A proof checker has signed off on a claim about the most famous open problem in mathematics, and nobody can yet read the proof. If you're weighing what AI reasoning is worth, the gap between machine-verified and human-understood is the part to watch.
Levent Alpöge, a mathematician at Anthropic, posted a four-page note on November 20 saying he believes he has found, working with Claude, a proof of the Riemann hypothesis, according to Phys.org. The conjecture dates to 1859, concerns how the prime numbers are spread out, and carries a $1 million Clay Mathematics Institute prize. The next day, November 21, the Lean FRO and the Mathlib Initiative put out a joint statement saying they have "full confidence that Alpöge's proof of the Riemann Hypothesis is correct," as reported by Officechai and ZeroHedge.
That's two groups, and they're the ones to know about. Lean is the programming language in which the whole proof is written, so a computer can check it line by line, and the Lean FRO and Mathlib Initiative maintain the language and its main mathematics library. According to the reports, dozens of people compiled Alpöge's code and reproduced the same proof export. Two independent tools, called comparator and con-leche, confirmed that the theorem the library proves is exactly the Riemann hypothesis as Mathlib states it. The export was then run through several checkers, including Lean's own kernel and others written by third parties in different programming languages.
The size of the thing is hard to picture. ZeroHedge puts the code at about 17 million lines, released as 14 compressed archives rather than a paper. Alpöge said the AI-written informal explanation runs past 250,000 pages and isn't worth publishing, and he has promised a human-readable account later. He also said the method extends to the generalized Riemann hypothesis, and credited Claude with the step he expects "to make mathematicians fall out of their chairs."
MIT mathematician Andrew Sutherland compiled the code early and told MLQ's reporting that "the chance that anything is fishy here is approximately zero." The later Lean statement closed the two gaps that were still open at that point: whether the formal statement matches the real theorem, and whether any symbols had been quietly redefined to cheat.
How One Judge's Split Ruling on Anthropic Became AI's Copyright Rulebook Judge Alsup ruled that training AI on books is fair use, but that Anthropic's piracy of those books from shadow libraries was not, a split that cost the company $1.5 billion. Publishers are now using the same legal seam to sue Google over how Gemini was trained. - how judges ruled on AI training copyrighted books - why anthropic paid 1.5 billion for copyright infringement
Frankly, that settles the narrow question. If the statement is right and the kernel accepts the proof, the proof holds, and that's a much stronger guarantee than any referee report. It doesn't settle the Clay prize. Under the institute's rules a solution must be published in a qualifying outlet, wait at least two years and win general acceptance. Polymarket traders, per ZeroHedge, gave only 2% odds of Clay declaring the problem solved by the end of 2026.
Brian Conrey, one of the leading Riemann specialists, told Phys.org's source coverage he "really did not see this coming at all," and called the work seriously complicated. A proof nobody can hold in their head is a new kind of object for mathematics, and it'll take people, not checkers, to say what it means.
Don't mix this up with Fermat #
This isn't the first big Claude and Lean result, and the two are easy to confuse. On September 4, Anthropic published a Lean formalization of Fermat's Last Theorem that Claude agents wrote in about 11 days, around 13 million lines of code. That was a machine rewriting Andrew Wiles's existing proof into code, and Kevin Buzzard of Imperial College London, who has led a human formalization effort since 2024, called it an "extraordinary autoformalization achievement." Nothing new was proved. The Riemann claim is the reverse: if it stands, the mathematics itself is new.
There's also a precedent from August. Anthropic reported that an unreleased Claude had raised the proven lower bound for the share of zeta zeros obeying the Riemann hypothesis from 41.6% to 67.2%, with outside experts Brian Conrey and Dan Goldston reviewing it. The authors then wrote that they didn't expect those techniques to prove the hypothesis itself. Three months on, Alpöge says something did.
The timing, and a rival claim #
The announcement came one day after Anthropic's Nasdaq debut, and Alpöge addressed that head on. He wrote that "the precise timing is horrible," while insisting the result emerged on its own schedule and that leadership approved posting it, according to MLQ. If you hold the stock, treat the certification as evidence about capability, not about earnings. Nothing in the reporting ties the proof to the company's financials.
Hours after Alpöge posted, Stanford postdoc Alessandro Masullo published a human-written draft proof of the generalized Riemann hypothesis, labeled "all human, with no AI." It has no Lean formalization and hasn't been formally checked. Anthropic hasn't issued a separate statement confirming the proof, per the reports.
The human write-up Alpöge promised is the next thing that has to land.
Anthropic's $1.5 billion copyright settlement gets final approval but authors aren't cheering
Judge Araceli Martínez-Olguín granted final approval of the $1.5 billion Bartz v. Anthropic copyright settlement on July 20, 2026, the largest in U.S. history. Authors are split on the outcome: around $3,100 per work, with over $100 million going to lawyers, and Anthropic admitting nothing. The case sets a financial benchmark for AI copyright... - anthropic copyright settlement lawsuit authors - how much authors getting paid settlement
Also read: How to Build a Golden Dataset for AI Agent Evaluation From Your Logs • Etsy backs off AI training on seller work after shop owners walk out • How to write an AI agent memory retention policy before it bites
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.