Picking the model per task
Top-tier models for design and debugging; Sonnet for the loop; Haiku for the grunt work.
4 min
Price tiers are an order of magnitude apart
Fable-class models list at roughly five times Sonnet 5; Opus at two and a half times. A mechanical turn (run the tests, read the diff, apply the obvious fix) produces the same tool call on any of them.
ReadyClaw flags mechanical turns: top-tier turns that emitted only tool calls and under 300 output tokens. In a typical coding session they are a fifth to a third of all turns.
A working split
Open the session on the model the hardest part needs: architecture, a gnarly bug, a security review. Hand the mechanical middle to subagents with an explicit model (Sonnet for edits and test loops, Haiku for searches and renames). Come back to the top model to review.
Check CLAUDE_CODE_SUBAGENT_MODEL in ~/.claude/settings.json. Left pointing at a superseded model, every subagent quietly pays last year's price for last year's quality.
npx readyclaw scan