An unreleased research build of Claude raised the proven lower bound on the fraction of Riemann zeta zeros lying on the critical line from 41.6% to 67.2%, according to a research blog post Anthropic published Monday. The jump is the largest single-step advance on the bound in decades, and Anthropic is careful to note upfront: “We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis.” The Clay Mathematics Institute’s $1 million bounty, one of seven Millennium Prize purses, isn’t in play. What’s in play is the credibility of AI-generated mathematics as such.
The origin story is almost aggressively casual. Jarred Sumner, an Anthropic staffer who isn’t a mathematician, prompted the model to “take a real stab” at the problem, and then, per Anthropic’s own writeup, was “mostly limited to sending Claude messages of encouragement (mostly variants of ‘keep going’ or ‘believe in yourself’).” The run lasted roughly 36 hours across two sessions inside Claude Code, burning 31 million output tokens.
Under the hood, it was a swarm. Claude generated 650 ideas that failed, then orchestrated approximately 60 subagents that ran 2,400 shell commands and wrote hundreds of Python scripts. Anthropic’s post-hoc anatomy of the swarm is worth reading closely: 2 agents produced the load-bearing mathematical ideas, 13 contributed supporting ideas, 13 acted as validators, and 30 tried and failed to generate anything usable. The distribution looks a lot like a human research group.
The mathematical move is combinatorial, not foundational. Claude stitched recent work by Siegfred Baluyot, Daniel Goldston, Ade Irma Suriajaya, and Caroline Turnage-Butterbaugh to Enrico Bombieri’s 2000 paper, then redeployed techniques Hugh Montgomery introduced in 1973 (originally conditional on Riemann holding) in an unconditional setting. Anthropic describes the construction as “a suitable space of functions with quadratic form induced by Weil.” Brian Conrey and Dan Goldston, both outside number theorists, examined the paper on short notice. Anthropic’s mathematicians Levent Alpöge and Ralph Furman shepherded the work internally, and pieces of the proof were formalized in Lean.
The context that makes this land is arriving from the opposite direction. Earlier this month, Scientific American reported that mathematicians examining OpenAI’s Astra writeup found it recycled arguments from prior papers, including 2016 work by Yeshiva University’s Steven Miller, who called the pattern “research misconduct.” Anthropic’s release is legible against that backdrop as a deliberate counter-posture: named validators, a Lean formalization, a public technical account.
The posture isn’t airtight. Anthropic hasn’t released the model checkpoint or full process transcripts, and AI Weekly notes a SymPy script is missing from the Lean repository. Claim and audit trail are diverging even as the field’s narrative-management demands tighten around them.
Sources
- https://www.anthropic.com/research/riemann-zeta
- https://www.scientificamerican.com/article/openais-latest-math-breakthroughs-commit-research-misconduct-experts-say/
- https://theaiinsider.tech/2026/08/11/anthropic-says-claude-improved-a-longstanding-bound-tied-to-the-riemann-hypothesis/
- https://www.neowin.net/news/unreleased-claude-model-makes-breakthrough-on-century-old-riemann-hypothesis-math-problem/
- https://aiweekly.co/alerts/anthropic-unreleased-claude-improves-zeta-bound-to-672