Save tokens in Claude Code: 10 ways to make your allowance last the day
Why Claude Code burns tokens and 10 ways to cut it: /clear, /compact, the right model, /effort, a short CLAUDE.md, Haiku subagents, usage per key.
Updated 25 Oct 2026
A RocketKey plan gives you a daily USD allowance that resets at 00:00 (GMT+7). Once it is used up, requests get error 402 until the next day. The quickest way to make it last is to know where Claude Code spends tokens, then cut the waste.
Why does Claude Code use tokens so fast?
Each time Claude replies or uses a tool (reads a file, runs a command), Claude Code sends the whole conversation of the session to the model. A session open since morning drags all of that history along, even when your last question is one line. Three things drain the allowance most:
- Long context: one session for unrelated tasks, big files read in, long logs pasted.
- An expensive model for easy work: Opus or Fable where Sonnet or Haiku would do.
- Thinking tokens: the 5.5 models always reason before answering, and that counts as output.
The tips below follow Claude Code's cost documentation and apply to RocketKey keys.
1. /clear when you switch tasks
When a task is done, type /clear to start fresh, so the old context stops riding along with every request. To come back later, name the session with /rename before clearing and reopen it with /resume.
2. /compact with instructions when a session gets long
To keep the thread of a long session, run /compact with what to keep, for example /compact keep the code changes and test results. Compacting reads the whole conversation to summarize it, so it is a large request itself. If you do not need continuity, /clear is cheaper.
3. Match the model to the task
Sonnet 5.5 handles most coding at half the price of Opus 5.5. Type /model to switch to Opus for the hard part and back to Sonnet after. The comparison is in Choose a Claude model for coding.
4. Lower reasoning with /effort
For simple work, lower the effort with /effort (or in /model). Per Claude Code's documentation, thinking cannot be turned off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models, so lowering effort is how you cut thinking tokens there.
5. Keep CLAUDE.md short
CLAUDE.md is loaded into every session. Anthropic suggests keeping it under about 200 lines with only the essentials. Move instructions for specific workflows (PR review, migrations…) into skills, which load only when used.
6. Turn off MCP servers you do not use
Run /context to see what fills the context and /mcp to disable unused servers. When a CLI does the same job (such as gh for GitHub), it usually takes less context than an MCP server.
7. Send verbose work to Haiku subagents
Running tests, reading long logs and fetching docs produce lots of output. Hand them to a subagent so that output stays in the subagent's context and the main session gets a summary. Set model: haiku in the subagent's configuration for simple jobs: Haiku 5.5 is about 20 times cheaper than Sonnet 5.5.
On the other hand, Anthropic notes that agent teams use about 7 times the tokens of a normal session, so turn them on only when you need them.
8. Write specific prompts
"Improve this codebase" makes Claude scan many files. "Add input validation to the login function in auth.ts" sends it straight to the right place. For big tasks, press Shift+Tab into plan mode so Claude proposes an approach before editing. When Claude heads the wrong way, press Esc to stop at once, and use /rewind to go back to an earlier point.
9. Filter output before Claude reads it
A 10,000-line log can cost tens of thousands of tokens. Filter it first with grep (or a Claude Code hook) so Claude sees only the error lines.
10. Track usage to see where the money goes
- In Claude Code,
/usageestimates the tokens and cost of the current session. It uses Anthropic's list prices, so treat it as a guide. - What RocketKey actually charges is under Dashboard → Usage, updated every 2 minutes. The Breakdown section has By model and By key views, showing which model and which machine spend the most. See Manage API keys.