Anthropic researchers describe using Claude in a search-and-evaluation loop to propose quantum-circuit changes. The technical preprint appeared on arXiv October 15, with the company’s research explanation dated October 16.
Across eight optimization tasks, Anthropic reports lower-cost circuits on seven and highlights roughly a 12-fold reduction for one algorithm step. These are author-reported benchmark outcomes, not measurements from an operational quantum computer.
The work applies AI to a specialized engineering design space, where a model proposes candidates and a task-specific evaluator checks them. If independently reproduced, this approach might lower resource estimates for some circuits.
The paper is a preprint and has not been peer reviewed. Reported gains depend on task selection, cost metrics and evaluator design; independent replication was not established in this review.