Model Cascading in Practice: Let a Cheap LLM Handle 80% of Requests and Escalate Only When It's Unsure