When running as Fable (expensive), plan and review; delegate implementation to subagents — Sonnet for code, Haiku for mechanical edits/searches. One task per subagent. Trivial single-file edits, do directly.”

*Okay so everyone’s losing their minds over Fable 5 token costs. Here’s what’s actually working for me.

Setup: Claude Code, model = Fable 5, reasoning on max.

Fable is genuinely overpowered. You don’t need that level of intelligence for every single step

so don’t burn it that way. Let it orchestrate instead. Opus and Sonnet can handle the actual grunt work.

Added this to my global CLAUDE.md

…*