, a lower bound on the proportion of non-trivial zeros of the zeta function, counted with multiplicity, that are both "simple" and on the critical line. Simple means the zero has multiplicity 1. The bound obtained is at least 2/3, reaching 67.25% when the Montgomery–Taylor window function is used. These are asymptotic guarantees about the proportion among all zeros, not observed ratios from examining a finite number of zeros.
According to Anthropic's announcement of August 10, an unreleased research version of Claude was used for that result, consuming a total of 31 million output tokens across two sessions. Review by human mathematicians and formalization in Lean followed. The November statement, by contrast, does not reveal the name of the Claude model used, the computational cost, or any quantitative breakdown of the human and AI contributions.
Whereas the August result concerned the proportion of zeros on the critical line, this claim is that no zero at all lies off it. Even if the proportion of zeros on the critical line were to reach 100% asymptotically, a single exception would mean the Riemann hypothesis fails. This difference determines the significance of the present claim.
How Far Does the Generalized Riemann Hypothesis Extend?
Alpöge claims to have proved not only the ordinary Riemann hypothesis but also the generalized Riemann hypothesis (GRH). However, the published code covers the generalized Riemann hypothesis for Dirichlet L-functions. The list of theorems in the GitHub repository for independent verification includes, in addition to the ordinary Riemann hypothesis, statements about L-functions corresponding to Dirichlet characters.
According to verification notes published by the verifiers, the scope is not limited to primitive Dirichlet characters; it covers all Dirichlet characters for all moduli. One of the main statements says that in the critical strip, the region where the real part is greater than 0 and less than 1, every zero of an L-function has real part 1/2. It also includes the statement that L-functions do not vanish where the real part is greater than 1/2, with the trivial character at treated as an exception.
Broader classes of L-functions, such as Dedekind zeta functions, are not included in the published theorems. Alpöge's statement also mentions the extended Riemann hypothesis (ERH), but he is only at the stage of hoping that further research will prove it. One must distinguish between the scope the published code currently proves and the scope the author expects to extend to in the future.
How Do You Verify a Proof of About 17.4 Million Lines?
In a public GitHub Actions CI run executed on November 20, 2026, two checks passed. CI (continuous integration) is a mechanism that automatically runs predefined processes and records the results. The two that succeeded were confirmation that the distributed source code matches the archive published by the author, and confirmation that Lean can successfully compile the theorem statements prepared by the verifiers. Both are preparatory steps that precede verification of the full proof.
The second check used not the author's proof code but theorem statements that the verifiers wrote independently using only the mathematics library Mathlib. A theorem statement specifies what is to be proved, while the proof code describes by what logic the conclusion is derived. Successful compilation of the statements confirms that the mathematical propositions to be verified can be correctly expressed in Lean; it does not guarantee the correctness of the submitted proof.
According to the verification notes as of November 20, the build processing the entire proof was still running, and no report of completed independent re-verification had been published. The verifiers manually checked the theorem statements and also searched all source files by string, reporting that they contain no sorry, which marks an incomplete proof, and no axiom, which declares an additional axiom. However, finding no problems through string search is not the same as having Lean check the entire proof under its inference rules and confirm its correctness.
In the verifiers' tally, the published code consists of 175,187 Lean files totaling 17,374,215 lines. The verification repository separates the checking of theorem statements from the processing of the submitted proof, and further provides a mechanism for examining in an independent environment whether the proof and the statements match. The design uses a network-isolated environment and multiple checkers, and also checks that no unproved assumptions have slipped into the proof. Only the three standard axioms commonly used in Lean are permitted, and Alpöge's claim of "no assumptions" also presupposes this logical foundation.
Mechanically checking a proof of about 17.4 million lines and having human mathematicians understand its argument serve different roles. Independent verification checks whether the published code actually proves the proposition corresponding to the Riemann hypothesis and whether it depends on unproved assumptions or inappropriate definitions. A human-readable paper, on the other hand, is needed to understand what new methods resolved the obstacles that conventional mathematics could not overcome, and to assess the mathematical significance.
On the Clay Mathematics Institute's official page, the Riemann hypothesis is still listed as an unsolved problem. The institute's rules for the Millennium Prize Problems set three preconditions for recognizing a solution: publication in a qualifying outlet, at least two years having passed since publication, and general acceptance in the global mathematical community.
For now, the paper for mathematicians that Alpöge has announced he will publish, and the results of the independent verification group's check of the full proof, will be the key materials for evaluating this claim. The first thing to confirm is whether the published Lean code derives the ordinary Riemann hypothesis and the generalized Riemann hypothesis for Dirichlet L-functions without any additional unproved assumptions.
