Home › Guides › Token usage

How to control Claude Code token usage

Updated: August 2026 · 6 min read

You're mid-task and the dreaded message shows up: you've hit your plan limit, come back in a few hours. If you use Claude Code seriously, sooner or later it happens. The good news: a big share of that usage is controllable once you understand what drives it.

How the limits work

On a subscription (Pro or Max), your usage is measured in tokens — chunks of text the model reads and writes. Everything counts: your request, every file Claude reads, every command it runs, its internal reasoning and the response. Plans renew the quota in windows of hours (plus weekly caps under heavy use): run out and you wait for the reset. So the goal isn't "use Claude less" — it's cutting the spend that buys you nothing.

What burns the most

• Giant context: huge projects where Claude re-reads dozens of files just to get oriented.
• Endless conversations: the longer the session, the more history gets re-processed on every turn.
• Vague requests: "improve the app" forces exploring everything; "fix the date bug in utils/date.js" goes straight there.
• Retrying failed tasks: three attempts at the same thing cost three times a well-formed request.

7 techniques that work

1. Be specific and give context. Name the file, paste the error message, say what you expected. Every fact you provide is exploration Claude doesn't spend.
2. One task, one conversation. Feature done? Close it and start a fresh session for the next thing. Dragging old history means paying to re-read it.
3. Use CLAUDE.md. A file at the project root with your key conventions and structure: Claude reads it on start and stops rediscovering the same things.
4. Pick the model per task. Renaming variables or writing a simple test doesn't need the most powerful model; save it for architecture and hard bugs.
5. Check the cost before approving long plans. If a task proposes touching 30 files, consider splitting it into stages.
6. Avoid pasting whole files into the chat when Claude can read them itself: reads from disk leverage caching; pasted text is new content re-processed every turn.
7. Measure your usage. What isn't measured can't be optimized: /cost in the CLI gives a snapshot of the moment, but won't tell you which request ate your plan yesterday.

The blind spot: real measurement

The terminal shows you the current session's spend and nothing else. No per-request history, no breakdown (was it the reading? the reasoning? the changes?), no "you're at 80% of your plan" before starting something big. Claude Station solves exactly that: a usage panel with each request's spend, its stage-by-stage breakdown, your plan percentage always visible — plus auto-resume: if the plan runs out mid-task, it resumes on its own when the quota renews.

Keep reading: What is Claude Code and how do you install it? · How to use Claude Code from your phone