Two Rival Claims on the Riemann Hypothesis Surface Hours Apart: One Built With Claude, One Without
Two independent GRH claims appeared hours apart; both face serious unresolved questions

On the morning of November 20, 2026, Levent Alpöge, a mathematician at Anthropic, quietly uploaded a compressed archive to Zenodo. The file weighed 239 megabytes. It contained roughly 175,000 files. Its README said the contents, when decompressed, amounted to something in the neighborhood of 250,000 printed pages of Lean 4 proof code—a machine-verifiable argument, he claimed, for the Generalized Riemann Hypothesis.
GRH is among the most consequential open problems in mathematics. It extends Bernhard Riemann's 1859 conjecture from the zeta function to the full family of Hecke L-functions, asserting that every nontrivial zero of these functions lies on the critical line Re(s) = 1/2. A proof would reshape number theory, cryptography, and our understanding of how prime numbers are distributed across the integers.
By late afternoon, a second claim had appeared. Ravi Vakil, the president of the American Mathematical Society and a Stanford algebraic geometer, posted a "hurried announcement" on his WordPress blog about a 940-page human-written manuscript titled The Music of the Primes by Alessandro Maria Masullo. Vakil said the proof was entirely human-written and that Stanford's Brian Conrad, a co-prover of the 1999 Modularity Theorem with Breuil, Diamond, and Taylor, had already reviewed it and provided feedback. By that evening, Vakil had updated his post to report that "several people have contacted me to report problems in the proposed proof."
The twin announcements crystallized, within a single day, the central challenge facing mathematics in the age of AI: how do you evaluate claims you cannot easily read?
The Machine-Checked Claim
Alpöge's archive is, in a narrow technical sense, unprecedented. The Lean 4 proof assistant checked it. That means a formal logic kernel verified that every step follows from the axioms. No human checked 250,000 pages; the machine did.
Mathematician and former Epoch AI lead mathematician Elliot Glazer confirmed that the central Lean challenge statement compiled successfully. Terence Tao, whose blog has become the field's de facto bulletin board for evaluating major mathematical claims, described the proof as "enormous" and noted that the strategy "appears to use the full power of modern developments on the Langlands program and arithmetic geometry." Tao also observed, with some relief, that "in this case at least we have a human author to communicate with"—a distinction he stressed in light of AI-generated proofs that arrive without an accountable author.
The technical approach Alpöge reportedly deployed is a generalization of the Stepanov method, using long chains of approximately √X theta lifts—morphisms between automorphic forms that thread through the Langlands program—to transfer a technique that had worked over function fields into the harder setting of number fields. It is, by any measure, a sophisticated strategy.
But Lean compilation is not the same as mathematical correctness. A proof can compile while relying on lemmas that have not themselves been verified, or while encoding a claim that is subtly different from the one it purports to prove. Compilation is necessary; it is not sufficient.
Read more: Anthropic's Claude produces first machine-checked proof of Fermat's Last Theorem
The Human-Written Claim
Masullo's 940-page manuscript takes a classical approach. Its core is Theorem 12.5.1.1, a trace formula linking Hecke L-functions to global operator-theoretic data. Chapter 13, as described by those who have read it, is meant to deduce GRH from that theorem.
Community review moved quickly. The AI math-auditing organization Refine published its assessment the same day, citing what it called "fundamental problems" in the manuscript. A separate paper posted to PhilArchive by researcher Jenny Lorraine Nielsen argued that Masullo's application of Montel's theorem in Sections 1.4.2 through 1.4.3 is flawed: Nielsen showed that nonnegative test-function approximants produce a divergent Archimedean integral for the trivial Dirichlet character over Q, contradicting the uniform boundedness the proof requires.
A Lean formalization attempt, placed on hold by its maintainer mccorvie on November 20, noted that Chapter 13 contained no formal statement of its main theorem. That is a significant structural gap in a 940-page argument whose final chapter is meant to be the proof.
The problems do not establish that Masullo's strategy is fundamentally wrong. They establish that the manuscript, as currently written, does not constitute a proof.
An Afternoon of Contrasts
The juxtaposition was hard to miss. One proof arrived as a Lean archive so large that most mathematicians cannot open it. The other arrived as a human-readable manuscript that mathematicians immediately began reading—and found problems within hours.
Both teams face versions of the same obstacle. For Alpöge, the question is whether the Lean code actually proves what its readme claims, and whether any human can verify the mathematical content of 250,000 pages at a reasonable pace. For Masullo, the question is whether the gaps that reviewers found can be closed.
Tao, writing on his blog, applied a phrase he had used at a Caltech talk weeks earlier about the broader pace of AI-driven mathematics: "We have to slow down. It's insane, the pace." His post on the dual claims extended that concern to proof evaluation itself—the field may be generating claims faster than it can responsibly assess them.
What the Clay Prize Requires
A correct proof of GRH would qualify for the Clay Millennium Prize, a $1 million award administered by the Clay Mathematics Institute. The prize rules require that a proof be accepted and published in a qualifying mathematical journal, then survive two years of community review before the prize committee acts.
Neither claim currently meets that bar. Alpöge's Lean archive, if the mathematics is correct, would still require human exposition of the core ideas before any journal could evaluate it. Masullo's manuscript has open counterarguments. Both proofs are, at this moment, unresolved.
A Field Reckoning With Its Own Speed
The dual announcements arrived five weeks after the Simons Foundation announced a $15 million initiative with Anthropic and OpenAI to help mathematicians navigate the AI era—each AI company contributing $5 million. The initiative includes workshops, proposal funding, and AI tool access for researchers. Its proposal deadline is December 1, 2026.
Read more: Anthropic prices at $83 and raises $81B, setting a new IPO record on Nasdaq
The funding reflects a consensus that the field faces an infrastructure crisis: institutions built to evaluate one proof per decade are now confronted with machine-generated archives measured in gigabytes and human manuscripts measured in hundreds of pages arriving in clusters.
Whether either GRH claim survives peer review, the events of November 20 made the nature of the problem concrete. Mathematics now has to decide not just whether a proof is correct, but who—or what—is qualified to say so.



