The stage where a trained model is actually run to produce an answer for a new request; it is what you pay for on each use.
A model's life has two phases: training, where it learns from data once or periodically, and inference, where it runs to answer real requests. Every time a user of your app asks a question, an inference call happens on the model provider's servers or on yours.
In your app, inference determines ongoing cost and response speed. Larger models are often smarter but slower and more expensive per request, while smaller models are faster and cheaper and may be enough for simple tasks such as classification or short summaries.
Practical tip: do not use the most powerful model for everything. Split the work: a small, fast model for simple repetitive tasks and a stronger one only for complex tasks. Track the number and size of requests so you understand where the cost comes from.