A strong model plans, a cheap model executes the bulk of the work.

Anthropic reduces agent costs via its advisor tool and Claude Managed Agents by routing heavy execution to cheaper models while keeping capable ones for planning.

Anthropic documents two ways to reduce agent costs by pairing a capable model with a lighter model, with the majority of generation billed at the latter's rate.

The first is via the advisor tool, an API function in beta. A faster, cheaper executor model consults a more capable advisor model during generation: the advisor reads the entire conversation, returns a plan or course correction, and then the executor takes over. The exchange occurs server-side in a single request, with the executor choosing when to call and the server providing the context. The advisor works without tools and only returns its advice text. The savings come from it not producing the final output: the executor handles that, at its lower rate.

The second involves multi-agent orchestration, via Claude Managed Agents. A coordinator model plans and then delegates to worker sub-agents, each operating in a thread with an isolated context. For a web search task, the workers handle the token-intensive page reading and only send summaries back to the coordinator: the pages browsed never enter its context. Here again, the bulk of the volume is billed at the workers' rate.

Both approaches stem from the same observation: an agent combines a small part of planning and judgment with a large part of mechanical reading and execution, which can be directed to the appropriate model. Both the advisor tool and Claude Managed Agents remain in beta on the API.