The work you don't see still has a cost.
The visible answer is not always the whole billed output. Reasoning models can perform additional work, and your application can create extra requests while trying to get the result right.
Read the usage record
Inspect the provider's documented usage fields and current billing rules. Track reasoning where the service exposes it, alongside ordinary input and output. Compare invoice totals with application logs so hidden retries are easier to spot.
Use effort with a purpose
Try lower effort or a simpler model on routine tasks, if the API supports the option. Set output and retry limits. Evaluate the result before reducing the budget for work where extra reasoning materially improves accuracy.
Budget for useful outcomes
Calculate spend per completed task, not answer length. Some short answers require substantial work; some long answers add little value. Monitor both quality and usage when changing models, so a smaller bill doesn't quietly create more manual correction.