CVCOVANewsComputing, Opportunities, Value & Analysis
Artificial Intelligence6 min read

Mathematician says he has proved the Riemann hypothesis with help from Claude; it still needs to be verified

Levent Alpöge has published a purported proof of one of the most famous problems in mathematics, produced together with Anthropic's AI. Early reviews are encouraging, but the community has not yet validated it.

By COVA News newsroom
Hand writing mathematical formulas on a blackboard with chalk

Photo: Vitaly Gariev on Unsplash

*Update (November 20, 2026): added the timeline of the discovery according to Alpöge's document, the specific idea he attributes to Claude, Elliot Glazer's estimate of a possible Lean bug, and the pending status of independent reviews.*

Since 1859, mathematicians have been trying to solve a puzzle about prime numbers, the numbers divisible only by 1 and themselves. On November 20, 2026, mathematician Levent Alpöge published a note of just four pages in which he says he believes he solved it together with Claude, Anthropic's artificial intelligence model. The proof is not a text that a person can read straight through. It consists of millions of lines of code written for a computer that checks each logical step. So far no one has definitively validated it, although the announcement has already reached outlets such as The New York Times, The Wall Street Journal and The Economist.

The problem is the Riemann hypothesis. In simple terms, it seeks to pin down as precisely as possible how many prime numbers there are of a given size. Primes appear to show up at random, but the hypothesis proposes that they follow a very precise order, tied to a mathematical tool called the zeta function. There is also a “generalized” version, which extends the same idea to a whole family of similar functions. The Clay Mathematics Institute lists it among its seven Millennium Prize Problems and offers US$1 million to whoever solves it.

There is a nuance about what exactly would have been proved. In his document, Alpöge says he has proved the classical version and maintains that his method extends to the generalized one. He even mentions that he hopes to reach an even broader variant, the so-called “extended” hypothesis. Scientific American, by contrast, emphasizes confirmation of the generalized version. Most outlets present Alpöge as an Anthropic mathematician, although Futurism links him to Harvard.

The document itself offers a brief account of how he got here. According to that timeline, the work with Claude began on October 6, 2026. Alpöge places the breakthrough on the night of Sunday, November 15, and says that same night he wrote to “Dario” to ask whether he could announce it. The timeline, generated by Claude, ends on November 16 with an entry titled “GRH,” the English abbreviation for the generalized Riemann hypothesis.

A proof checked by a machine

The proof is written in Lean, a “proof assistant”: a program that verifies that each step follows correctly from the previous one. The code was published as a compressed archive split into 14 parts. According to computer scientist Scott Aaronson, who wrote about the matter on his blog, the code runs to 17.4 million lines. In a footnote, Alpöge says that one version of the full text at one point exceeded 250,000 pages. In his document, he attributes to Claude one of the central ideas of the argument: using a chain of mathematical constructions called “theta lifts” whose length grows as the square root of X, the quantity used to measure the size of the numbers involved. It has not been specified which Claude model was used, nor how much of the work was done by the machine and how much by the person.

The early signs are positive. Scientific American reports that independent researchers, including Andrew Sutherland of MIT, ran the code without errors. Sutherland estimates the probability of a flaw at “approximately zero” and speaks of “a historic moment.” An independent GitHub repository called Hertaspace reviewed the project's 175,187 files. It did not find the instructions that typically hide shortcuts, that is, steps taken as true without being proved or that delegate verification to less reliable computations. That same repository, however, listed two reviews as still pending or merely planned: one of the axioms, which are the starting points accepted without proof, and one that checks that the statement is clean.

Even so, the code running is not enough. Mathematician Kevin Buzzard warned in the Lean community chat that a successful run of “comparator” is still missing. That tool confirms that no one changed the meaning of the symbols in the statement or exploited an internal Lean bug. Buzzard also wonders whether the formalized theorem is truly the Riemann hypothesis or a slightly different statement. Elliot Glazer, according to the outlet 36Kr, said on X that his preliminary inspection suggests it is indeed the generalized version, and estimated the probability of a bug in Lean itself to be well below 1%.

Why we have to wait

Aaronson is betting the proof will hold up. He notes, however, that Lean has had bugs that made it possible to “prove” that 0 equals 1, although he sees no sign of that happening here. Moreover, mathematicians will not rest easy until they can explain and reconstruct the ideas themselves, something that, according to him, “will probably take years.” Clay itself requires that the proof be published and withstand at least two years of scrutiny. According to Finchannel, as of November 20 the institute was still listing the problem as unsolved.

The announcement follows other results by Alpöge with Claude. In July he presented a counterexample to the Jacobian conjecture, and on August 19 another to the Carathéodory conjecture, although in those cases it is also unclear what each party did: Scientific American refers to a solution generated by Claude for the Jacobian conjecture, and Wikipedia says the AI “found a counterexample,” without clarifying what the human contribution was. That same August, Anthropic reported partial progress on this very hypothesis, achieved with a model it has not released and also verified in Lean. The result concerns the so-called nontrivial zeros of the zeta function, the key points of the problem: the hypothesis holds that all of them lie on the same line, the “critical line,” and the advance showed that at least a large share of them do. The report makes clear that this does not prove the hypothesis. The figures vary by source: a survey on arXiv describes going from around 42% to more than 66%, while Kingy.ai cites at least 67.25%. According to SiliconANGLE and science communicator Pasquale Pillitteri, the company had not issued an official statement on the new proof as of November 20.

If the proof holds up, it would be one of AI's greatest achievements in mathematics, and would show that a machine can contribute ideas to problems that have been open for more than a century and a half. What to watch now is the outcome of “comparator,” the review of the axioms, the check that the statement is the correct one, and the opinion of number theory specialists, who have not yet weighed in on the method. It also remains to be seen whether a human-readable version will appear and whether Anthropic will say anything officially. This is a theoretical advance, with no product behind it, and there is no information about concrete effects for Latin America. For now, the most honest approach is to view it with enthusiasm and patience.

Related news

Artificial Intelligence

OpenAI cuts GPT-6.2 Astra API pricing by 20%: what changes for those paying in dollars

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol. The flagship model's usage-based pricing drops, but the cheaper model's stays the same, and the fine print will determine how much users in Latin America actually save.

Artificial Intelligence

Nvidia reports earnings without Jensen Huang: what's at stake for AI chips

Five days after its founder's resignation, the company that dominates artificial intelligence chips faces Wall Street. For Latin America, open questions outweigh measurable effects for now.

Artificial Intelligence

Alibaba launches Qwen4-Max: a giant, cheaper model, still without independent testing to back it up

The Chinese company unveiled a frontier artificial intelligence model with lower prices and compatibility with OpenAI and Anthropic tools. For developers in Latin America it is another option, though questions remain about real-world performance, privacy and access from the region.