If I was being charged for the raw, output/input token count, excluding thinking/reasoning token costs, then sure. But at least via the API, you pay for tokens you cannot see.
There are features of input and output that are opaque to you, but that you pay for. Part of how model providers chose to run their service.