Per-request math is usually small. A typical question and answer might use a few thousand tokens, a fraction of a cent. Costs become real at scale, or when you feed in large documents on every call. Caching, picking a cheaper model for simple steps, and trimming context all cut the bill.
The bigger line item is almost always the build: designing the feature, connecting your data, testing for accuracy, and shipping something reliable. The model API is often the cheapest part of an AI project.
Updated July 2026
Go deeper