60-Agent Claude Swarm Proves 67.2% Riemann Zeta Lower Bound

Anthropic’s unreleased research build of Claude established a new lower bound for the Riemann hypothesis, proving that at least 67.25% of nontrivial zeros lie on the critical line. The result advances a 167-year-old mathematical problem, moving the established bound up from the previous record of 41.6%.

The breakthrough centers on the proportion of zeros of the Riemann zeta function that satisfy the condition predicted by the hypothesis, which states they must have a real part of 1/2. The previous 41.6% record was established by mathematicians Kyle Pratt, Nicolas Robles, Alexandru Zaharescu, and Dirk Zeindler.

Multi-Agent Research Topology

The proof required 36 hours of continuous autonomous execution. The model operated within Claude Code using a multi-agent topology of approximately 60 subagents. Across two sessions, the swarm consumed roughly 31 million output tokens. Generating 31 million reasoning tokens requires significant compute isolation, state management, and error recovery mechanisms.

During execution, the model issued 2,400 shell commands, wrote hundreds of Python scripts, and reviewed 54 arXiv papers. Developers orchestrating parallel subagents in Claude Code can observe the necessity of decoupled exploration for deep reasoning tasks. Human involvement during the run was minimal, consisting primarily of prompt continuations from project initiator Jarred Sumner after an initial failed attempt generated 650 dead-end ideas.

The Mathematical Methodology

The model did not invent entirely new mathematical machinery. It identified a novel combination of Bombieri’s 2000 work with a recent series of papers by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh. This synthesis allowed techniques introduced by Montgomery in 1973, which previously required assuming the Riemann hypothesis was true, to be applied unconditionally.

The resulting paper also claims that at least 83.625% of the zeros are distinct. It provides analogous results for fixed primitive Dirichlet L-functions.

Verification and Lean Formalization

Anthropic validated the results through internal staff mathematicians Levent Alpöge and Ralph Furman, followed by external review from Brian Conrey and Dan Goldston. To ensure strict mathematical validity, the model produced a machine-checkable Lean formalization of the proof. This mechanical verification removes human intuition from the validation loop, a critical requirement when evaluating AI output for exact sciences.

The achievement coincides with the emergence of Riemann-Bench, a new private benchmark for research-level mathematics. On August 5, 2026, a model identified as Claude Opus 5 reportedly led this benchmark with a score of 79.0%. This indicates a sharp increase in reasoning depth and multi-step logic over 2024 and 2025 baseline models.

If you build multi-agent systems, this execution demonstrates that inference scaling yields results traditional single-shot prompting cannot reach. Production deployments for complex reasoning require infrastructure designed for days of execution, structured formal verification pipelines, and parallel subagent topologies.