Anthropic's unreleased research version of Claude has solved a major piece of one of mathematics' oldest unsolved puzzles, advancing the Riemann hypothesis by proving that at least 67.25% of nontrivial zeros lie on the critical line. This represents a significant jump from the previous record of 41.6%, set by mathematicians Kyle Pratt, Nicolas Robles, Alexandru Zaharescu, and Dirk Zeindler.

The breakthrough is notable not just for the mathematical result, but for how it was achieved. Rather than a single AI model working alone, Anthropic deployed approximately 60 Claude subagents working in parallel, operating continuously for 36 hours. The swarm consumed roughly 31 million output tokens, issued 2,400 shell commands, wrote hundreds of Python scripts, and reviewed 54 research papers from arXiv. This approach demonstrates that scaling inference, not just model size, can unlock reasoning capabilities that single-shot prompting cannot reach.

What Makes This Different From Previous AI Math Breakthroughs?

The Claude swarm did not invent entirely new mathematical machinery. Instead, it identified a novel combination of existing techniques: Bombieri's 2000 work paired with recent papers by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh. This synthesis allowed methods introduced by Montgomery in 1973, which previously required assuming the Riemann hypothesis was true, to be applied unconditionally. The model essentially found a creative bridge between established mathematical frameworks that human researchers had not yet connected.

To ensure the result was mathematically valid, Anthropic took an unusual step: it produced a machine-checkable Lean formalization of the proof. This mechanical verification removes human intuition from the validation process, a critical requirement when evaluating AI output in exact sciences. The work was then reviewed by internal staff mathematicians Levent Alpöge and Ralph Furman, followed by external validation from Brian Conrey and Dan Goldston.

The paper also claims that at least 83.625% of the zeros are distinct and provides analogous results for fixed primitive Dirichlet L-functions, extending the implications beyond the core Riemann hypothesis result.

How to Build Multi-Agent Systems for Complex Reasoning Tasks

• Design for Extended Execution: Production deployments for complex reasoning require infrastructure designed to sustain days of continuous execution, not just minutes or hours. The Claude swarm ran for 36 hours straight, requiring robust state management and error recovery mechanisms.
• Implement Formal Verification Pipelines: When AI output must meet exact standards, build verification systems that can check results mechanically. Machine-checkable proofs like Lean formalization remove ambiguity and provide confidence that results are correct.
• Use Parallel Subagent Topologies: Rather than relying on a single model instance, orchestrate multiple specialized subagents that can explore different approaches simultaneously. This decoupled exploration enables deeper reasoning than sequential processing.

Human involvement during the run was minimal, consisting primarily of prompt continuations from project initiator Jarred Sumner after an initial attempt generated 650 dead-end ideas. This suggests that once a multi-agent system is properly configured, it can operate with light human guidance, making course corrections only when necessary.

What Does This Mean for Claude's Capabilities?

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%, indicating a sharp increase in reasoning depth and multi-step logic compared to 2024 and 2025 baseline models. Claude Opus 5 launched on July 24, 2026, and represents Anthropic's most capable reasoning model to date.

The practical implication is clear: if you are building systems that require sustained, multi-step reasoning over extended periods, inference scaling yields results that traditional single-shot prompting cannot reach. The swarm approach works because it allows the model to explore multiple reasoning paths in parallel, validate intermediate results, and recover from dead ends without losing progress.

For developers and organizations considering AI for research, mathematics, or other deep reasoning tasks, this result suggests that the bottleneck is no longer model intelligence alone. Instead, it is the infrastructure and orchestration needed to let intelligent models think for hours at a time, with proper validation and error handling built in.