Anthropic's unreleased Claude research model pushed a key bound on the Riemann Hypothesis past prior human results, guided by encouragement alone.
Did Claude solve the Riemann Hypothesis? #
No. The Riemann Hypothesis remains unproven. What an unreleased Anthropic research version of Claude did was push forward a specific numerical bound tied to the problem, moving the zero-free region estimate from 41.6 to 67.2, a threshold that had stood as the best human result. That is not “67% of the proof.” It is progress on one narrow, technical piece of a much larger unsolved problem, achieved through a long, brute-force research process rather than a flash of insight.
TL;DR #
Claude did not prove the Riemann Hypothesis, but an unreleased Anthropic research model improved a specific zero-free region bound from 41.6 to 67.2, beating the best prior human-derived result.A non-mathematician at Anthropic, staffer Jared Sumner, drove the session largely by prompting Claude to attempt the hypothesis and then repeatedly telling it to keep going.The model spun up 60 Claude sub-agents, generated 31 million output tokens, ran about 2,400 shell commands, and wrote hundreds of Python scripts during the process.Most of the ideas failed: the sub-agents tried roughly 650 approaches, and none of them worked before the model found the one that did.** Encouragement mattered mechanically, not just emotionally**: messages like “keep going” and “believe in yourself” appear to have helped Claude push past early skepticism that it could make any real headway.This sits inside a broader wave of frontier models (including xAI’s Grok 4.5 in other recent cases) being used to find counterexamples and incremental results on open math conjectures.The result says more about process than proof: it shows how AI research agents can grind through massive search spaces on hard problems even when the human “prompting” them has no domain expertise.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
What is the Riemann Hypothesis, and why does this bound matter? #
The Riemann Hypothesis is one of math’s most famous unsolved problems. It concerns the distribution of prime numbers and, in essence, proposes that prime numbers are spread out with maximum regularity. Proving it would confirm that all the “non-trivial zeros” of the Riemann zeta function lie on a specific line. Mathematicians haven’t proven that, but they have proven partial results, including bounds on regions where zeros cannot exist (the “zero-free region”).
The number that moved, from 41.6 to 67.2, refers to one of these bounds. Improving it doesn’t get anyone closer to a full proof in a linear sense. Proving that all zeros lie on the critical line is a fundamentally different, much harder claim than tightening a region where they can’t lie. But it’s still a legitimate, citable improvement over the best previously known human result, and that’s notable regardless of who or what produced it.
How did a non-mathematician get Claude to make progress? #
According to Anthropic’s own account of the work, the person driving the session, staffer Jared Sumner, is not a mathematician. His role wasn’t to guide the math. It was to keep the model working. He prompted Claude to take a genuine attempt at the Riemann Hypothesis, then mostly sent messages of encouragement rather than technical direction. Phrases like “keep going” and “believe in yourself” showed up repeatedly.
Anthropic’s writeup credits this encouragement with helping Claude get past an early phase of skepticism about whether it could make meaningful progress at all. That’s a strange thing to read literally: a language model needing a pep talk to keep trying. But functionally, it lines up with something a lot of people who use these models heavily have noticed. Reasoning models can talk themselves into giving up on hard framing early, and a simple nudge to continue can reopen paths they’d otherwise abandon.
This matters for how people think about the “prompt engineer” role that got hyped a few years ago. It turns out that in at least this case, the highest-value skill wasn’t deep subject knowledge. It was persistence and the willingness to keep pushing a model that seemed stuck.
What actually happened during the research process? #
The scale of the effort is the part worth sitting with. Anthropic’s account describes an unreleased research version of Claude that:
- Produced 31 million output tokens over the course of the work.
- Spun up 60 separate Claude sub-agents to parallelize different lines of attack.
- Ran roughly 2,400 shell commands and wrote hundreds of Python scripts.
- Had its sub-agents run thousands of numerical checks to test hypotheses.
- Generated and tested about 650 distinct ideas, the overwhelming majority of which failed.
That’s not one model having an insight. That’s an enormous, largely automated search process, with a swarm of sub-agents trying numerical experiments, checking results, discarding failures, and iterating, all orchestrated with minimal human mathematical input. The eventual improvement in the bound came out of that grind, not out of a single elegant proof step suggested by a person.
#
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Is this part of a bigger trend in AI and mathematics? #
Yes. This isn’t an isolated event. Frontier models have been increasingly used to find counterexamples to conjectures and make incremental progress on open problems in math, sometimes within days of each other. Separately from the Anthropic work, other researchers have used xAI’s Grok 4.5, run inside coding tools like Cursor, to find a counterexample disproving a different mathematical conjecture. The frequency of these results has picked up enough that keeping track of which model produced which result has become genuinely confusing, since new claims surface on a near-weekly basis.
The pattern across these cases is similar: a capable reasoning model, given enough compute, time, and a research-style workflow (sub-agents, scripts, numerical verification loops), can chip away at problems that are too vast for a single mathematician to brute-force by hand, even without being “right” about the big unsolved question itself.
Is this a meaningful AI capability milestone or just a headline? #
It’s genuinely useful evidence, but it needs a fair frame. The model didn’t demonstrate novel mathematical insight in the way a human breakthrough proof would. It demonstrated an ability to sustain a massive, structured search across thousands of numerical checks and hundreds of failed ideas until one line of attack produced a result that beat the prior best bound. That’s valuable. It’s also a very different kind of achievement than “solving” a Millennium Prize-caliber problem.
What makes it worth paying attention to is less the specific number and more the workflow: an unreleased research model, a non-expert human simply keeping it motivated, dozens of sub-agents working in parallel, and a brute-force numerical search that ground through hundreds of dead ends before finding something that worked. That’s a plausible template for how AI-assisted math research looks in practice right now: less “AI discovers elegant proof,” more “AI runs an enormous, patient search that a human keeps alive.”
Frequently Asked Questions #
Did Claude prove the Riemann Hypothesis?
No. It improved a specific numerical bound (the zero-free region estimate, from 41.6 to 67.2) beyond the best previous human result. The core hypothesis, that all non-trivial zeros lie on the critical line, remains unproven.
Who prompted Claude during this research, and were they a mathematician?
Jared Sumner, an Anthropic staff member, drove the session. He is not a mathematician. His main contribution was prompting Claude to attempt the problem and then repeatedly encouraging it to keep going.
How many ideas did Claude try before finding one that worked?
Anthropic’s account describes roughly 650 generated and tested ideas across the research process, the vast majority of which did not work.
What tools did Claude use during the process?
The model used 60 Claude sub-agents, ran about 2,400 shell commands, wrote hundreds of Python scripts, and had its sub-agents perform thousands of numerical checks, all totaling 31 million output tokens.
Is this Claude model publicly available?
No. The version used for this work is described as an unreleased research model, not a publicly shipped Claude product.