Tech • AI • Robotics • Game

ENFR

Daily Podcast full article

Anthropic takes on Riemann hypothesis: Claude’s claimed proof meets the peer-review test

Anthropic mathematician Levent Alpöge says Claude helped him produce a proof of the Riemann hypothesis, one of mathematics’ most famous open problems. The claim is extraordinary, but not yet accepted: a Lean formalization, a rival human draft, and early expert scrutiny have shifted the question from “can AI score on benchmarks?” to “can mathematicians verify and understand what it produced?”

Generated November 21, 2026 at 6:19 AM≈ 6 min read
AI-generated illustration

A claim big enough to stop the room

The working headline is the story: Anthropic takes on Riemann hypothesis. On November 20, Levent Alpöge, a mathematician at Anthropic, posted a four-page note saying he believes that he and Claude have discovered a proof of the Riemann hypothesis . The claim goes further than the classical statement: Alpöge says the method “easily extends” to the Generalized Riemann Hypothesis, and that he and Claude have formalized the arguments without assumptions .

That is why the announcement immediately landed beyond the usual AI hype cycle. The Riemann hypothesis, formulated in 1859, is not just another hard exercise: it concerns the zeros of the zeta function and the fine structure of prime numbers, and it remains one of the Clay Mathematics Institute’s Millennium Prize Problems, with a $1 million prize attached . AP reported hours after the post that the Clay site still listed the problem as unsolved and quoted Clay president Martin Bridson warning that the path from announcement to real understanding would be “long and difficult” .

The most important word in Alpöge’s statement is therefore not “proof” but “believe.” The public status, as of the first day of scrutiny, is a claimed proof, not an accepted theorem.

What Alpöge says Claude did

Alpöge’s own note is unusually personal for a mathematical announcement. He writes that Riemann had been a guiding ambition since his student days and that, near the finish line, Claude “far outpaced” him . The mathematical sketch points to a descendant of Stepanov-style arguments and to a mechanism involving chained theta lifts, with Claude proposing a chain length tied to a von Mangoldt sum .

For non-specialists, the key point is simpler: Alpöge is not describing Claude as a calculator that filled in routine algebra. He is describing a collaboration in which he set directions and analogies, while the model allegedly supplied crucial mathematical moves deep inside the argument . If that account survives review, it would mark a stronger role for generative AI than “assistant,” closer to “research collaborator” in foundational mathematics.

At the same time, the note makes clear that the human-readable proof is not yet in a normal publishable form. Alpöge says the argument is “incredibly long,” that producing a huge unreadable document would be routine, and that his next task is to understand, simplify and explain the proof properly . AP reported that Alpöge referred to more than 250,000 pages emerging from the work with AI . SiliconANGLE similarly reported that no full paper had been released, while a Lean version had been posted instead .

Why Lean matters, and why it is not the end

The unusual feature of this claim is the formalized proof archive. Alpöge’s note links the proof in fourteen compressed Lean archive parts . Lean is a proof assistant: if the theorem is stated correctly, the axioms are standard, and the archive builds cleanly, the software can check the logical steps at a level no human referee would attempt line by line.

That is a major reason mathematicians are treating the claim differently from the many failed Riemann proof announcements of the past. Lean4.dev, a site focused on the Lean ecosystem, described the release as a claimed proof of the Riemann hypothesis and of GRH for Dirichlet L-functions, delivered as a Lean archive rather than a paper . Its early inspection found a vast archive: 175,187 Lean files, 17.4 million lines, and about 819,000 theorems and lemmas, pinned to Lean 4.33.1 and a fixed Mathlib commit .

The same analysis also states the caveat clearly: early builds and checks of the formal statement had been reported, but there was not yet expert review of the mathematics, and the page did not assert that the hypothesis had been proved or that Clay had accepted a solution . A clean Lean build can tell the community that a formal object checks against a kernel; it cannot, by itself, tell mathematicians whether the definitions capture exactly the intended theorem, whether hidden assumptions have slipped in, or how to turn the result into comprehensible mathematics.

That distinction is now the center of the story. The decisive test is no longer a benchmark score or a demo. It is the slower machinery of formal verification, independent rebuilding, statement auditing, and peer review.

The human rival: Masullo’s draft

The plot became even stranger because another proof claim surfaced the same day. Stanford mathematician Ravi Vakil posted a hurried announcement that Alessandro Masullo had a proposed proof of the Generalized Riemann Hypothesis, that Brian Conrad had been examining it closely for months and had given extensive feedback, and that the work was “all human, with no AI” . Vakil said the plan had been to bring Masullo to Stanford for a year to polish the writing, but that the current draft should now be made public for checking, formalization and gap-finding .

This creates a rare and uncomfortable priority moment: an AI-assisted formal archive and a human-written draft appeared in public within hours of each other. But the two artifacts are not directly comparable. Alpöge’s release is machine-checkable in principle but largely unreadable as mathematics at its current scale. Masullo’s draft is meant for humans but, by Vakil’s own framing, is still a draft needing outside checks and polish .

The fair conclusion is not “Claude solved it” or “the human proof wins.” The fair conclusion is that two extraordinary claims entered the verification pipeline at once.

What would acceptance require?

For Clay-level acceptance, a Riemann proof must become more than a viral PDF or a massive codebase. It must be published, understood, and accepted by the mathematical community over time. AP’s report that the Clay site still listed the problem as unsolved on November 20 is not bureaucratic foot-dragging; it is the normal posture toward a claim of this magnitude .

The Lean archive may accelerate part of the process. If independent Lean experts can build it reproducibly, confirm that only standard axioms are used, inspect the theorem statement, and replay the result in independent tooling, that would provide formidable evidence . But human understanding still matters. Mathematics is not only the binary fact of validity; it is the map of why a result is true, which ideas generalize, and what new structures have been exposed.

Alpöge himself seems to recognize this. His note says he intends to speak about the argument and write it properly . That future write-up may be as important culturally as the proof archive itself, because it will determine whether the result becomes part of mathematics rather than an opaque artifact accepted by a checker.

The AI milestone, if it holds

If the proof is valid, the consequences for AI research are hard to overstate. The Riemann hypothesis has functioned for generations as a symbol of mathematical depth. A Claude-assisted solution would suggest that frontier models can contribute not merely to coding, tutoring, or theorem-proving benchmarks, but to the creation of new research at the highest level.

Still, the most sober lesson is that AI does not remove verification. It moves verification to the center. Claude may have “flown,” in Alpöge’s memorable phrasing, but the landing still has to be inspected by mathematicians, proof engineers and institutions . Until then, this is a historic claim under review, not a closed chapter.

In other words: even Deep Thought would need more than 42 citations. The answer is not the announcement. The answer is what survives the proof check, the expert reading, and the long aftershock of understanding.

Comments

Be the first to comment.

Sources from the last 72 hours

  1. [1]RemarkNov 20, 2026, 1:00 AM
  2. [2]The Riemann Hypothesis: a claimed proof, delivered in LeanNov 20, 2026, 1:00 AM
  3. [3]Hurried Announcement: Alessandro Masullo GRHNov 20, 2026, 1:00 AM
  4. [4]Anthropic researcher claims to have proven Riemann hypothesis using ClaudeNov 20, 2026, 10:28 PM
  5. [5]Anthropic worker says AI helped him solve Holy Grail of math problemsNov 20, 2026, 1:00 AM

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