The hidden cost of AI coding is context you forgot to remove
One thing I keep noticing in AI coding sessions: the expensive part is not always the hard prompt.
It is the old context that keeps riding along after it stops being useful.
A few examples:
- terminal logs from a bug that is already fixed
- files pasted in because they were useful 30 minutes ago
- summaries that compress confusion instead of clarity
- retry loops where the model is solving yesterday's version of the task
- staying on a bigger model because restarting feels slower
The problem is that this waste is invisible while you are working. By the time you check a dashboard or invoice, the behavior already happened.
That is why I built TokenBar as a tiny macOS menu bar token counter. I wanted the AI session to have something closer to a speedometer.
Not because every token is bad. Tokens are useful when they move the task forward.
But when the count keeps climbing and the answer is not getting clearer, that is usually a workflow signal:
- restart with a cleaner brief
- trim stale context
- switch back to a cheaper model
- stop asking the same fuzzy question in five different ways
The useful metric is not just total spend. It is tokens per useful step.
If you build with AI every day, live token visibility changes the moment you decide to reset the session.
TokenBar: https://tokenbar.site/
