Hey @AndrewC, thanks for sharing the breakdown and the before/after screenshots.
The bot’s explanation is right. The Tokens number on each row counts everything the model read on that call, including the bot’s instructions, memory, tool list and recent chat. Most of that is cached context, which is billed at a much lower rate than new input. If you hover over a token count on the usage page, you’ll see it split into input, output, cache read and cache write.
All of these rows are counted against your plan’s included usage, so nothing extra has been charged.
The changes you made are the right levers: fewer event-triggered wake-ups, fewer bot-to-bot messages, and a leaner background for each bot. Your second screenshot already shows per-call size dropping to roughly a third. Separately, the composer-2.5 rows are full Cloud Agent coding jobs, so they will always be much larger than a single bot reply.