A claim large enough to shake mathematics
The Riemann Hypothesis has long occupied a special place in modern mathematics: it is a compact statement about the zeros of the zeta function, but its consequences reach deep into the distribution of prime numbers and the architecture of analytic number theory. On November 20, 2026, that problem entered a new and highly unusual phase when Levent Alpöge, a mathematician at Anthropic, posted a short note saying he had found a proof while working with Claude, Anthropic’s AI system .
The headline is simple; the reality is not. Alpöge’s note says the method does not merely resolve the classical Riemann Hypothesis, but also extends to the Generalized Riemann Hypothesis, a stronger family of statements concerning Dirichlet L-functions . He also says the argument has been formalized “without assumptions” in Lean, a proof-assistant language designed to let computers check whether every logical step follows from accepted definitions and prior results .
If confirmed, this would be one of the largest mathematical events of the century and arguably the most dramatic example yet of AI-assisted discovery. But as of the latest reporting within the first news cycle, the word “claim” remains essential. The proof has not yet become a theorem accepted by the mathematical community, and the path from a Lean archive to a settled Millennium Prize Problem is steep .
What Alpöge released
The release is not a conventional mathematics paper. Alpöge posted a four-page statement on his personal website, along with fourteen archive files containing the Lean formalization . In the note, he writes that producing a fully written human-readable version would be routine in principle but that the result would be huge and unreadable; he says an informal written rendering exceeded 250,000 pages .
That choice matters. Mathematicians usually judge a major proof not only by whether each step is valid, but also by whether the argument creates understanding: why the theorem is true, which structures matter, and how the proof fits into existing theory. In this case, the public mathematical object is primarily code. Quanta reported that Alpöge had made available about 17 million lines of Lean, which he says formalize a proof of the stronger generalized result .
A public GitHub mirror of the release describes the archive as an unmodified copy of Alpöge’s Lean code and identifies the main top-level targets: a theorem named riemannHypothesis, statements about zeros of the Riemann zeta function, and statements about zeros and nonvanishing for Dirichlet L-functions . The mirror also states that the project contains 175,187 Lean files and roughly 17.4 million lines, underscoring the scale of the verification problem .
Why Lean changes the first question
Lean does not make a proof automatically meaningful, elegant, or explanatory. It does, however, change the first question reviewers ask. In a traditional proof, experts must read prose and symbols, infer intended definitions, and test the argument for gaps. In a formal proof, reviewers first ask whether the code compiles against the proof assistant’s kernel and whether the final theorem is stated correctly.
That is why the initial verification can move faster than peer review, but also why it can be misunderstood. If Lean accepts a proof of a theorem, that is very strong evidence that the theorem follows from the formal definitions and axioms supplied to Lean. It is not, by itself, a guarantee that the theorem formalized in code perfectly matches the mathematical claim everyone has in mind. Reviewers must therefore check both the mechanics and the translation.
The GitHub mirror explicitly warns that, as of mirroring on November 20, no independent build or review of the code had been published and that the release should be treated as a claimed proof until it has been compiled and checked . Quanta likewise reported that Lean experts were racing to verify whether the code works as claimed, and that the mathematical community would still need to determine what the formal result means .
What Claude is said to have done
The most striking part of Alpöge’s account is the role assigned to Claude. He says he guided the work toward a setting that combined two ideas from a toy model, but that Claude accelerated near the end and supplied decisive technical mastery in areas where Alpöge says he was less expert . The note’s internal timeline runs from early October to November 16, when it marks “GRH,” suggesting that the claimed breakthrough came after a burst of work over roughly six weeks .
The described mathematics is highly advanced. Alpöge’s sketch invokes Stepanov-style arguments, theta lifts, automorphic forms, Iwasawa towers, trace formula ideas, Bost slope inequalities, and other tools from modern number theory and arithmetic geometry . Those references are not a readable proof; they are a compressed map of what the proof is supposed to contain. Alpöge himself acknowledges that parts of the summary may be rough and that the Lean formalization is meant to clarify the details .
This is why the announcement is so different from an ordinary AI benchmark. The claim is not that a chatbot answered a contest problem. It is that a frontier model functioned as a collaborator on a sprawling research program, helping produce a formalized object large enough that no person has yet publicly absorbed it in full.
The current state: excitement, caution, and verification
The current state is best summarized in three points. First, a real, named mathematician at Anthropic has publicly claimed an AI-assisted proof of the Riemann Hypothesis and the Generalized Riemann Hypothesis . Second, the public proof object is Lean code rather than a conventional paper, and independent experts must verify both that the code compiles and that its formal statements correspond to the intended theorems . Third, no report in the first wave of coverage established that the claim had already passed the mathematical community’s full review .
The caution is not merely formalism. History is full of claimed proofs of famous problems that did not survive scrutiny, and Riemann is in the most dangerous category of all: a famous, compact, extremely hard conjecture where a single mistranslated definition, hidden assumption, or gap could invalidate everything. The presence of Lean reduces some risks but introduces others, especially the risk that observers confuse compilation of some files with acceptance of the entire mathematical claim.
AFP-based French coverage reported that Brian Conrey, a major figure in work on the Riemann Hypothesis, called the moment monumental while also not yet having verified the proof . That combination—astonishment without certification—is the right tone for the story.
Anthropic, timing, and the company context
The announcement also arrived under unusual corporate circumstances. SiliconANGLE reported that Anthropic had not published an official announcement about Alpöge’s claim, and noted that the company had floated on Nasdaq a day earlier in a public offering that raised more than $93 billion . Alpöge’s own note acknowledges that the timing was awkward and says he consulted Anthropic CEO Dario Amodei before announcing .
That context should not be used to dismiss the mathematics, but it should inform how the story is read. A claimed proof of Riemann made with a company’s own AI system is simultaneously a scientific claim, a reputational event, and a market-sensitive narrative. The right response is not cynicism by default; it is rigorous separation between publicity and verification.
Anthropic’s earlier public work on Claude and the zeta function also forms part of the backdrop. The company had already presented Claude as capable of making progress on related mathematics, but the new claim is qualitatively different: moving from progress on a related bound to a claimed proof of Riemann and GRH is a leap from impressive assistance to possible landmark discovery .
What would make the claim official?
There are several layers of acceptance. The first is technical: independent Lean experts need to reproduce the build, inspect the axioms used, and verify that the final statements are the intended Riemann and generalized Riemann statements . The second is mathematical: number theorists need to understand the strategy, identify the core innovations, and determine whether the formal proof corresponds to a robust conceptual argument. The third is institutional: for the Clay Millennium Prize, a solution normally must be published and withstand sustained scrutiny, not merely appear as a code release .
Numerama emphasized that the million-dollar prize is not immediate even if the proof ultimately stands, because the Clay process requires publication in a recognized venue and time for acceptance by the community . That is important for readers outside mathematics: a proof can be formally checked before it is prize-eligible, and it can be exciting before it is official.
Why this story matters even before the verdict
Even if the proof fails, the event matters. It shows how close mathematics now sits to a new research workflow in which human taste, machine search, formal verification, and massive code artifacts interact. Alpöge’s note presents a model in which the human mathematician supplies taste and direction, while Claude navigates technical terrain at a speed or breadth the human collaborator describes as hard to follow .
If the proof succeeds, the implications are profound. The Riemann Hypothesis would move from conjectural foundation to theorem; many conditional results across number theory would need to be reclassified; and AI-assisted formal mathematics would gain its strongest example to date. If the proof fails, it may still teach researchers where formalization pipelines, model-generated arguments, or mathematical translations break down.
For now, the responsible headline remains the one that began this story: an Anthropic researcher claims to solve the Riemann problem with Claude. The next chapter belongs not to headlines, but to Lean logs, theorem statements, expert audits, and eventually a human-readable explanation that tells mathematicians not only that the proof works, but why.
