Match each part of an OpenAI tool-call round-trip to the usage field that bills it

OpenAI bills two things on every API call: the **prompt tokens** you sent in, and the **completion tokens** the model wrote out. Tool calling does not change that rule. It just adds more pieces to each turn. A tool-using request typically takes two turns. On turn one, you send the conversation plus the schemas describing what tools exist. The schemas count as prompt tokens. The model replies with either text or a `tool_calls` array containing the function name and JSON arguments. The arguments are tokens the model generated, so they count as completion tokens. On turn two, your code runs the actual tool, gets a result, and sends a new request that includes the prior messages, the assistant's tool-call message, and your tool-result message. All of that is now prompt tokens on turn two. The model finally writes a user-facing answer; that text is the second turn's completion tokens. The cumulative bill for an agent loop is the sum across every turn, not just the last one. Multi-step agents that take five hops pay five separate prompt plus completion bills.

**Tool calling** in the OpenAI API is just a structured conversation, so the billing rule never breaks: anything in the request body is `prompt_tokens`, anything the model produces is `completion_tokens`, on a per-turn basis. The tool schema (the `tools` parameter with each function's name, description, and JSON Schema) is part of the request, so it bills as input tokens on every turn it is sent. Verbose schemas add up across long sessions. The assistant message containing `tool_calls` (the function name plus JSON arguments) was generated by the model, so it counts as `completion_tokens` on the turn it was emitted. On the next turn you send the conversation back including that assistant message and a `role: "tool"` message with the function's return value. The whole transcript, schemas included, is now prompt tokens for turn two. The model's final text reply is turn two's completion tokens. The trap is the cumulative cost. A five-step agent loop sends the growing transcript five times, so prompt tokens grow roughly quadratically with steps unless you summarize or use prompt caching. Engineers who quote only the final-turn usage understate the bill by a factor equal to the number of hops.