By how much? At least TFA provided numbers for one example, and they disagree with you (by a lot).

I ran a fairly large experiment last week, and the token usage wasn't bad at all. What softs of use cases are you seeing large token usage by GLM 5.2?