An AI didn't solve the Riemann hypothesis. Here's what it did instead.

Anthropic asked Claude to "take a real stab" at the Riemann hypothesis. Claude gave up on the famous conjecture on its own, then turned to a different, named problem about how prime numbers are distributed, and produced a publishable new result there. The hypothesis itself, still open after nearly 170 years, is not what got advanced this week.

Prime numbers are the indivisible building blocks of arithmetic, the atoms every other integer is made from. The Riemann hypothesis, the most famous unsolved problem about how those atoms are spaced, would settle a long-standing puzzle inside the Riemann zeta function, the 19th-century object that encodes how primes are distributed. A $1 million prize for proving it has been on the table since 2000. Nobody has collected it. Anthropic's team asked Claude, an unreleased research version of the model, to work on the hypothesis. Claude walked away from the main question and produced work instead on a related, unresolved problem about the distribution of those primes, a piece of the prime-number machinery that sits next to the hypothesis without being equivalent to it. Proving anything precise about how primes thin out is hard, because the distribution is irregular enough to defy the standard analytical shortcuts, and the field has been waiting for the kind of new idea that does not lean on the existing playbook.

The outside read on the result comes from Maynard. He called the work "a genuinely interesting mathematical contribution" and said the related problem "was in need of a new real idea, which this new result seems to provide." That is not a vendor claim, and the name attached to it is the strongest one a result of this kind could land on. Maynard has done the kind of foundational work on prime gaps that the field treats as a benchmark.

Mathematicians have used AI as a tool, the way physicists use it, for pattern-hunting and exhaustive search, for years. The category of advance Anthropic is claiming, a language model generating a publishable new idea on a named open problem with an outside mathematician calling it genuinely new, is a different thing. That is the part of the story the "solved" and "did not solve" framings both skip.

The wave of "AI solves Riemann" coverage, and the matching "AI did nothing" rebuttal, both flatten the same thing: an LLM producing a publishable new idea on a real, named piece of prime-distribution math, with an outside mathematician saying so on the record, has not happened before. The hypothesis itself is exactly as open today as it was a week ago. The space of problems sitting next to it just got one new entry.

What happens next is peer review. Anthropic's paper is self-published, and the standard for a result of this kind is independent mathematicians, people who did not consult on the work, going through the argument line by line. Maynard's verdict is the strongest outside signal so far. The corrective frame from Scientific American's piece lines up with it: math's most famous problem is no closer to being solved, and AI made its first real advance on the problem's subject matter, not the conjecture itself.