Blog

How to reduce Claude Code token usage without losing output quality

Token Coach · 2026-10-21

Short answer: Most Claude Code token spend comes from context, not from your prompt. Keep threads short, keep CLAUDE.md lean, match the model to the task, and let a coach tell you which technique is worth it for the task in front of you.

Where the tokens actually go

Every time you press enter, Claude Code sends the conversation so far, your CLAUDE.md, the tool definitions it has loaded, and your new message. Your prompt is usually the smallest part. That is why a two-hour thread costs far more per turn than a fresh one, even when you type the same sentence.

1. Start fresh more often

A long thread re-sends everything before it. When a piece of work is done, write a short handoff note (what was decided, what is next, which files matter) and start a new thread with /clear. You keep the state and drop the bill.

2. Keep CLAUDE.md lean

CLAUDE.md is loaded into every session. Long style guides, pasted docs and old decisions ride along on every turn. Keep it to rules Claude must follow, and move reference material into files Claude can read when it needs them.

3. Match the model to the task

Renaming variables does not need the strongest model. Use the lightest model that does the job well, and save the heavy one for design, debugging and anything with real ambiguity.

4. Let prompt caching work for you

Claude Code caches the repeated start of your context automatically, and cached tokens cost a fraction of fresh ones. Caches expire after a short idle period, so long pauses mid-thread and constantly editing CLAUDE.md both make you pay full price again. Steady work in one thread keeps the cache warm.

5. Send wide searches to a subagent

Exploring a codebase pulls a lot of file content into context. A subagent does that reading in its own context and returns a short summary, so your main thread stays small.

6. Trim tools you are not using

Every connected MCP server adds its tool descriptions to context. Disconnect the ones you are not using in this project.

7. Skip ceremony on small tasks

Plans, reviews and test suites are worth it on big, risky work. On a one-line change they are pure overhead. The hard part is knowing which is which in the moment.

Let CTO pick for you

CTO is a free token coach for Claude Code. Before a task it tells you which of these techniques is worth it at that size and which to skip, warns when a thread is getting expensive, offers the handoff note, and grades its own advice against your later sessions. It runs on your machine and stays quiet on small tasks.