Tech • AI • Robotics • Game

ENFR

Full article — scored 10/10

Anthropic Mathematician Claims to Solve Riemann Hypothesis With Help From Claude

Levent Alpöge, a mathematician at Anthropic, says he and Claude have produced a Lean-formalized proof of the Riemann Hypothesis and a broader generalized version. The claim is extraordinary, the first machine checks are encouraging, and the mathematical community is now separating three questions: does the code compile, does it state the right theorem, and can humans understand why the proof works?

Sign in to follow
Generated November 21, 2026 at 3:29 AM≈ 7 min read
Original source — The New York Times

The claim now shaking mathematics

An Anthropic mathematician has made one of the most consequential claims possible in pure mathematics: that the Riemann Hypothesis, the 167-year-old conjecture governing the hidden order of prime numbers, has been proved with the help of Claude. In a four-page note posted on November 20, 2026, Levent Alpöge wrote that he believed he had “working with Claude” discovered a proof of the Riemann Hypothesis, and that the same method extends to the Generalized Riemann Hypothesis for Dirichlet L-functions .

That sentence is not yet the same thing as a theorem accepted by mathematicians. But unlike many past claims around the Riemann Hypothesis, this one arrived with a substantial technical artifact: a multipart Lean archive that Alpöge says formalizes the argument without additional assumptions . Lean is a proof assistant, meaning that a proof can be checked step by step by a small trusted kernel rather than merely read as prose by specialists. That changes the usual verification path: the first wave of scrutiny is computational, while the deeper wave is conceptual.

The announcement immediately put number theorists, formalization experts and AI researchers into a rare shared emergency. If the proof stands, it would settle a Clay Millennium Prize problem and one of the defining questions of analytic number theory. If it fails, it will still become a landmark case study in the risks and power of AI-assisted formal mathematics.

What Alpöge says Claude did

Alpöge’s own note describes the work as a collaboration rather than a machine acting alone. He says he guided the approach from analogies with function-field proofs, especially the Stepanov tradition, and that Claude later outpaced him near the finish line . The mathematical sketch points to a “descendant of Stepanov’s” argument, with multiplicity over the integers supplied through chains of theta lifts between unitary groups .

The note is unusually candid about the limits of the human-readable explanation. Alpöge says the formalization is finished but that a prose version would be enormous and unreadable in its current form; he adds that he intends to understand and explain the argument properly . A footnote says an attempted written version ran beyond 250,000 pages . Quanta separately reported that the release consists of about 17 million lines of Lean code, which Alpöge says contains a formal verification of a stronger result, the generalized Riemann hypothesis .

The timeline included in the note runs from October 6 to November 16, 2026, ending with “GRH” . It lists ingredients that would normally belong to several different regions of modern number theory: theta lifts, Bessel periods, the Rallis seesaw, Iwasawa towers, the Bost slope inequality and relative trace formulas . That breadth is one reason the claim is so hard to digest quickly. It is not merely a short new trick that experts can inspect over coffee; it is, if correct, a huge synthesis.

What has already been checked

The strongest update since the announcement is that Andrew Sutherland of MIT posted a verification page on November 20, 2026, reporting that the published Lean source compiled without errors and that the five headline theorems depended only on Lean’s three standard axioms: propext, Classical.choice and Quot.sound . Sutherland also reported that all 175,187 modules compiled under Lean v4.33.1 and Mathlib v4.33.1, that every module was replayed through the Lean kernel, and that two independent kernels, nanoda and con-leche, accepted the full exported proof .

This is a major point in favor of the release, but it is not the same thing as a complete mathematical acceptance. Sutherland’s own caveats matter: a passing check shows that the formal statements, as written in Lean against Mathlib’s definitions, follow from the standard axioms according to the kernels used . In other words, the computer can say that a certain formal object is valid; mathematicians still need to confirm that the formal object exactly matches the intended theorem and that the surrounding definitions and dependencies behave as expected.

The theorem Sutherland identifies at the top of the submission uses Mathlib’s own definition of RiemannHypothesis, while four additional theorems include analogous statements for Dirichlet L-functions; on his page, “GRH” refers specifically to Dirichlet L-functions . That distinction is important because “generalized Riemann hypothesis” can be used in different scopes across number theory. The current claim is enormous even in that narrower Dirichlet sense, but careful language is essential.

The human review has barely begun

The first public reactions combine awe with restraint. AFP reported that the claim had not been independently verified and that the proof had not yet been reviewed by outside experts when its story was published . Brian Conrey, a leading specialist on the Riemann Hypothesis, told AFP that he had not yet seen a manuscript, only Alpöge’s remarks, while also calling the moment monumental if it holds up .

IBM Think quoted Fields Medalist James Maynard saying that if the result fully checks out, it would be a historic milestone in mathematics, while also noting that checking the Lean verification would take time . IBM also quoted Youness Lamzouri, who said the claim would be a huge achievement if true but warned that mathematicians need to understand the ideas well enough to explain, teach and build on them . That is the core tension: formal verification can establish correctness faster than the mathematical community can absorb meaning.

Quanta’s running update captured the same tension from another angle. It reported that Lean experts were racing to verify the code even though no one yet understood the many steps in the proof, and it quoted criticism that proof by Lean code can short-circuit the search for “why” by focusing on valid symbolic manipulation . That criticism does not invalidate formal proof. It does, however, explain why the next stage will be slower than the initial compile logs.

State of Proof, an independent tracking site for AI-assisted science, framed the open work as a bounded verification program: identify the Lean declarations for classical RH and Dirichlet GRH, compare their definitions and quantifiers with standard statements, inspect the dependency tree and replay the relevant declarations in isolation . Its intake record explicitly says inclusion is not validation and that the claim had not yet undergone a proof examination .

Why the Riemann Hypothesis matters

The classical Riemann Hypothesis concerns the zeros of the Riemann zeta function. In plain language, it predicts a precise hidden regularity behind the distribution of prime numbers. AFP summarized it as a problem first proposed by Bernhard Riemann in 1859 and widely regarded as the most important open question in pure mathematics .

The stakes are partly symbolic and partly technical. Symbolically, RH has served for generations as the icon of unsolved mathematics: simple enough to state in broad outline, deep enough to resist every known method. Technically, a proof would sharpen the error terms in prime-counting estimates and make many conditional results unconditional. Alpöge’s broader Dirichlet GRH claim would extend that power to primes in arithmetic progressions, a central theme in analytic number theory .

There is also the Clay Millennium Prize dimension. AFP noted that the Clay Mathematics Institute selected the Riemann Hypothesis as one of seven Millennium Prize Problems, each with a $1 million award, and that Clay’s rules require publication in a peer-reviewed journal and two years of scrutiny before a solution can qualify . Even a proof that compiles in Lean does not automatically trigger the prize process.

What this means for AI

If the claim survives, the AI story will be as big as the mathematical one. Alpöge’s note does not portray Claude as a passive autocomplete system. It credits Claude with key technical motion near the end of the argument, while Alpöge describes himself as steering the analogy and the overall path . The most plausible interpretation, pending fuller documentation, is neither “AI solved RH alone” nor “human used a better calculator,” but something stranger: a collaboration in which a human expert guided a model through a terrain too large for either party to summarize cleanly afterward.

That would mark a new stage for AI in research. Earlier AI-assisted mathematical results could be impressive yet bounded. This claim, by contrast, sits at the summit of mathematical prestige. It is not a benchmark puzzle; it is one of the problems that helped define 20th- and 21st-century number theory.

Still, the lesson should not be triumphalism. The decisive question is not whether Claude generated impressive mathematics. It clearly participated in producing an artifact that serious experts are now checking. The decisive question is whether that artifact proves exactly what it claims to prove, and whether mathematicians can translate it into durable understanding.

Where the story stands

As of the latest public checks within the first day, the situation is extraordinary but not final. Alpöge has claimed a Claude-assisted proof of RH and Dirichlet GRH, released as Lean rather than as a conventional paper . Sutherland’s MIT verification page reports successful compilation, kernel replay and independent-kernel acceptance for the headline formal proof artifacts, with some checks still in progress at the time of his update . Independent mathematical understanding, peer review and Clay-style acceptance remain ahead.

That leaves the correct headline exactly where it began: an Anthropic mathematician claims to have solved the Riemann Hypothesis with help from Claude. The claim is no longer just a press-cycle curiosity; it has machine-check evidence behind it. But until the formal statement, dependencies, axioms and mathematical ideas have been fully examined, the word “claims” remains essential.

Developments

  1. Anthropic Mathematician Claims to Solve Riemann Hypothesis With Help From Claude - The New York TimesThe New York Times · Nov 21, 2026, 12:23 AM · 9/10

Sources from the last 72 hours

  1. [1]remark.pdfNov 20, 2026, 1:00 AM
  2. [2]Verification files for the Lean proof of the Riemann HypothesisNov 20, 2026, 11:05 PM
  3. [3]Anthropic researcher says Claude solved math's greatest riddleNov 20, 2026, 8:32 PM
  4. [4]TransformationNov 20, 2026, 1:00 AM
  5. [5]Anthropic mathematician claims AI helped solve a 167-year-old math mysteryNov 20, 2026, 1:00 AM
  6. [6]A Claimed Proof of the Riemann HypothesisNov 20, 2026, 1:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.