0
I’m planning a feature that sends user requests to an OpenAI model, and I’m trying to budget for real usage rather than just test a handful of prompts. The requests vary in length, and some may need tool calls or a second model pass.
What’s a sensible way to estimate monthly spend before launch? I’m especially interested in how people account for input and output tokens, retries, caching, and changes in usage as a product grows. Do you build a small workload test first, or rely on expected average token counts?
