When I started using models with huge context windows, I assumed bigger meant easier.
What actually happened was I stopped noticing waste.
I would keep old specs, tool output, stack traces, and half dead prompts in the session because the model could still handle it.
The result was not just higher spend. It was slower decisions.
The pattern I watch now is simple: if the context window keeps growing but the task is not getting clearer, the session is probably getting worse.
A few habits that changed my workflow:
That is why I built TokenBar for macOS.
It keeps live token usage visible in the menu bar while I work.
For me, token counting is less about billing and more about knowing when my workflow is drifting.