Monday Jul 6
STRATEGYAnthropic

Willison Lets Claude Pick Its Model

6JUL
CLAUDESUBAGENT

A simple rule is cutting one operator's AI coding costs. Simon Willison, the Datasette creator, tells Claude to pick its own model per subagent task. His 'Fable allowance' is now shrinking slower than before.

The rule in one line: let the system decide which tier of itself should do the work. Design and review decisions stay at the top tier.

The cheapest option handles trivial edits. Heavier implementation work runs one tier down, inside a background process. Judgment and synthesis never leave the main thread.

The payoff shows up in the invoice, not the output. It's a pattern any team running coding agents could copy this week.

full brief & sources

⚡ Why this matters

  • Every team running coding agents faces the same tradeoff: quality versus cost per task.
  • Willison is one of the most credible operator voices in this space, and he's showing his actual math.
  • This is a concrete, copyable pattern, not a vague 'use AI wisely' platitude.

🔍 What happened

  • Posted July 3 on simonwillison.net under the title 'Fable's judgement.'
  • Rule: for coding tasks, let Claude use its judgement to pick an appropriate lower-power model and run it in a subagent.
  • Trivial edits route to Haiku. Substantive implementation routes to Sonnet in a subagent.
  • Design, auditing, data synthesis, and judgment-heavy work stay with the main, most capable model.
  • Willison reports his Fable usage allowance is shrinking more slowly since adopting the rule.

💬 Smart takes

  • Simon Willison: the approach has been working well, and his Fable allowance is shrinking less quickly than before.
  • Skeptic: this only works if you already trust the model's judgement about task difficulty, which took Willison years of hands-on use to calibrate.

🧭 Where this goes

  1. Likelymore coding-agent power users adopt explicit model-tiering rules like this one.
  2. LikelyIDE and agent tools start exposing built-in 'let the model choose its own tier' settings.
  3. Possiblethis becomes a documented best practice in Claude Code and Cursor onboarding guides.
  4. Wild Cardmodel routing becomes fully automatic and removes the need for this kind of manual rule within a year.

🥄 The Spoon Take

The cheapest way to cut AI coding costs isn't a smaller model. It's better delegation. Willison keeps judgment at the top and pushes routine work down the stack. That's just good management, applied to agents instead of people.

🤔 Pushback

This only works because Willison has years of calibration on when to trust the model's own judgement.

·Simon Willison - Fable's judgement