The cost per request is the average spend of a call, computed from input and output tokens and the model's price, or from the compute cost for a hosted model. It is tracked together with quality and latency to judge whether a solution remains viable at scale.
Example
average cost = total spend / number of requests