crypto

The AI That Cracked Maths' Biggest Prize. Maybe.

The Riemann hypothesis has sat unsolved for 167 years, resisting every brilliant mind thrown at it since Bernhard Riemann first posed the question in 1859....

21 November 2026
The AI That Cracked Maths' Biggest Prize. Maybe.

The Riemann hypothesis has sat unsolved for 167 years, resisting every brilliant mind thrown at it since Bernhard Riemann first posed the question in 1859. Last Friday, one day after Anthropic floated on the Nasdaq and raised $81.34 billion, one of the company's own mathematicians announced that he and an AI called Claude had cracked it. If true, it is the most significant mathematical result in living memory. If not, it is the most spectacular piece of accidental IPO timing in the history of tech. Either way, we need to talk about it properly.

What Actually Got Claimed Here?

Levent Alpöge, a mathematician at Anthropic, published a four-page note last Friday claiming that he and Claude had produced a proof of the Riemann hypothesis. That alone would be enormous. But Alpöge went further: he says the method also resolves the Generalized Riemann Hypothesis, the broader version of the problem that extends the original question to other mathematical functions beyond the Riemann zeta function. That is not a footnote. That is a significantly larger claim than even the headline suggests.

To understand why this matters, you need a rough sense of what the Riemann hypothesis actually is. Riemann's original conjecture from 1859 concerns where certain zeros of the zeta function fall on the complex plane. The reason mathematicians care so obsessively is that those zeros are directly linked to how prime numbers distribute themselves. Primes are not randomly scattered. There is a pattern, and that pattern is tied to those zeros. Centuries of mathematical research rest, at least partially, on this foundation. The Clay Mathematics Institute rates it as one of its seven Millennium Prize Problems, each worth a million dollars, and no problem on that list commands quite the same mystique as the Riemann hypothesis.

What Does "Lean Proof" Actually Mean?

Here is where it gets both more interesting and more complicated. Alpöge and Claude did not just hand-wave a solution. They encoded their argument in Lean, a specialised piece of software designed to check mathematical proofs step by step, verifying that each formal deduction follows the rules of logic. That sounds definitive, but it is not the same as mathematical acceptance. Not even close.

Alpöge's note links to 14 downloadable parts of the proof archive, which is an unusually generous level of transparency. The maths community can go digging. But what Lean can do is confirm that the formal deductions obey its own rules. What still needs to happen is that mathematicians reproduce the full archive themselves, verify that the formalised theorem actually maps correctly onto the original conjecture as Riemann stated it, and examine whether the logical chain holds up under scrutiny from people whose entire careers have been spent in this space. Lean is a tool. It is not a verdict.

The Timing That Nobody Can Ignore

Alpöge himself used the word "horrible" to describe the timing. That is quite an admission. He says he messaged Anthropic CEO Dario Amodei on the Sunday night before the Friday publication, specifically the evening of 15 November, to raise the possibility of an announcement. He then says he consulted company leadership, acknowledged there was "a constraint to respect" without spelling out what that constraint was, and published on Friday, the day after Anthropic began trading on the Nasdaq. Anthropic's IPO raised $81.34 billion. The company's stock is now publicly traded. And the following morning, one of its mathematicians announced a potential solution to the most famous unsolved problem in mathematics with a million-dollar prize attached.

Alpöge says this should not be read cynically. He stopped short of saying the IPO caused the delay. But the gap between messaging the CEO on Sunday and publishing on the Friday after the IPO is not nothing. It sits there, in his own words, unexplained. We are not accusing anyone of anything. We are saying that if you wanted to design a scenario that looked suspicious even if it was completely innocent, this is roughly what it would look like.

This Is Not Happening In a Vacuum

Alpöge is not new to the chaotic intersection of AI and high-stakes mathematics. Back in September, OpenAI claimed to have solved the Navier-Stokes Millennium Prize problem while Alpöge and NYU mathematician Tristan Buckmaster were pursuing related work on fluid equations. A credit dispute broke out over research priority. Then in October, OpenAI released 722 manuscripts generated by an unreleased model, grouped into 372 families of results, of which only 162 included Lean-checked main results. The mathematical community was not impressed. Demanding solid evidence rather than a flood of AI-generated claims is not obstructionism. It is how mathematics actually works.

The anxiety in that community is real and it is growing. Toronto mathematician Daniel Litt has warned that without serious adaptation, there might be "no more math in 50 years" if AI simply hoovers up the unsolved problems before human researchers can learn anything from the pursuit. At an MIT panel held this week, Buckmaster argued that mathematics still fundamentally requires human understanding. These are not luddite positions. They are serious questions about what it means to do mathematics at all when a machine can generate thousands of pages of formally-checked argument overnight.

What Would It Take to Actually Win the Prize?

Here is where the Clay Mathematics Institute's rules deserve a proper look. Winning the million dollars is not simply a matter of publishing a proof and waiting. Under the institute's prize rules, a proposed solution must be published in a qualifying outlet, survive at least two years of scrutiny from the mathematical community, and earn general acceptance before the prize can even be considered. Two years minimum. That is baked into the process because history is littered with claimed proofs of famous problems that collapsed under sustained examination.

Anthropic itself had previously reported progress on the zeta function, which is the mathematical object at the heart of the Riemann hypothesis. But that earlier work stopped well short of claiming a solution to the full problem. Alpöge's new assertion goes significantly beyond it. The honest answer right now is that we do not know whether this holds up. Independent verification is still pending. The archive is available. The community is presumably already looking. The next months will be considerably more interesting than the announcement itself.

Our Verdict

The cynical read is that this is an AI company, freshly floated on the stock market at a valuation that would make your eyes water, finding a very convenient way to generate headlines about its flagship product. The generous read is that a genuinely brilliant mathematician has been working on something extraordinary, found himself in an awkward position regarding timing, made a judgement call, and is now waiting for the community to do what it always does, which is check the work. The truth is probably somewhere in the middle, which is to say: complicated.

What we can say with confidence is this. If the proof stands up, the story of how we tell it will get harder. Claude will be in the first sentence of every account ever written about the Riemann hypothesis. That is worth $81.34 billion and then some. And if the proof falls apart, we will have learned something important about the difference between AI-assisted mathematics and AI-validated mathematics, a distinction the industry has not yet been forced to draw clearly. Either outcome matters. Watch this space, but do not hold your breath. The Clay Institute rules say we are waiting at least two years regardless. Maths does not care about IPO cycles.

Share:XFacebookWhatsApp