Claude raised the Riemann zeta zero bound from 41.6% to 67.2%: inside Anthropic's August 10, 2026 result

Anthropic announced on August 10, 2026 that an unreleased research version of Claude improved the lower bound on the fraction of Riemann zeta function zeros satisfying the Riemann hypothesis, raising it from 41.6% to 67.2%. Claude reached the result over two sessions in Claude Code using a total of 31 million output tokens, and two Anthropic mathematicians, Levent Alpöge and Ralph Furman, validated the paper while outside experts Brian Conrey and Dan Goldston examined it. The result is formalized in Lean and passes the standard validation tool comparator. ASAP works from Anthropic's own announcement and the accompanying proof note to separate what was proved from what was not.

What moved is the fraction, not the hypothesis unsolved since 1859

The Riemann hypothesis has stood unresolved since 1859, and what changed on August 10, 2026 is not its truth value but the lower bound on how many of the zeta function's zeros are known to satisfy it. The hypothesis asserts that every nontrivial zero of the Riemann zeta function lies on the critical line, and rather than prove that outright, mathematicians have spent decades pushing up the fraction of zeros that can be shown to sit there. Anthropic describes that accumulated constant as having reached 41.6% before this work.

Claude's theorem raises the figure to 67.2%. The constant written in the accompanying proof note is 3/2 minus the cotangent of one over the square root of two, which evaluates to 67.25%, and the note opens by summarizing the result in a single line: 67% of the zeroes are on the line. The share of zeros that could still sit off the critical line has been squeezed below one third.

The foundations are stated explicitly. Anthropic says Claude's approach starts from recent work by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh, and combines it with Bombieri's 2000 paper and techniques from Montgomery's 1973 work. This is one step placed on an existing lineage of research rather than a result produced from nothing, and the announcement says so plainly.

650 failed ideas and a day and a half of roughly 60 subagents

The process numbers weigh far more heavily toward failure than success. Claude generated and tried 650 ideas in the initial phase, and by Anthropic's account none of them worked. The turn came afterward: in the follow-up phase, roughly 60 Claude subagents coordinated over about a day and a half, executing 2,400 shell commands and hundreds of Python scripts.

Verification ran in parallel with the search. Claude checked its work numerically against known zeta zeros, searched arXiv and downloaded 54 papers for comparison, hunted for counterexamples, and produced independent re-proofs. The two sessions consumed 31 million output tokens in total.

These figures describe a cost structure for search rather than an elegant proof. Continuing after 650 ideas have all failed is not a condition a single researcher's calendar accommodates, and consuming failure in bulk worked here. The same structure also makes clear that none of the waste is identifiable as waste until the result arrives, since the overwhelming majority of those 31 million tokens left no line in the final proof.

The human input amounted to "keep going"

Human involvement in this case is recorded at an unusually low level. Jarred Sumner, an Anthropic staff member and non-mathematician, prompted Claude to take a real stab at the hypothesis itself and left the mathematical choices from there to the model. Anthropic writes that his subsequent input was mostly limited to messages of encouragement, mostly variants of "keep going" or "believe in yourself."

Mathematical expertise entered at review rather than at generation. Two mathematicians at Anthropic studied and validated Claude's paper, and Conrey and Goldston, both experts in the area, examined it on short notice. The formalization was produced with another staff member, Eric Easley, and the resulting Lean proof passes the standard validation tool.

Read the division of labor literally and it looks like this: a non-expert chose the problem and supplied encouragement, the model ran the search and wrote the proof, and experts plus a machine handled verification. That arrangement reopens the question of which part of mathematical research was actually the part a human had to hold.

Why the verification stack matters more than the result

The bottleneck for AI-produced mathematics is verification rather than generation, and Anthropic's August 10, 2026 release is built around that fact. A language model can assemble a plausible-sounding chain of argument at will, and an expert needs far longer to read one paper properly than a model needs to write one. Once generation outruns review, the credibility of any result rests on verification throughput rather than on generative capability.

That is where this announcement separates itself from earlier claims. Anthropic attached three layers at once. The human layer holds two in-house mathematicians and two outside experts, the machine layer holds a Lean formalization that passes the standard tool, and the model's own layer holds numerical checks against known zeros, counterexample searches, and independent re-proofs. The formalization matters most here, because Lean judges logical completeness rather than persuasiveness.

The stack still has a seam worth naming. A passing Lean formalization establishes that the formalized statement follows from the formalized hypotheses, and it does not by itself guarantee that the formalized statement is the same claim the announcement makes in prose. Checking that correspondence remains human work, and publishing the Lean repository alongside the proof note is what opens that check to outsiders.

Two ways to read 25.6 percentage points

By the numbers, this jump is larger than the field's usual increments. The bound moved from 41.6% to 67.2%, a gain of 25.6 percentage points that adds more than half again to a constant built up over decades. In an area where improvements measured in fractions of a percent constitute a paper, a move of this size is conspicuous on its own.

Overreading begins the moment the object of that bound is dropped. The 67.2% figure is the fraction of zeros proven to lie on the critical line, not a statement that the Riemann hypothesis is 67.2% true. The hypothesis demands 100%, the remaining 32.8% is still open, and Anthropic states directly that it does not expect the techniques Claude used to lead to a proof of the hypothesis. The party announcing the result drew its own ceiling, and that sentence deserves to travel with the headline number.

One further distinction concerns the kind of difficulty involved. This was a constant-improvement problem sitting on an existing research lineage, success could be judged numerically, and partial progress accumulated. Those conditions suit brute parallel search unusually well. Whether the same method transfers to problems with no crisp success criterion, or problems requiring a new concept, is not something a single case can settle.

What transfers to other organizations is the workflow, not the mathematics

For practitioners, the transferable part of this case is the task architecture rather than the zeta function. Three things were present together. The problem admitted machine-checkable success and failure. A parallel search structure could absorb 650 failures. And a verification layer confirmed the final artifact without depending on human reading.

The third element is the one most often missing when teams hand long tasks to agents. Tooling already supports widening a search by adding subagents, but confirming that the output is correct usually remains a person reading the output. Scaling search without scaling verification only lengthens the review queue. In software work, tests and type checking occupy the position Lean holds here; in data work, recomputation and cross-checking against an independent source do.

Problem selection changes accordingly. The one decision a human genuinely made in this case was which problem to attack, and after that a judgeable metric supplied the direction. The first question to ask about any task you plan to delegate to an agent is therefore not how hard it is, but whether success can be judged without you.

Open questions: an unreproducible model and the scope of review

The largest gap around this result is that the model is not public. The bound was raised by an unreleased research version of Claude, and no outside researcher can currently retrace the process. Anthropic published Claude's paper, an informal note, an explanation with appendix, detailed transcripts, and the Lean formalization repository, so the artifact is open to verification even though the process is not open to reproduction.

The scope of review also deserves stating precisely. Two outside experts examined the paper on short notice, which is not the same procedure as journal peer review. Mathematics normally settles a result through many researchers reading it carefully over months, and where this one lands will be decided in that process. The Lean formalization substitutes for a large share of that scrutiny, but confirming that the formalized statement matches the prose claim remains a human task.

What remains is the question of how far this generalizes. One success reveals nothing about a success rate, and there is no published account of how many unsolved problems the same model attacked before this one produced a result. If anything, a record of 650 failed ideas points toward a low hit rate for the method. The next real signal is a second case, and until one arrives this result is most accurately treated as a single occurrence with a sample size of one.

Source: ASAP analysis based on Anthropic's announcement "Learning more about Claude's mathematical capabilities" (August 10, 2026) and the proof note published alongside it