Tool call

When the model requests your code to run a function with arguments.

A tool call is the model's request for your application to execute a function (search, database query, API call) and return the result. Every tool-calling request includes the tool schemas as input tokens, adding 500-5000 tokens per call depending on how many tools are defined. Multi-step agent loops compound this: each loop iteration replays the full prompt, history, and tool schemas. This is why agent workloads cost 5-10x more than single-turn chats at equal user-facing output length.