PRODUCTION SCENARIO
Your orchestrator on Claude Opus 5 fans out simple lookups, fetch a record and extract three fields, to 20 subagents per request. Subagent tokens are roughly 90 percent of the bill. The orchestrator's own planning quality has to stay exactly where it is, and your platform team keeps the whole fleet on one model id so evaluation results stay comparable across services.
How should effort be set across the orchestrator and its subagents?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Leave the orchestrator at its high default and run each of the 20 subagents at low effort
Low effort is documented for simpler tasks that need the best speed and lowest costs, naming subagents specifically, while high stays right for complex reasoning and agentic orchestration; that maps onto a bill where the subagents are 90 percent of the spend and the planning has to hold. Lowering the orchestrator is the right move when its own generation dominates and its reasoning is not load-bearing, which is the reverse of this split. A uniform medium is reasonable when both halves tolerate the same level, but it saves less on the fan-out and still steps the planning down. A smaller model for trivial lookups is a sound design wherever a second model id is acceptable, and this fleet is held to one.
Official Source: https://platform.claude.com/docs/en/build-with-claude/effort ↗