Per-token pricing is fine at demo scale and terrible at real scale. Small task-specific models can be ~100× cheaper per conversation. Mature teams route by difficulty: narrow models for the high-volume well-defined slice, frontier models only for the hard fraction. Report cost per correct answer.