AI Model Claude Increases Known Bound for Riemann Hypothesis Zeros from 41.6% to 67.2%

Six hundred and fifty failed mathematical approaches were discarded in a day and a half. A swarm of 60 artificial intelligence subagents churned through 31 million tokens of output, writing hundreds of Python scripts and executing thousands of shell commands. At the end of this 36-hour brute-force sprint, an unreleased research version of Anthropic's Claude produced a paper that moved a stubborn mathematical wall. It pushed the proven lower bound of Riemann zeta function zeros on the "critical line" from 41.6% to 67.2%.[4][5]

The result represents the largest single-step advance in the 167-year history of the Riemann hypothesis. For 37 years, human mathematicians had managed to move this specific bound by less than a single percentage point. Claude increased it by over 25 points in a single weekend, acting as a universal reader capable of synthesizing decades of dense, specialized literature.[1]

But the data requires immediate qualification. Claude did not prove the Riemann hypothesis. The hypothesis, which carries a $1 million Clay Mathematics Institute prize, demands proof that 100% of the function's non-trivial zeros lie on the critical line. Proving that two-thirds of them do is a monumental achievement in analytic number theory, but Anthropic itself notes that the techniques used are unlikely to yield a full proof.[2][6]

To understand the evidence, one must first understand the mechanism. The Riemann zeta function is a mathematical tool that encodes the distribution of prime numbers. Primes appear unpredictable one by one, but their overall distribution follows a pattern. The "zeros" of the zeta function—the inputs that cause the function to output exactly zero—act as the fine print correcting our estimates of where primes appear.[2][3]

The Riemann hypothesis claims that all the interesting, "non-trivial" zeros fall perfectly along a single vertical axis in the complex plane, known as the critical line. If true, it means prime numbers are distributed as smoothly and predictably as mathematically possible.[3][5]

Computers have already calculated trillions of these zeros, and every single one has landed exactly on the critical line. However, finite computation cannot prove an infinite rule. There is always the possibility that some zero, infinitely far up the number line, breaks the pattern.[3][6]

Because mathematicians cannot check infinity, they instead prove asymptotic lower bounds—rigorous guarantees that, no matter how far out you go, at least a certain percentage of zeros will always sit on the line. Decades of human refinement, using a technique introduced in the 1970s, had painstakingly pushed this guaranteed floor to 41.66%, or exactly five-twelfths.[6][7]

Claude's breakthrough was not a flash of novel human-like intuition, but an act of massive literature synthesis. The AI model held disparate threads of specialized mathematics simultaneously, finding a connection that had eluded human specialists.[1]

Specifically, Claude found a way to stitch together techniques introduced by Hugh Montgomery in 1973—which previously only worked if one assumed the Riemann hypothesis was already true—with a 2000 paper by Enrico Bombieri and recent unconditional work by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh.[2][4]

By treating a certain space of functions with both positive and negative definite parts at once, Claude bypassed the need to assume the hypothesis. It swapped out a 53-year-old conditional assumption for a rigorous linear-algebra argument, pushing the floor to 67.2%.[2][7]

The process behind the proof is as significant as the math itself. The session was driven by Jared Sumner, an Anthropic staffer who is not a mathematician. Sumner prompted the model to attempt the hypothesis, and then essentially acted as a manager, repeatedly telling the system to "keep going" and "believe in yourself" when it hit roadblocks.[5]

The AI spun up a swarm of specialized subagents within a Claude Code environment. These agents wrote scripts, tested theories, and failed 650 times before finding the 651st approach that successfully connected the disparate papers.[5]

To ensure the result was not a hallucination, Claude was tasked with translating its argument into Lean, a rigorous programming language used for machine-checking mathematical proofs. The Lean formalization compiled successfully, mechanically verifying the logical steps of the new bound.[3][6]

Anthropic then brought in human oversight. Internal mathematicians Levent Alpöge and Ralph Furman reviewed the work, followed by external analytic number theory experts Brian Conrey and Dan Goldston, who examined the paper on short notice and validated the methodology.[2][3]

The evidence for the 67.2% bound is therefore unusually strong for a same-day AI announcement, backed by both machine formalization and domain-expert review. However, the exact AI model used remains an unreleased research version, meaning independent labs cannot yet reproduce the exact 36-hour generative run end-to-end.[4][6]

The achievement reshapes how the mathematical community views artificial intelligence. It suggests that AI's near-term role in advanced mathematics will not be replacing human creativity, but rather executing exhaustive, brute-force searches through the space of known literature to find connections humans missed.[1][8]