You pick one Claude model and send everything to it. Usually the biggest one, because bigger feels safer.
That costs ten times the price on jobs that need none of it.
It burns your limits by lunch on the days you needed the headroom most.
The spec sheet picks wrong. Every picker you have seen is copied off the model card, and benchmark scores answer which model is best. Your question is which one is enough.
Sort by job class instead. One test decides it: what it costs you when the answer is wrong.
Which model for which job
Haiku for sorting
$1 in, $5 out per million tokens. 1x. High volume, one right answer, and a bad call is obvious at a glance.
Receipt: my topic system reads creator posts every week and rules each one worth a breakdown or not.
Hundreds of one-bit calls, all pinned to the cheap tier.
Sonnet for building
$2 in, $10 out. 2x. The spec already exists and the job is executing it faithfully.
Receipt: this email was written on Sonnet and reviewed by a second Sonnet in a fresh context.
The seat that writes the HTML behind my graphics runs there too.
Opus for judgment
$5 in, $25 out. 5x. No single right answer, and a wrong call compounds quietly for weeks.
Receipt: every judgment seat in my graphics pipeline sits here. Strategy, layout, pre-flight, scoring, the critic.
Nothing ships below 95 out of 100.
Fable for overturning
$10 in, $50 out. 10x. A second opinion on a conclusion you are about to bet on.
Receipt: any post of mine above 50,000 impressions gets re-read on the top model, whose only job is to attack the earlier verdict.
It ran once, on a post that settled at 128,280, and it overturned both write-ups that came before it.
The tier under Haiku
Zero. If you can write down the pass condition, write the script instead.
Receipt: two gates in my own pipeline are plain code with no model calls, and they are built to overrule the writer.
Do this now
- List the five Claude jobs you ran most in the last two weeks.
- Mark each one cheap to catch or expensive to catch when it goes wrong.
- Move every cheap-to-catch job down one tier, and leave exactly one job at the top.
- Run the five for a week. If the output holds, the step-down is real.
↳ Note: those prices are current API rates, checked this week. On a subscription you are not paying per token, and the same split still decides how fast your limits drain.
Reply and tell me the one job you are still sending to the biggest model. I read every reply and I will tell you where it belongs.