The cheapest model call can produce the most expensive agent run

I used to think agent cost tracking started with token usage. Record the model, multiply input and output tokens by the provider price, then compare routes. That works for reconciling a bill. It break