the token costs are real. we switched to smaller models for 80% of tasks and barely noticed