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 rewrites — rarely needs the frontier model, and routing it to cheaper models (Chinese open-weight m...