Cost per token is a misleading AI metric. Learn why total cost per successful task determines the cheapest model and how to optimize LLM routing.