Tokens per task is the AI cost metric I actually use
The first useful AI cost metric for me was not dollars. It was tokens per task.While building TokenBar, I noticed I was checking AI cost too late.
I would look at the bill after a long coding session and think, yeah, that got expensive. But that never told me where the workflow actually started going wrong.
What changed for me was watching token usage live per task.
A session that burns a lot of tokens can still be fine if progress is moving. The bad pattern is when token usage keeps climbing while the task stays fuzzy, the context gets messier, and I start retrying instead of tightening the ask.
loop:
- restart earlier when a session gets bloated
- stay on smaller models longer
- stop dragging old tool output into new tasks
- separate exploration from execution
That is a big part of why I built TokenBar for macOS.
It sits in the menu bar and makes token and cost usage visible while I work, which has been more useful for behavior change than any dashboard I check later.
