On November 20, Eastern Time—the second day of Anthropic's (ANT.US) U.S. stock market debut—the company's top mathematician Levent Alpöge published a four-page note on his personal website, claiming that he and Claude had jointly found a proof of the Riemann Hypothesis, a problem that has vexed mathematicians for 167 years, and extended the conclusion to Dirichlet L-functions.
He publicly released Lean formalized source code split into 14 compressed volumes, totaling roughly 17 million lines. A separate manuscript, which he said exceeds 250,000 pages, was not made public due to its sheer length and unreadability. Anthropic declined to comment on the claim.
On the same day, as U.S. and global tech stocks sold off broadly on Friday, Anthropic closed at $88.91, down 8.50%, with its intraday low coming within just 5 cents of its $83 IPO price.
On one side is Claude's demonstrated capability in frontier scientific research; on the other is the question investors still need answered—whether this increasingly powerful general AI capability can be delivered repeatedly at scale, at pricing and subscription costs that customers are willing to bear.
This result has already moved past the stage of "the author claims the code runs." According to a verification log published by MIT number theorist Andrew Sutherland as of 06:05 Beijing time on November 21, all 175,187 modules passed item-by-item review by the Lean kernel, and two independently written verification kernels both accepted the fully exported proof.
The five core theorems rely only on Lean's three standard axioms, and independent cross-checks also confirmed they match the pre-listed target propositions. The classical Riemann Hypothesis directly adopts the existing definition in the public mathematics library Mathlib. This significantly strengthens the credibility of the formal proof.
What mathematicians need to explain next is which new ideas make the proof work, and which methods can be applied to other hard problems. The Clay Mathematics Institute still lists the Riemann Hypothesis as an unsolved problem; its million-dollar prize will only be considered after the result has been published in a qualifying channel for at least two years and has gained broad recognition from the international mathematics community.
According to the verification log released by MIT number theorist Andrew Sutherland, two independently written verification kernels both accepted the complete proof, and the formulation of the classical Riemann Hypothesis is consistent with the existing definition in the public mathematics library.
For investors focused on the AI supply chain, the most notable aspect of this historic advance—Alpöge and Claude finding a candidate proof for the Riemann Hypothesis that has puzzled mathematicians for 167 years—is that Claude can, under expert guidance, propose key derivations and deliver results that third parties can independently test.
How much research capability can support Anthropic's valuation further depends on whether this performance can be reproduced in other high-value R&D tasks, and whether the compute and expert hours required for each success can make customers pay while allowing the company to continue achieving strong profitability.
Anthropic researcher says he found a proof of the Riemann Hypothesis together with Claude
After Anthropic completed a record-breaking initial public offering, a researcher at the company said he believes he used the company's artificial intelligence software to solve a famous number theory problem that has troubled mathematicians for more than 150 years.
Mathematician Levent Alpöge posted a note on his personal website on Friday, saying he had found a proof of the Riemann Hypothesis. He shared links to code written in the Lean programming language and said the code provides evidence for his conclusion.
According to his note, Alpöge messaged Anthropic CEO Dario Amodei on Sunday night, asking whether he could announce the research results to the public. Alpöge described the timing of the announcement as "terrible," but said: "If I waited any longer, I really couldn't stand it."
Anthropic declined to comment, and Alpöge did not respond to requests for comment. The Riemann Hypothesis concerns the distribution of prime numbers and has long been regarded as one of the hardest problems in mathematics.
For years, OpenAI and Anthropic have sought to demonstrate the capabilities of their most advanced models by tackling difficult problems in mathematics and the life sciences. If Alpöge's result is confirmed, the Riemann Hypothesis would become the second Millennium Prize Problem potentially solved with the help of AI. In September, OpenAI said its model had produced a solution to the 90-year-old Navier–Stokes problem.
Andrew Sutherland, a senior research scientist at MIT, said he verified Alpöge's work on the Riemann Hypothesis on his own computer and also saw at least three verifications completed by other professional mathematicians on different systems. Sutherland said the research by Alpöge and Claude builds on a great deal of prior work, including generations of mathematicians' research on the Riemann Hypothesis as well as accumulated work in the field of mathematical formalization.
Looking ahead, Sutherland said: "I am both very nervous and very excited." He explained that mathematicians will have many years of work ahead, needing to study the proof in depth and figure out how it can be applied. "This is a day worth celebrating," Sutherland said. He also acknowledged that some in the mathematics community worry AI may disrupt mathematicians' careers and other jobs.
Noah Giansiracusa, an associate professor of mathematics at Bentley University and an affiliated scholar at Harvard University, added that if the solution holds, it would be a big deal for AI. "It sounds like this proof is indeed very likely to be airtight and completely valid," Giansiracusa said.
Alpöge described the experience of trying to prove the conjecture as "a stream-of-consciousness ramble." He wrote: "Over the past few months, Claude and I explored such a vast mathematical landscape together, and we enjoyed it immensely. But this time, nearing the end, Claude left me far behind."
From prime distribution to research AI pricing, what is truly scarce is verifiable discovery
The Riemann Hypothesis asserts that all nontrivial zeros of the Riemann zeta function lie on the line where the real part equals ½. It matters because these zeros determine how far the actual prime count can deviate from its smooth estimate. The Riemann zeta function is the Riemann zeta function.
In August this year, Anthropic reported that it had raised the lower bound on the proportion of zeros that can be proven to lie on this critical line from 41.6% to 67.2%. This time, Alpöge described a different route. He steered the research toward Stepanov's classical method in function fields, while Claude proposed stringing together about \(\sqrt X\) theta lifts for the von Mangoldt weighted sum with summation length \(X\), so that each lift simulates the role of one "multiplicity" in the classical proof. According to the author's account, Claude's most original contribution was proposing this step.
Terence Tao preliminarily judged that the argument may use modern results from the Langlands program and arithmetic geometry, and that it will still take time to clarify the full picture. Competition in research AI therefore now has a more concrete yardstick. Models can try many proof routes, Lean can verify the final derivation step by step, and researchers must judge whether the proposition is valuable and which lemmas are worth reusing.
For investors, a more explanatory metric is how much inference compute and expert time are required in total for each independently verified result that customers can use. This research has not yet disclosed the full cost of discovery. Open-source routes are also competing for the same group of research users.
Sergey Gukov, a professor at Caltech, and students trained specialized mathematical models from scratch; some models excel at specific problems but still have limited cross-domain ability. Terence Tao's Open Math OS integrates open-weight models, cloud computing, and verification tools, with research documents and conversations private by default and used for training only with researchers' explicit consent. For labs with unpublished results, whether a model can solve a problem, whether a proof can be independently reviewed, and who controls the research data will jointly determine procurement and renewals.