Morning Edition · Saturday, November 21, 2026Published at 1:18 AM EST · New York
Early independent checks report that the Lean code compiles on the standard axioms and encodes the Generalized Riemann Hypothesis. On the same day, the American Mathematical Society's president posted a separate, human-written proof draft.

Levent Alpöge, a mathematician employed by Anthropic, posted a four-page note on November 20 saying he had, "working with Claude, discovered a proof of the Riemann hypothesis," according to Phys.org. Coverage of the note says the method extends to the Generalized Riemann Hypothesis (GRH). The note itself is not a readable proof. The argument exists in full only as a Lean formalization, and that is the version Alpöge chose to release. Terence Tao, the University of California, Los Angeles mathematician, wrote on his blog that Alpöge used an internal AI model to produce an "enormous" Lean-only proof of GRH. Tao pointed readers to the Lean Zulip discussion forum as the best place to follow verification.
A machine-written formal proof can fail in two ways. Either the code does not check, or it proves something other than the intended statement. Verification so far has addressed both possibilities. Andrew Sutherland of the Massachusetts Institute of Technology told Scientific American that the code compiles and rests only on "the three standard axioms." The mathematician Elliot Glazer reported, after a preliminary look, that the formalized proposition is indeed GRH. Reviewers still have to confirm that every definition the statement depends on matches the classical objects, and that check is not yet complete. The Clay Mathematics Institute still lists the Riemann hypothesis as unsolved. Coverage credits Claude with the key step, a long chain of theta lifts, but the note does not say which model or how much compute was used.
On the same day, Ravi Vakil, president of the American Mathematical Society (AMS), posted a "hurried announcement" of a proposed GRH proof by Alessandro Masullo. Vakil wrote that Brian Conrad of Stanford University had read the work for months and that it is "so far looking good." He had planned to bring Masullo to Stanford for a year to polish the write-up before any announcement. The draft is reportedly about 940 pages and under major revision. Two GRH claims surfaced within hours of each other, one in Lean and one in a traditional manuscript, and reviewers will check them by different methods on different timelines.
Early reporting contains errors. The Russian-language Telegram channel AI ML Big Data attributed the Claude-assisted claim to the mathematician Alexander Polishchuk. The note and English-language reporting identify Alpöge instead. Incentives matter here too. Alpöge works for a lab that listed on Nasdaq the day before. A confirmed result would be the strongest public evidence so far that a frontier model can contribute a decisive step on a Millennium Prize problem. The Lean kernel's check does not depend on Alpöge's reputation, which is why the formal release can be verified independently in a way a press statement alone cannot be.
Part of a tracked trend
AI Moves Into Autonomous Scientific Discovery and Clinical Care
Over the next 3-9 months, AI systems move beyond text tasks into running real scientific experiments and managing clinical care, backed by peer-reviewed and benchmarked evidence of chemist- and physician-level performance.
Start a discussion in Townsquare.
More from this edition
If the proof holds, Anthropic gains the strongest public evidence yet that its models can do decisive research, and that evidence reached the market the day after its Nasdaq listing, when investors were setting a value on its stated capability lead.
Scientific American and Nacionale confirm the code compiles and the statement appears to be GRH, but whether it matches the classical result is still listed as unconfirmed, and one report puts the code at about 17 million lines, so trust rests on the Lean kernel and the definitions because no person can read it, and readers should not confuse it with Anthropic's August result, which raised a bound on the share of zeros on the critical line and did not prove the hypothesis.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
Formal verification shifts the test of trust from the model itself to the Lean kernel and the definitions in the statement, giving labs with formal-math pipelines a type of proof benchmarks cannot supply. Lean community review will likely settle whether the statement matches GRH within weeks, unless a misdefined term surfaces and overturns the claim.
What to watch
Observations to monitor, not financial advice.
Synthesized from: Levent Alpöge (note) · Hacker News / Math 216 blog (Ravi Vakil) · Polylog editors · Scientific American · Terence Tao's blog
Comments
0No comments yet.