Token usage higher than expected

Token usage often climbs quietly because the system keeps adding context, repeating instructions, or carrying too much history into each request. The surprising part is that the model may still look efficient while the bill steadily grows underneath.

It helps to inspect usage by request type instead of looking only at the total. Some flows are naturally expensive, but others may be wasting tokens because the prompt is bloated, the retrieval is noisy, or the response format is too verbose.

Once the biggest token drains are visible, teams can cut them with smaller prompts, tighter context windows, better summarization, and smarter routing. Small reductions at scale usually matter more than dramatic one-time optimizations.