The most capable LLM is not the right model for every request. Sending every task to a frontier model drives up cost. Sending everything to a smaller model buys a different bill: failed tool calls,...