Anthropic's Levent Alpöge Claims a Claude-Assisted Riemann Proof, and Lean Checks Are In
A.I. / news
Anthropic's Levent Alpöge Claims a Claude-Assisted Riemann Proof, and Lean Checks Are In
Imperial's Kevin Buzzard and MIT's Andrew Sutherland have examined the formalization, according to Scientific American. No journal review has been reported.

Levent Alpöge, a mathematician at Anthropic, said on Friday, November 20, 2026 that he believes he has proved the Riemann Hypothesis while working with Claude. Independent mathematicians have since examined a Lean formalization of the proof, according to Scientific American, which updated its report on November 21.
Alpöge made the claim in a document hosted on his own site, alpo.ge/remark.pdf. The PDF could not be read as text when fetched for this report, so the quotes below come from the excerpt Simon Willison published the same day.
What Alpöge wrote
The opening line reads: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He also wrote that "Claude and I have also formalized the arguments without assumptions" and that "Claude far outpaced me near the finish line."
The excerpt adds that the method "easily extends much further, resolving the so-called Generalized Riemann Hypothesis". That second claim has no independent check in either source.
The hypothesis was posed by Bernhard Riemann in 1859. It concerns the pattern of the prime numbers.

Who has checked the Lean proof
Scientific American names two outside mathematicians and a joint statement from the Lean and Mathlib communities.
| Checker | Affiliation | What the report says |
|---|---|---|
| Andrew Sutherland | MIT | Compiled the Lean proof without issues |
| Kevin Buzzard | Imperial College London | Questioned whether the Lean statement matches the theorem |
| Lean and Mathlib communities | Joint statement | "Full confidence" the statement and proof are correct |
Sutherland said the Lean kernel rechecked the proof using only the three standard axioms. His words, as Scientific American reports them: "the chance that anything is fishy here is approximately zero."
Buzzard raised the question that matters for any formal proof. A file can compile and still prove a different statement from the one intended. He called for a comparator run, a check for hacks, symbols given new meanings and kernel-specific bugs that typically takes hours. The article said that run was pending, and its November 21 update says it has been completed.
The Lean and Mathlib communities went further than either individual. Their joint statement says they have "full confidence" in the proof and certifies that both the statement and the proof are correct, according to Scientific American.
Scientific American also reports that Alpöge first checked the work by writing the proof in Lean. Alpöge's own document says he and Claude formalized the arguments without assumptions.
What is still open
Scientific American calls the story developing. It does not mention peer review or a journal submission, and neither fetched page links a Lean repository.
The Generalized Riemann Hypothesis claim has not been tested by anyone named in the reporting. Buzzard's own wording on the comparator result was not available in the article text fetched for this report.
Anthropic has not been quoted on the claim in either source. For context on the company, see The Terminal's reports on Anthropic's IPO and on Anthropic and OpenAI cyber access tiers. The next checkpoint is a journal or arXiv submission, which neither source reports as filed.
Sources
More in A.I.
- 01Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 02Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 03Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.
- 04OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%The figures are from OpenAI's own red-team evaluation, and the post names DeepSeek and Moonshot AI as distillers of U.S. models.