Claude pushed the Riemann zeta zero-on-line lower bound from 41.6% to 67.2%: 60 subagents, 31M output tokens, and a Lean proof — the first draft of an "AI math construction site"
The Riemann hypothesis is not solved. But around it, the "construction site" itself has changed how math gets done. On August 10, Anthropic published a blog post disclosing that an unreleased research version of Claude, while taking a swing at the Riemann hypothesis, accidentally pushed a long-standing related lower bound from 41.6% to 67.2% (anthropic.com/research/riemann-zeta). The interesting part isn't the number — it's how the number was produced. A non-mathematician Anthropic staff member, Jarred Sumner, typed one prompt — "take a real stab" — and Claude burned 31M output tokens, coordinated roughly 60 subagents, ran 2,400 shell commands, cross-validated thousands of known zeros against each other, and finally wrote a Lean formal proof that passes a standard comparator check.
Not "AI doing science" but a coordinated "AI construction site"
In the first session Claude tried 650 ideas, all of them dead-ends. The second session re-shaped itself into a multi-agent cluster: 2 agents responsible for the core mathematical ideas, 13 contributing ideas, 30 attempting but unable to develop new ones, 13 acting as validators checking the correctness of arguments, and a final 2 drafting the initial paper — 60 subagents in total. Subagents refereed each other's work, searched for counterexamples, and independently re-derived the finding from scratch (anthropic.com/research/riemann-zeta). Throughout, Jarred's input is wryly described as "mostly variants of 'keep going' or 'believe in yourself'."
The signal here is not the upper bound on a single model's intelligence, but the fact that the combination of multi-agent orchestration, tool use, long-context coherence, and formal proof closure has, for the first time, produced visible output on a concrete mathematical sub-problem.
What 67.2% actually means
A small amount of math background is helpful here. Among the zeros of the Riemann ζ function, mathematicians had, over decades, incrementally pushed the known minimum proportion of zeros lying on the critical line from zero up to 41.6%. The new bound of 67.2% is not a proof of the Riemann hypothesis; it is a substantial refresh of "the lower bound on how many of the zeros we can prove sit on the line." Drawing on the work of Aryan (2019), Baluyot–Goldston–Suriajaya–Turnage-Butterbaugh (2023, 2025), and Bombieri (2000), Claude's key move was "treating the entire quadratic-form space — positive-definite and negative-definite parts together, with non-diagonal quadratic forms allowed." That "courage to treat the whole space at once" is, in Anthropic's words, what allowed Claude to clear the prior bound (anthropic.com/research/riemann-zeta).
A separate piece of engineering deserves highlighting. Anthropic's own mathematicians, Levent Alpöge and Ralph Furman, reviewed the work and produced an informal note for experts. In parallel, Claude worked with engineer Eric Easley to produce a Lean 4 formalization of the result, which passes the standard comparator validation tool (github.com/anthropics/zeta-23-lean). The Lean step turns the proof into something a machine can independently re-check at any time — "sounds right" gets locked into "verifiably right."
Why this reads more like a workflow demo than a model breakthrough
Read against the broader LLM capability curve, two points stand out. First, Anthropic is explicit at the end of the post: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis." The company did not position this as an AGI leap; it honestly framed it as "the latest example of the speed of progress in AI models' mathematical capabilities" (anthropic.com/research/riemann-zeta). Second, Anthropic admits Claude was initially skeptical it could make meaningful progress — likely because the training data taught it both that "open math problems are hard" and "AI models have limits." A few rounds of encouragement (the post notes the same kind of encouraging prompt was used earlier to help Claude disprove the Jacobian conjecture) pushed it past that self-doubt and onto the result.
This points to an angle worth more discussion than "AI doing math": what current-generation LLMs actually lack, in open-ended math research, isn't depth of inference — it's an end-to-end workflow that can sustain progress, cross-validate, and produce a publishable artifact. Drop the subagent orchestration, drop the long-horizon memory, drop the tool use, drop the Lean check, and the 650 dead-end ideas from the first session become the default outcome.
What the industry can take from this
From the perspective of an engineer writing LLM workflows, the parts worth copying are not the raw 31M tokens or the 60 subagents — they are the three closing loops layered on top. First, subagents referee one another, rather than relying on a single agent's internal self-check. Second, the system downloaded 54 arXiv papers to do a literature sanity check — explicitly asking "has someone already done this?" Third, Lean formalization acts as a final gate: even after human expert review, a machine-checkable proof locks the conclusion in place (github.com/anthropics/zeta-23-lean). That three-layer closure is, plausibly, the standard action set for AI-assisted scientific research over the next stretch.
As for the Riemann hypothesis itself: this clearly didn't solve it. But a 67.2% lower bound, accompanied by a Lean formalization, signed off by two staff mathematicians, and confirmed against arXiv, is one of the few AI outputs in serious math that can actually go through peer review. The next thing worth watching is which lab productizes this multi-agent "math construction site" workflow into something a non-mathematician engineer can pick up and use.
References: Anthropic blog post, Lean formalization repo.