Most teams discover their agent's cost after the first invoice. The number is rarely a surprise if you know the four places tokens are spent, because three of them are fixed before the user says anything.
The four consumers
**System prompt.** Fixed per request. Usually the smallest of the four and the one people over-optimize.
**Tool definitions.** Every connected server contributes a schema, charged on every request whether or not the model calls it. This is the one that surprises people: ten servers can consume a fifth of a small window before the conversation starts.
**Conversation history.** Grows linearly, and is re-sent each turn. A long thread is expensive in a way a long single message is not.
**Retrieved content and output.** The part you actually want, and usually the smallest share.
Measure before you optimize
Paste a representative prompt into the context budget checker and see what it consumes. Under 30% of the window, stop reading — you do not have a context problem. Over 70% before the conversation gets interesting, the toolset is the first suspect, not the model.
For the money rather than the window, the cost calculator takes tokens per request, requests per month, and price per million. It deliberately ships no model table: prices change quarterly and a stale default is worse than no default.
Three levers that actually work
**Remove servers you do not use.** List what you connected and ask, for each one, whether it was called this week. Disconnecting two is the highest-return change available and it takes a minute.
**Narrow the tools, not the prompt.** A server exposing forty tools costs more and chooses worse than one exposing four. This is why the directory labels each server with what it is for.
**Cut history deliberately.** Summarize older turns instead of resending them. The trap is summarising the wrong part: keep the decisions, drop the exploration.
The output premium
Output tokens typically cost three to five times input tokens. If your bill is higher than the calculator predicted, look at answer length before you look at prompt length — long answers are the expensive half.