lockstep
5 min readNaman Jain

How to reduce Claude Code token usage: 12 practical ways

Ways to reduce Claude Code token usage, from Anthropic's docs: /usage, /clear, focused compaction, model and effort, MCP, hooks, skills, and subagents.

A long Claude Code session that started with one question can end up costing far more than the work justifies. The cause is usually not one big request. It's the whole conversation being re-sent with every message, plus background work nobody noticed.

This guide lists twelve ways to cut usage, from Anthropic's cost documentation and best practices, checked on 16 November 2026. It doesn't quote prices or plan limits, which change; check your own with /usage.

Why usage climbs

Claude Code sends your full conversation with every request, and each time Claude uses a tool, it sends another request with the results. Prompt caching re-reads that history at a discounted rate while the cache is warm. Usage grows faster than activity because of:

  • Long context: every message re-sends everything before it.
  • Cache misses: the first message after a break longer than the cache lifetime reprocesses the whole context.
  • Background work: scheduled tasks, goal check-ins, and messages from other sessions start new turns while you're idle.
  • Subagents and agent teams: each sends its own requests.
  • Compaction: /compact reads the whole conversation to summarize it.

1. Start with /usage

On Pro, Max, Ultra, Team, and Enterprise plans, /usage breaks usage down by skill, subagent, plugin, and MCP server, and recommends a change for any behavior behind 10% or more of recent usage. /context shows what is in the window right now. A status line can keep the context percentage on screen all the time.

2. Clear between unrelated tasks

/clear costs nothing. Old context re-sent on every message of a new task is the most common waste. Use /rename first if you want to come back to the old session with /resume.

3. Compact with a focus

When you need continuity, /compact focus on the API changes and the test commands controls what the summary keeps. You can also put compaction guidance in CLAUDE.md under a "Compact instructions" heading. Remember that compacting a large context is itself a large request.

4. Match the model to the task

Anthropic's cost guide says Sonnet handles most coding tasks well and costs less than Opus, and suggests saving Opus for complex architecture and multi-step reasoning. Switch with /model. Subagents that inherit your session's model inherit that choice too.

5. Lower effort for simple work

Thinking tokens are billed as output. For simple tasks, lower the effort level with /effort. Some current models always use extended thinking; for those, effort is the control.

6. Prefer CLI tools over MCP where they exist

Claude Code already defers MCP tool definitions, so only tool names and server instructions load until a tool is used. The docs still note that CLI tools such as gh, aws, and gcloud are more context-efficient because they add no tool listing. Disable servers you aren't using with /mcp. See Claude Code MCP setup.

7. Filter verbose output with a hook

A hook can cut output before Claude reads it. Anthropic's example is a PreToolUse hook that rewrites test commands to keep only the last 100 lines of output, turning tens of thousands of tokens into hundreds. See Claude Code hooks: 6 examples.

8. Move procedures from CLAUDE.md into skills

CLAUDE.md loads in full every session. Detailed procedures for PR reviews or migrations cost tokens even when you're doing something else. Skills load only when used. See CLAUDE.md examples and Claude Code skills: 6 examples.

A skill that describes your architecture, key directories, and naming conventions saves Claude reading many files to work them out. For typed languages, code intelligence plugins replace grep-and-read with precise go-to-definition.

10. Send noisy work to subagents, on a smaller model

Running tests, fetching docs, or reading logs can flood the main conversation. A subagent keeps that output in its own context and returns a summary. Its requests still count, so give simple subagents a smaller model. See Claude Code subagents: 5 examples.

11. Use agent teams sparingly

Anthropic's docs say agent teams use roughly 7 times the tokens of a standard session when teammates run in plan mode. Keep team tasks small and self-contained, and use a team only when parallel work pays for it. See Claude Code agent teams.

12. Write specific prompts, and stop early

"Improve this codebase" makes Claude scan broadly. "Add input validation to the login function in auth.ts" doesn't. For bigger tasks, plan first, press Esc the moment Claude drifts, and give it a test to pass so it doesn't need several rounds of corrections. More in Claude Code best practices.

Watch background usage

Some usage happens while you're away: session recaps, idle compaction, prompt suggestions, scheduled tasks, and goal check-ins. The docs explain when each runs and how to turn off the ones you don't need.

The tokens spent re-explaining

One cost doesn't show up in /usage as its own line: the tokens spent re-explaining the same team decisions in every new session, and the rework when an agent builds against one it never heard about.

Where Lockstep fits

Lockstep briefs each Claude Code session on the team's approved decisions that apply to the work, so they don't need to be pasted into prompts or kept in an ever-growing CLAUDE.md. Decisions are captured from PRDs, Slack, and pull requests and become binding when their owner approves them; the briefing is ranked so the most important ones come first.

See how it works.

Sources checked 16 November 2026

For what loads into context and when, see context engineering for coding agents and Claude Code. For how memory loads, see how Claude Code memory works. For what fills the window and survives compaction, see Claude Code context window. For how caching works under the hood, see prompt caching explained.

Frequently asked questions

Why does Claude Code use so many tokens?
Claude Code sends the full conversation with every request, and each tool use sends another request carrying the results. Prompt caching makes repeated history cheaper, but long sessions, cache misses after a break, background tasks, subagents, and compaction all add usage.
How do I see what is using my Claude Code tokens?
Run /usage. On Pro, Max, Ultra, Team, and Enterprise plans it breaks usage down by skill, subagent, plugin, and MCP server, and recommends a change for any behavior behind 10% or more of recent usage. /context shows what is in the context window right now.
Does /clear or /compact save more tokens?
/clear costs nothing and starts fresh, so use it when switching to unrelated work. /compact keeps continuity but reads the whole conversation to summarize it, so on a large context it is itself a big request. Add a focus, such as /compact focus on test output, to control what it keeps.
Do MCP servers use a lot of tokens in Claude Code?
Less than before. Claude Code defers MCP tool definitions, so only tool names and server instructions load until a tool is used. CLI tools such as gh or aws are still more context-efficient, and you can disable servers you aren't using with /mcp.
Does using subagents save tokens?
They keep verbose output out of your main conversation, but their own requests still count toward your usage. Pick a smaller model for simple subagent work. Agent teams use roughly 7 times the tokens of a standard session when teammates run in plan mode, according to Anthropic's docs.

Keep reading