Cost Management

Claude Code charges by API token consumption. Here’s how to track, manage, and reduce your costs.

Typical Costs (Enterprise Reference)

Average: ~$13 per developer per active day
Monthly: $150–250 per developer
90th percentile: under $30 per active day

These vary widely based on model choice, codebase size, and how many parallel sessions you run.

Track Your Costs

`/usage` command

/usage

Shows:

Token usage for current session
Dollar estimate (computed locally — check Claude Console for authoritative billing)
On Pro/Max/Team/Enterprise: breakdown by skills, subagents, MCP servers
Press d for last 24h, w for last 7 days

Status line (continuous)

Configure your status line to show context window usage at all times. See Keyboard Shortcuts & Status Line.

Claude Console (authoritative)

For authoritative billing: https://platform.claude.com/usage

For Teams (API users)

Set workspace spend limits at: https://platform.claude.com/workspaces

Rate Limit Recommendations for Teams

Team size	TPM per user	RPM per user
1–5	200k–300k	5–7
5–20	100k–150k	2.5–3.5
20–50	50k–75k	1.25–1.75
50–100	25k–35k	0.62–0.87
100–500	15k–20k	0.37–0.47
500+	10k–15k	0.25–0.35

TPM per user decreases with team size because fewer users use Claude concurrently in larger orgs.

Reduce Token Usage

1. Clear Between Tasks

/clear

Start fresh when switching to unrelated work. Stale context wastes tokens on every subsequent message.

Before clearing, rename the session so you can resume later:

/rename auth-module-debugging
/clear
# ... later ...
/resume auth-module-debugging

Custom compaction instructions:

/compact Focus on code samples and API usage

Or in CLAUDE.md:

# Compact instructions
When compacting, preserve test output and code changes.

2. Choose the Right Model

Model	Best for
Haiku	Simple subagent tasks, lookups, fast checks
Sonnet	Most coding tasks — best cost/quality balance
Opus	Complex architectural decisions, multi-step reasoning

Switch mid-session: /model Set default: /config For subagents: add model: haiku in the subagent frontmatter

3. Reduce MCP Server Overhead

MCP tool definitions are deferred by default (only names in context until used). But unused servers still add overhead.

/mcp         # See configured servers
/context     # See what's consuming context space

Prefer CLI tools when available (gh, aws, gcloud, sentry-cli) — they don’t add any context overhead vs MCP servers.

4. Move Workflow Instructions to Skills

Problem: CLAUDE.md is loaded into context at session start — even when you’re doing unrelated work.

Solution: Move specialized instructions (PR review checklists, deploy steps, migration patterns) into skills — they only load when invoked.

Rule of thumb: Keep CLAUDE.md under 200 lines. Only put always-needed context there.

5. Use Hooks to Preprocess Data

Instead of Claude reading a 10,000-line log file, a hook can grep for ERROR and pass only matching lines:

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

#!/bin/bash
input=$(cat)
cmd=$(echo "$input" | jq -r '.tool_input.command')

if [[ "$cmd" =~ ^(npm test|pytest|go test) ]]; then
  filtered_cmd="$cmd 2>&1 | grep -A 5 -E '(FAIL|ERROR|error:)' | head -100"
  echo "{\"hookSpecificOutput\":{\"hookEventName\":\"PreToolUse\",\"permissionDecision\":\"allow\",\"updatedInput\":{\"command\":\"$filtered_cmd\"}}}"
else
  echo "{}"
fi

6. Delegate Verbose Operations to Subagents

Running tests, fetching docs, or processing logs generates verbose output. Delegate these to subagents — the verbose output stays in the subagent’s context, and only a summary returns.

7. Write Specific Prompts

Vague: “improve this codebase” → broad scanning, many file reads
Specific: “add input validation to the login function in auth.ts” → minimal file reads

8. Use Plan Mode for Complex Tasks

Press Shift+Tab to enter plan mode. Claude explores first and proposes an approach for your approval — preventing expensive re-work when the initial direction is wrong.

9. Course-Correct Early

If Claude starts going in the wrong direction:

Press Escape to stop immediately
Press Escape twice (or /rewind) to restore to a previous checkpoint
Redirect before Claude does more work in the wrong direction

10. Adjust Extended Thinking

Extended thinking improves quality but costs more (thinking tokens billed as output tokens).

For simpler tasks:

/effort low          # Reduce thinking effort
/model               # Disable thinking in model settings

Or set environment variable: MAX_THINKING_TOKENS=8000

Note: Fable 5 always uses extended thinking — can’t be disabled.

11. Agent Team Cost Control

Agent teams use ~7x more tokens than standard sessions (each teammate has its own context window).

Tips:

Use Sonnet for teammates (not Opus)
Keep teams small
Keep spawn prompts focused (CLAUDE.md + MCP load automatically — don’t duplicate in prompts)
Shut down teammates when their work is done

Enable agent teams: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Background Token Usage

Even when idle, Claude Code uses a small amount of tokens:

Conversation summarization for /resume
Some command processing (/usage, etc.)

Typical background cost: under $0.04 per session — minimal.

How Prompt Caching Reduces Costs

Claude Code automatically uses prompt caching to reduce costs for repeated content:

System prompts
CLAUDE.md files
Tool definitions
Conversation history that hasn’t changed

You don’t need to do anything — it’s automatic. See how prompt caching works.

🧠 Quick Recall

Trick: the four levers — “cache, model mix, delegation, lean CLAUDE.md.”

👨‍🏫 Teach It

Show: /usage plus a statusline with cache counters.
They’ll trip on: blaming the model for cost when the cache prefix is churning (see file 54).

Learning path: ⬅ 40-agent-sdk.md · Index · ➡ 42-large-codebases.md

Written by Fenil Patel