Most of my AI token waste happens after the hard part is done

I kept assuming the expensive part of an AI workflow was the hard prompt.

It usually wasn't.

The bigger leak was what happened after I got unstuck:

• I stayed on the expensive model for cleanup work
• I kept old tool output in context even after the decision was made
• I let the session keep running while I was half coding and half chatting

That last 20 percent of the task was often burning tokens like the first 80 percent.

What changed for me was watching token usage live instead of checking a dashboard later.

A few habits came out of that:

1. Switch down once the hard reasoning step is over
   If the model has already found the bug or the plan, I do not need frontier-model pricing for renaming, formatting, or wrap-up work.

2. Restart when the chat becomes a scratchpad
   Once a session turns into pasted logs plus abandoned branches, it stops being context and starts being luggage.

3. Track tokens per finished task, not per day
   A daily total does not tell me much. A single bugfix that costs as much as a feature writeup does.

4. Treat context windows like budget, not capacity
   Just because the model can hold more does not mean the extra context is helping.

I built TokenBar for this because I wanted the number visible while I work, right in the macOS menu bar.

Curious how other people handle this:
Do you explicitly switch models mid-task, or do you mostly stay in one lane until the session ends?