Most AI model providers charge per token, usually quoted per million tokens, with separate rates for input and output. Rates differ by model, and larger and more capable models cost more. Some offer discounts for cached input or for work submitted in batches that does not need an instant answer.
The cost of a request is the input tokens times the input rate plus the output tokens times the output rate. That makes it easy to estimate for a single call and harder to predict for a product. A chat that resends its history, a retrieval system that attaches long documents or an agent that makes dozens of model calls per task can each cost far more than a quick test suggests.
Pricing changes often, and competitors cut prices and launch new tiers.
Estimate costs with real traffic, set budgets and alerts, use smaller models for simple steps, and check the current price list before you commit to a plan.