Anthropic Multi-Agent Model Tiering: Opus Orchestrator + Sonnet Workers Achieves 90.2% Accuracy Improvement

Anthropic Engineering·high signal

Anthropic's engineering blog documents the production architecture behind Claude Research: a lead Opus 4.6 agent decomposes queries, spawns parallel Sonnet 4.6 subagents per sub-question, then synthesizes results. This configuration outperformed single-agent Opus 4.6 by 90.2% on their internal eval. Token cost scaling: single-agent chat is 1x, single agentic is 4x, multi-agent is 15x — but quality gains justify the cost for complex research tasks.