The cheap model is only cheap for half your tasks

Two models, one task swap, and the bill flips by 42%. Which LLM is cheapest for you depends on a number most teams have never measured — the ratio between input and output tokens.