How are people estimating LLM API costs before putting an AI application into production?

I usually start with a spreadsheet that estimates average input and output tokens per request, then multiply those figures by the expected monthly request volume. I plug in each model’s current pricing to compare costs. I also account for conversation history, retries, and peak usage, since they can significantly affect the final bill. For example, if an AI application uses the Nonprofit Check Plus API for verification, I’d include those API calls in the overall usage estimate. Running a small test with real prompts before launch helps validate the estimates.