How to reduce Claude Code token usage (and stop hitting limits)

First, see where your tokens go

Why usage climbs in a long session

• Long context: every request carries the full conversation, and every round of tool calls sends another request. Caching makes re-reads cheaper, not free.
• Cache misses: your first message after a break longer than the cache lifetime reprocesses the whole context.
• Things that run while you're idle: session-scoped scheduled tasks and /loop, messages from your other sessions, goal check-ins, session recaps and idle compaction.
• Parallel workers: each subagent, workflow agent and agent-team teammate sends its own requests. Agent teams use about 7x the tokens of a standard session when teammates run in plan mode.
• Compaction itself: /compact has to read the conversation it summarizes. /clear costs nothing.

12 ways to reduce Claude Code token usage

1. Clear between unrelated tasks

   /clear starts a fresh conversation, so stale context stops riding along. Run /rename first if you want to /resume it later.

2. Compact with instructions

   /compact Focus on code samples and API usage tells Claude what to keep. You can also put standing compact instructions in CLAUDE.md.

   # Compact instructions
   
   When you are using compact, focus on test output and code changes.

3. Use the right model

   Sonnet handles most coding and costs less than Opus; keep Opus for architecture and hard reasoning. Switch with /model, and give simple subagents model: haiku.

4. Lower the effort for simple work

   Thinking tokens bill as output and can run to tens of thousands per request. Lower the level with /effort for routine tasks.

5. Keep CLAUDE.md short

   It loads in full every session. Stay under about 200 lines and move workflow-specific instructions into skills, which load only when used. See the CLAUDE.md guide and Claude Code skills.

6. Trim MCP servers

   Tool definitions are deferred by default, but every server still adds names and instructions. Disable unused ones in /mcp, and prefer CLIs such as gh, aws or gcloud, which add no per-tool listing.

7. Filter noisy output with a hook

   A PreToolUse hook can rewrite test commands to keep only failures, turning tens of thousands of tokens of logs into hundreds. Anthropic's docs include a ready-made filter script.

   {
     "hooks": {
       "PreToolUse": [
         {
           "matcher": "Bash",
           "hooks": [
             { "type": "command", "command": "~/.claude/hooks/filter-test-output.sh" }
           ]
         }
       ]
     }
   }

8. Delegate verbose work to subagents

   Running tests, reading docs or processing logs in a subagent keeps the output in its context; only a summary returns. Give those subagents a smaller model.

9. Install code intelligence plugins

   For typed languages, a language-server plugin answers "go to definition" in one call instead of a grep and several file reads.

10. Write specific prompts

    "Add input validation to the login function in auth.ts" reads a few files; "improve this codebase" scans everything.

11. Plan first, stop early

    Use plan mode for big changes so a wrong direction is caught before code is written; press Escape the moment Claude heads off course, and /rewind to a checkpoint.

12. Watch what runs while you're away

    Each scheduled task, /loop run or idle message sends your full context. Stop loops you don't need, set crossSessionInbound to hold to queue messages from other sessions, and check the Loops rows in /usage.

Source: Claude Code docs, Manage costs effectively

Usage limits and team costs

On Pro, Max and Ultra, usage counts against your plan's limits rather than a bill; see Claude Code vs Codex for how the 5-hour and weekly limits work. On the API, Anthropic reports an enterprise average of around $13 per developer per active day and $150 to $250 a month, with 90% of users under $30 a day.

Admins can set spend limits for teams, and a modelPricing table in managed settings makes /usage show contracted rates instead of list price.

Agents that work on a schedule, without burning tokens all night

Anything that wakes a session sends its whole context, so an always-on agent needs limits. Crewmark turns a Claude Code, Codex, DeepSeek Harness or OpenCode session into an AI employee that wakes on a heartbeat only during the working hours you set.

1. Set working hours and a rhythm

   Choose every 15, 30 or 60 minutes, and the hours it's on the clock; outside them it sleeps. See how to make Claude Code proactive.

2. One issue at a time

   Each heartbeat takes the next issue from the board, so the session works on a focused task instead of open-ended exploration.

3. On the plan you already pay for

   Employees run on your existing subscription, counted against your plan's limits. Read what an AI employee is.

Claude Code token usage FAQ