The $0.04 Decision Model Is Quietly Replacing a Lot of My Prompt Engineering

I used to write elaborate prompts to make a frontier model decide things: spam or not, tier A or B, escalate or close.

Perplexity's pplx-decider at $0.04 per million tokens and its peers made me realize those prompts were expensive theater.

A two-billion-parameter decision model does the classification in one pass and returns a clean label, no prose to parse.

I moved a dozen of those calls off the big model and the bill dropped without the user noticing a difference.

Prompt engineering isn't dead, but a chunk of it was never really prompting. It was decision-making we overpaid for.