IA · 21 November 2026 · 4 min read
Has the Riemann Hypothesis Fallen After 167 Years? The Claude Proof Formalized in Lean
In brief: Anthropic mathematician Levent Alpöge has announced a comprehensive proof of the Riemann Hypothesis developed in collaboration with Claude. Completely formalized in the interactive theorem prover Lean, the work tackles the most prestigious open challenge in mathematics, dating back to 1859. The milestone highlights a shift where frontier AI reasoning moves past statistical approximations and subjects itself to deterministic verification by code compilers.
by Team Mocchi's
The Riemann Hypothesis, first proposed in 1859 by German mathematician Bernhard Riemann and long regarded as the pinnacle of pure mathematics, may have finally met its match. Levent Alpöge, an accomplished mathematician and researcher at Anthropic, published a detailed proof on his website developed in close partnership with Claude. Unlike past claims that relied solely on conventional academic prose, Alpöge’s proof was fully drafted and formalized in Lean, an open-source programming language designed to verify mathematical logic mechanically.
The hypothesis asserts that the distribution of prime numbers is intimately tied to the non-trivial zeros of the Riemann zeta function. Unsolved for over sixteen decades, it stands as one of the seven Millennium Prize Problems designated in 2000 by the Clay Mathematics Institute, each accompanied by a one-million-dollar reward.
A 167-year quest and a proof written in Lean
In an essay detailing the breakthrough, Alpöge revealed that he contacted Anthropic CEO Dario Amodei late last Sunday night to plan how to communicate the findings. He elected to wait until Anthropic had concluded its historic initial public offering on Wall Street before unveiling the work. His claim extends beyond the traditional formulation: Alpöge stated that the framework extends naturally to settle the Generalized Riemann Hypothesis (GRH) and could soon encompass the Extended Riemann Hypothesis as well, without unproven assumptions.
As reported by TechCrunch, the mathematics community immediately converged on Zulip discussion channels. Prominent number theorists and Lean Mathlib contributors, including East Anglia’s Chris Birkbeck and MIT’s Andrew Sutherland, began parsing the repository. The definitive advantage of a Lean formalization is transparency: if the interactive theorem checker compiles the code without syntactic or logical violations against core axioms, human reviewers can confirm the internal consistency of the derivation far more reliably than reading hundreds of pages of dense manual scratchwork.
Brute force versus human-machine collaboration
This development stands in sharp contrast to recent high-profile attempts to deploy artificial intelligence within advanced mathematics. As noted by The Verge, OpenAI claimed significant headway on the Navier-Stokes existence and smoothness problem this past September by marshalling a compute-heavy cluster of ten thousand autonomous agents, a maneuver that triggered debate within academia over unpublished research uploaded to generative tools.
By comparison, Alpöge’s work with Claude mirrors an intimate partnership between an experienced scholar and an advanced reasoning model used to explore lemmas and translate arguments into rigorous code. Notably, Anthropic itself struck a cautious note merely months ago: in August, the lab acknowledged Claude’s experiments on zeta functions while expressly warning that current techniques were unlikely to crack the full hypothesis anytime soon. Alpöge’s paper directly upends that modest forecast.
Peer review and the Clay Institute timeline
Despite the immediate excitement, official recognition will require patience. The Clay Mathematics Institute mandates that any proposed Millennium Prize solution must be published in a peer-reviewed journal of international standing and undergo a mandatory two-year public review period to ensure broad community acceptance and rule out hidden cyclical logic.
Over the past twenty-five years, only one Millennium Prize problem has crossed that finish line: the Poincaré conjecture, resolved in 2002 by Russian mathematician Grigori Perelman, who famously declined both the Fields Medal and the cash prize. If the formalized Lean codebase provided by Alpöge and Claude withstands the scrutiny of global peers, it will represent the most consequential intersection of automated deduction and frontier AI to date.
Mocchi's take
This breakthrough offers a vivid blueprint for the future of mission-critical software engineering and artificial intelligence. For teams building production architectures, the real lesson here is not simply that a large model produced novel ideas, but that those ideas were held accountable to a deterministic verification system like Lean. We cannot cure hallucination and non-determinism merely by expanding model parameters; the answer lies in marrying probabilistic exploration with strict, formal compilers. When building advanced enterprise software, this dual-engine approach is the gold standard: allow intelligent agents to conceptualize and propose, but mandate that formal verification layers certify every critical step before deployment.