On November 20, 2026, Levent Alpöge — a number theorist on Anthropic’s technical staff — announced a proof of the Riemann hypothesis found in collaboration with Anthropic’s Claude model. The release exists only as a Lean formalization: 175,187 files, roughly 17 million lines, with no human-readable paper. The claim covers the Generalized Riemann Hypothesis for Dirichlet L-functions as well, with the Extended Riemann Hypothesis expected to follow with more time.1 Within a day, the Lean and Mathlib communities had machine-certified both the formalized statement and the proof, and independent checks at MIT and Waterloo reproduced the verification. What remains open is not machine correctness but human understanding.
The announcement
The claim arrived as a four-page untitled note on Alpöge’s website (alpo.ge/remark.pdf). Its core sentence: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” No paper accompanies it — Alpöge’s own draft write-up passed 250,000 pages, which he judged it “obviously inappropriate” to post as a “massive unreadable travesty of ‘slop’”, so only the Lean code was released, in fourteen archive parts.2
Alpöge describes the collaboration frankly: he set the direction by combining two ideas from a toy setting — an argument descended from Stepanov’s function-field proof, where the multiplicity in Stepanov’s divisor is hard to reproduce over the integers — and Claude supplied steps beyond his own expertise, including the suggested length of the chain of theta lifts for the von Mangoldt sum of length X, namely √X. He credits Claude with the final steps (“Claude far outpaced me near the finish line… Claude flew, and danced, and sparkled near that finish line”) and states plainly that he is presenting a complete argument before having mastered the mathematics around it. A timeline compiled by Claude runs from October 6 to November 16, 2026.3
The method, as far as the note describes it, transplants Stepanov-style determinant arguments into an automorphic setting: iterated theta lifts between totally definite unitary groups win the multiplicity that Stepanov’s argument needs, with norms and divisibility controlled via Chowla–Selberg, Weyl, and Siegel mass formulas. Michael Rubinstein’s working guide (a 164-page draft, itself written with Claude, Codex, and Gemini as study aids) reports that the proof deduces the Riemann hypothesis from one inequality between weighted sums over primes — labeled (M) — plus two estimates for Dirichlet characters, with (M)‘s proof taken from the Lean development.4
Verification
On November 21 the Lean and Mathlib communities issued a joint statement, endorsed by the Lean FRO and the Mathlib Initiative, certifying that “both the statement and the proof of the Riemann Hypothesis are correct” against four criteria: the Mathlib RiemannHypothesis statement is the intended one and expert-vetted; the library compiles (dozens of independent compilations reproduced the same raw proof export); the code proves exactly that statement (two independent comparison tools, comparator and con-leche –compare, confirmed the match); and the exported proof is valid — checked by Lean’s kernel, by the con-leche and con-ron checkers (which carry formal proofs, relative to set theory, that they cannot accept a false proof), and by third-party checkers in other languages. The checks ran on an inert export of the proof, out of reach of any executable code in the Lean sources.5
Independent replications followed the same day. Michael Rubinstein (Waterloo) reports four checks with no errors: full compilation of all 175,187 modules; an axiom audit (the five headline theorems depend only on Lean’s three standard axioms); a replay of every module through Lean’s kernel; and a check of all 1,484,729 dependent declarations by nanoda, an independently written kernel. Andrew Sutherland (MIT) ran the same class of audit, concluding the theorems depend only on the standard axioms; further independent receipts are tracked on a community dashboard (Naskręcki’s riemann-rev).6 Sutherland’s outside assessment was blunt: “Honestly the chance that anything is fishy here is approximately zero.” James Maynard put his confidence at “98 or 99 percent”. Terence Tao wrote that understanding the proof will take time and that early analysis suggests it uses the full power of modern number theory.7
Verified but not understood
The event splits Tao’s five-stage problem-solving pipeline cleanly: generation and verification are now machine-complete, while exposition, community acceptance, and canonicalization — the human stages — have not begun. Alpöge says he intends to understand and simplify the argument and write it properly; his note itself is a declaration of the gap (“I may have baked lots of strange suboptimal choices into the argument”). This is the Leiden Declaration’s “informal communication” threat in its most extreme form: a personal website plus a community blog post, with peer review replaced by machine verification. It also inverts the October 2026 OpenAI bulk-release failure mode: there the model was unreleased and the artifacts unverifiable, while here the artifact is public and independently checkable without trusting anyone’s model. The tension is real enough that Brian Conrad (Stanford) called releasing a proof its own author does not fully understand “frankly completely outrageous”. The Clay Mathematics Institute continued to list the Riemann hypothesis as unsolved in the days after the announcement.8
Tao’s own stated rule — that a result whose authors cannot give a clear, expert-level talk should not be published — is in direct conflict with this release, and that conflict is now the live question in ai-mathematical-practice: whether machine verification plus eventual human assimilation is an acceptable substitute for the understanding-first sequence the field’s norms assume.
Timing and reception
The announcement came one day after Anthropic’s record IPO, and Anthropic’s shares fell on the Friday it was made as an AI sell-off spread. Alpöge pre-answered the cynical reading in the note itself: the breakthrough arrived late on a Sunday night, he asked CEO Dario Amodei about announcing, “there was a constraint to respect”, and “there is an obvious cynical view, to do with the timing, but it is false”.9 The same day, AMS president Ravi Vakil posted a separate all-human proposed proof by Alessandro Masullo; within about eight hours he reported that several people had found problems in it — with AI models.10 Alpöge had earlier made headlines in July 2026 with a Claude-found counterexample to the Jacobian conjecture, and on the announcement day he and Tristan Buckmaster — the same analyst who had disputed OpenAI’s Navier–Stokes claim on priority grounds — claimed that the Navier–Stokes equations can blow up unforced.11
Open questions
- Whether the Clay Millennium Prize process will recognize the result, given its rules and the absence of a conventional publication — the prize requires two years after publication of a qualifying solution.
- Whether Alpöge’s expected “massive simplification” arrives and produces a proof humans can actually digest; Rubinstein’s 164-page working guide is the first step in that direction.
- What it means for a field to accept a proof whose statement and validity are machine-certain while the ideas remain unassimilated — the concrete instance of the assimilation-funding question AGMAI raised in September.
- Whether the theta-lift/Stepanov combination transfers to other L-function settings (the Extended Riemann Hypothesis is Alpöge’s own next target).
Related
- ai-mathematical-practice — the institutional framework this event stress-tests: Tao’s pipeline, the Leiden Declaration, AGMAI, and the OpenAI bulk release as the counter-case
- satisfiability-modulo-theories — the formal-verification substrate that makes machine certainty possible here
- anthropic-ipo — the market context of the announcement’s timing
Sources
- Levent Alpöge 2026 — Riemann Hypothesis announcement note
- 2026 — Lean verification of the Riemann Hypothesis
- Michael Rubinstein 2026 — Alpöge/Claude proof of the Generalized Riemann Hypothesis
- 2026 — Anthropic mathematician says he and Claude proved the Riemann hypothesis
Footnotes
-
Michael Rubinstein 2026 — Alpöge/Claude proof of the Generalized Riemann Hypothesis ↩
-
Michael Rubinstein 2026 — Alpöge/Claude proof of the Generalized Riemann Hypothesis ↩
-
2026 — Anthropic mathematician says he and Claude proved the Riemann hypothesis ↩
-
2026 — Anthropic mathematician says he and Claude proved the Riemann hypothesis ↩
-
2026 — Anthropic mathematician says he and Claude proved the Riemann hypothesis ↩
-
2026 — Anthropic mathematician says he and Claude proved the Riemann hypothesis ↩