Claude Code burns tokens fast for one core reason: it resends your entire conversation history every turn. The longer the context, the faster it drains — and that "it got dumber after 20 minutes" feeling is usually the same thing. Saving tokens = managing context. Ranked by payoff:
1. /clear between tasks, /compact at phase breaks
Don't let the last bug's debug logs sit in context forever. /clear for a new task, /compact when a phase is done.
2. Don't let tool output blow up the window
Reading a file or running tests dumps the whole result into context — a single test run can cost ~5,000 tokens even if you needed one line. Read targeted slices, not the whole repo.
3. Use smaller models for simple work
Haiku / Sonnet for routine edits instead of Opus — same task, 75%+ cheaper per token.
4. Use Plan mode before editing
Let it propose a plan you approve, then edit. Cuts the trial-and-error and repeated file reads.
5. Push stable facts into CLAUDE.md / skills
Project conventions, commands, style — write once in CLAUDE.md or a skill instead of re-explaining in natural language every turn.
6. Tighten your prompts
State the goal clearly in one shot. Fewer round-trips = shorter history = fewer tokens.
Combined, these cut typical usage 60-80% and roughly double a session's useful life.
Where the credits come from
When you're burning tokens anyway, a free relay station's signup credit stretches further than watching the official meter — you only change base_url and the key. List with models and tested dates: AI relay station list
Signup links carry my referral param; strip
?aff=if you prefer.
Top comments (0)