Single-shot vs agent loops
A chat completion often bills one input and one output. An agent may call tools many times, re-sending the system prompt and accumulating tool results each turn—so input tokens grow with every step.
What usually dominates the bill
Large tool payloads (search pages, code files, JSON APIs) and long system instructions resent every turn. Prompt caching can cut the system-prompt portion when the provider supports it.
How to plan spend
Estimate tool calls per run, average tool-result size, and monthly run volume. Use the Agent cost estimator to compare models before you lock architecture.
FAQ
Does every tool call create a new model turn?
In most tool-calling designs, yes: the model emits a tool call, you execute it, then you send the result back for another model turn until a final answer.