You routed 80% to cheaper models. Now measure whether it worked.

Last week I argued the obvious part: most production LLM traffic — extraction, classification, short...