Let the model pick the model

The usual advice for taming an expensive coding model is to write rules: use the cheap model for edits, the pricey one for design. Simon Willison, distilling a tip from Anthropic's Claude Code team at a conference fireside chat, drops the rulebook and delegates the decision instead. His standing instruction to Claude Code is a single sentence: for all coding tasks, use your judgement to pick an appropriate lower-power model and run it in a subagent.

The result, as he describes it, is a self-organising hierarchy. Fable, the top-tier model that bills at roughly twice the rate of Opus, keeps the design, audit, and judgement calls, and spins up Sonnet for real implementation work and cheap Haiku for mechanical edits. The point isn't the tiering, which is old. It's who does the routing: not a human writing per-task rules ahead of time, but the model deciding mid-loop, task by task, which tier each piece of work deserves.

The pattern's real parent is a testing tip. The Anthropic team's advice was to stop scripting rigid rules like 'only write tests for larger features' and instead tell the model to use its own judgement about when tests are worth it. Willison's move was to point that same judgement-delegation at model choice. It doesn't come with a measurement: he reports his Fable allowance 'shrinking less quickly than before' but never quantifies the saving, and offers no check on whether the model's routing calls are any good or whether the cheap output needs redoing. But the shift it points at is the real payload — routing is becoming a judgement the model makes mid-loop, task by task, rather than a rule a human writes in advance.