The model wrote 23 tokens and billed 150 — reasoning tokens are output you never see