Usage grows with the length of the conversation, the model and the effort level. Everything below cuts one of those three. Run /usage before and after to see what changed. (Anthropic: manage costs)
1. Clear between tasks
/clear starts fresh, so the next task doesn't carry the last one's history. Use /compact instead when you're continuing the same task. To come back to a conversation later, name it with /rename first and return with /resume.
2. Compact at natural breaks
/compact replaces the conversation with a summary. Tell it what to keep, as in /compact focus on the auth bug fix, or put standing instructions in CLAUDE.md under a "Compact instructions" heading. Compacting is itself a big request, so do it when you finish a step, not in the middle of one, and before you step away rather than after. To drop only the last few turns, /rewind keeps the rest.
3. Use Sonnet for most work
Opus uses your limit faster than Sonnet (its API price per token is twice Sonnet's), and it's Claude Code's default model on paid plans. Sonnet handles most coding well. Switch with /model, or set your default in /config. /model opusplan plans with Opus and writes the code with Sonnet.
4. Lower the effort level
Thinking is billed as output, and a lower effort level means less of it. Set it with /effort (low, medium, high, xhigh or max) or in /model. Choose the model at the start of a session: switching midway means the next request can't use the cache and sends the whole conversation again. On the 5.5 models you can change effort without losing the cache.
5. Keep what's always loaded small
- Keep CLAUDE.md under about 200 lines. Move instructions you only need sometimes into skills, which load only when used.
- Run
/contextto see what's taking up room. - Run
/mcpand turn off servers you aren't using. Where a command-line tool such asghdoes the job, Anthropic suggests it over an MCP server.
6. Keep noisy output out of the conversation
Test runs, logs and long command output fill the conversation fast. Hand that work to a subagent, so only its summary comes back, and for simple subagent tasks set model: haiku in its settings, since subagents draw on your limit too. Agent teams can use about 7x the tokens of a normal session when teammates run in plan mode, so keep them small.
7. Let the cache work
Claude Code caches the conversation so each turn doesn't pay for it in full. On a subscription within your plan's limits, the main conversation's cache lasts an hour and subagents' five minutes; with an API key or on usage credits, five minutes. Switching model and compacting start it over, and so does changing effort on older models. Resuming an old session does too, since its cache is long gone. From v2.1.251, /usage shows a "Prompt cache (main)" line, and from v2.1.242 the promptCacheTtl setting picks 5m or 1h. (Anthropic: prompt caching)
8. Watch what runs while you're away
A session left open can keep spending: scheduled /loop tasks and session recaps use tokens while you're away. Start a fresh session for recurring work. Idle compaction also runs while you're away, but it makes your first message back cheaper, so turn it off with idleCompaction only if you'd rather keep the full conversation.
When the waste is a bug
Sometimes it's not you. These are live on the Claude Code RAP Sheet, each with its reports and any known workaround:
- Prompt cache misses and re-writes: the cache drops mid-session and the whole conversation is written again, sometimes for many turns in a row.
- Auto-compaction loses session context: compaction fires early or without warning, and the session forgets what it was doing.
- Unrequested agent fan-outs drain usage: Claude starts many subagents for a small task, with no estimate or cap.
- Usage token totals inflated: the Stats tab and the desktop usage card count each call about twice, so local totals look higher than what you used.
- Cut per-request context overhead: a request from people who want each call to carry only what the work needs.
More guides
The tips come from Anthropic's own pages, manage costs, prompt caching, context window and models, usage and limits, checked Nov 2, 2026. Commands change often; when this page and theirs differ, theirs is right.