Skills
Route models by role to cut subagent bills 40–70%: Opus plans, Haiku/Sonnet execute
Subagent-heavy Claude Code workflows can burn roughly 7x the tokens of a single thread because each subagent re-reads context in its own fresh window and bills its own calls. The biggest lever is role-based model routing — keep the planner/reviewer on Opus but run the worker subagents on Haiku (5x cheaper than Opus) or Sonnet — which practitioners report cuts focused-task cost 40–70%, with the largest savings when a context-mode subagent absorbs heavy MCP tool use. Push the heavy context-reading down into the workers so your main session stays cheap and focused on direction and review.
↳ Follow the thread