What does running several subagents at once, each returning a detailed result, cost you?

The principle — delegation moves a cost rather than removing it, and the two costs move in different directions.

Why the key is correct — Anthropic states both halves plainly. Running several sessions or subagents at once multiplies token usage, and running many subagents that each return detailed results can consume significant context. The isolation buys you the worker's raw reading; it does not buy you its conclusions for free.

Why the others are wrong — workers share no budget, so the cost is not paid once. The token bill grows rather than staying flat, because each worker gathers context of its own. And a separate window protects you from what the worker read, not from what it sends back.

Remember this — delegation is a context saving bought with tokens, and enough returned summaries will spend the saving.

Sources — Anthropic's Claude Code agents and subagents documentation.