Pick the model by what a wrong answer costs

"Use medium effort. Never max." was the Claude tip all over my feed today. It's a good default. As a rule for every task, it's wrong in both directions. It overpays for work a script should do, and it underpays for the work where a mistake is expensive.

I run a small set of scheduled agents for my own work. Before anything runs, each task gets one question: what does a wrong answer cost?

The routing table

Task	If it's wrong	What runs it
Counting, dates, limits, duplicates	It shouldn't be wrong at all	A plain script, no model
Many small yes/no checks	One bad item	A small decision model, every item in one batched call
Reading pages, filling forms	A retry	Sonnet
Pulling facts out of long pages	A cluttered main context	A Haiku sub-agent that hands back a short summary
Planning and hard calls	Days of rework	Opus
Code that ships	A bug in production	Opus with high effort

A script first

Anything with an answer you can compute never goes to a model. Dates, caps, word counts, duplicate checks. A model can get these wrong. A script can't, and it costs nothing to run again.

Small models for volume

Yes/no questions over many items go to a small decision model in one call. This afternoon it screened 40 LinkedIn posts for $0.0016. It only filters and ranks. It never makes the final call on anything that goes out in my name.

The big model where a mistake is expensive

Planning and code that ships get the strongest model and more effort. The extra tokens are cheap next to a bad plan or a production bug. A sub-agent (a helper with its own context window) does the long reading, so the expensive session only sees the summary.

Before you pick a setting

• Can a script answer it? Then no model.
• Is it many small items? Batch them to a small model and spot-check a sample.
• Would a wrong answer reach a person or production? Pay for the big model.
• Check the routing again when a model or a price changes.

Medium is a good default. The routing is the rule.