Price the clock, not the model

A chat costs you once. A background agent costs you every time it acts — every scheduled run, every loop it takes pursuing a goal. And most of that work is checking: did anything change, is there anything to do? Most of the time the answer is "no" — and you paid for the looking.

That's why the biggest lever on cost isn't the model, it's how often the agent works:

• On Grok Bot, this is literal: a routine on the 5-minute floor wakes ~8,640 times a month; the same routine hourly wakes 720 times — about 12x less work for usually a sliver less freshness. (Grok Bot caps you at 50 routines per Bot and a 5-minute minimum for a reason.)
• On Muse and Dots, you don't set a cron, but a goal that says "keep watching X continuously" burns usage the same way. A goal scoped to "check once a day and tell me" costs a fraction of "monitor this in real time."

All three bill against a plan allowance or usage pool rather than a per-task invoice, so you won't see a line item for a wasteful agent — you'll see your allowance drain faster, or heavier tiers become "necessary." The fix is the same everywhere: start with the least frequent / least continuous version of the task that still does the job, and only tighten it if a real miss proves you need to.

A quick gut check before you set any recurring or always-on task: what would I actually do with fresher information that I can't do on a slower cadence? If the honest answer is "nothing," you've found your setting. "Just in case" is the most expensive phrase in agent-running — because "just in case" gets paid for thousands of times a month.