百年黎曼猜想被Claude破了新纪录！是个未公开新模型

TL;DR - Anthropic announced that an unreleased Claude research model, running as a ~60-agent swarm, raised the proven lower bound on the proportion of Riemann zeta non-trivial zeros lying on the critical line from 41.6% to 67.2% — a long-stalled record — without proving the Riemann Hypothesis itself.

• Setup: given only "try to prove the Riemann Hypothesis," Claude first generated and failed ~650 approaches; a simple "try again" prompt triggered a second run using ~60 sub-agents in Claude Code over ~1.5 days, 2,400 shell commands, hundreds of Python scripts, and ~31M output tokens.
• Agent division of labor: 2 agents produced the key mathematical ideas, 13 contributed supporting ideas, 30 explored dead ends, 13 verified arguments, and 2 drafted the paper; the human's role was mostly encouragement.
• Result came from recombining existing published results rather than new machinery; Claude self-checked via cross-review agents, counterexample search, 54 arXiv papers for novelty checking, and an independent re-derivation.
• Validation: reviewed by Anthropic mathematicians Levent Alpöge and Ralph Furman plus analytic number theorists Brian Conrey and Dan Goldston, and formalized in Lean (passing Comparator). Anthropic notes it is not traditional peer review and the method likely cannot extend to a full proof. The model's identity (speculation about "Fable 5.1/6") is unconfirmed rumor.