Routing read/classify steps to a cheap model cut cost with no quality loss

Summary: Not every phase needs the frontier model. Reserving the strong model for reasoning/generation and sending the read-and-classify steps to a cheap model kept quality where it mattered and cut cost and latency where it didn't.

What was encountered

Every phase could have gone to the strong model, but several phases only read text and classify — e.g. "did the PM's comment answer question Q1?". Paying frontier prices for that kind of step is pure waste with no quality gain.

How it was solved

Split the pipeline by a single question — is this reasoning, or is this reading? Sonnet (strong) writes SQL: the clarify and draft phases. Haiku (cheap, fast) reads and classifies: preprocess, auto_answer, validate.

Reference: Model routing — a reusable instance of it. Caveat worth remembering: if a "classification" step is actually subtle judgment, the cheap model will quietly get it wrong and poison everything downstream — promote that step back to the strong model.