Tracking Costs

Claude Code uses tokens, and tokens cost money. Here’s how to track and optimize your usage.

Understanding Tokens

Tokens are the units Claude uses to process text:

• ~4 characters = 1 token (English)
• ~100 tokens = 75 words
• A typical code file = 500-2000 tokens

Cost factors:

• Input tokens (what you send to Claude)
• Output tokens (what Claude generates)
• Context (accumulated conversation history)

Using CC Usage

The best way to track costs is ccusage:

1. Install ccusage

   npx ccusage

   Or install globally:

   npm install -g ccusage

2. View your usage

   npx ccusage

   This shows:

   • Total tokens used today
   • Estimated cost
   • Breakdown by session

3. Check specific time periods

   npx ccusage --days 7     # Last 7 days
   npx ccusage --month      # This month

Sample Output

╭─────────────────────────────────────────────────────╮
│ Claude Code Usage Report                            │
├─────────────────────────────────────────────────────┤
│ Period: Today (Jan 5, 2026)                         │
│                                                     │
│ Sessions: 12                                        │
│ Input tokens: 145,230                               │
│ Output tokens: 42,891                               │
│ Total tokens: 188,121                               │
│                                                     │
│ Estimated cost: $2.34                               │
╰─────────────────────────────────────────────────────╯

Token Usage by Activity

Different activities use different amounts of tokens:

Activity	Typical Token Usage
Simple question	500-1,000
Explain a file	2,000-5,000
Bug fix	5,000-15,000
Feature implementation	20,000-50,000
Full project scan	50,000-200,000

Caution

Long conversations accumulate context. A 30-message conversation might use 100k+ tokens just in context, even if your messages are short.

Optimizing Token Usage

1. Start Fresh Sessions

Don’t continue conversations forever. Start new sessions for new tasks:

# New task? New session.
claude "new task: add user authentication"

2. Use /compact

When context grows large, compress it:

> /compact

This summarizes the conversation, preserving important context while reducing tokens.

3. Be Specific

Vague prompts require more back-and-forth:

Expensive:

> fix the bug
> no, the other bug
> in the auth module
> the one about passwords

Cheaper:

> fix the password validation bug in auth.py line 45

4. Use the Right Model

Model	Cost	Use For
Haiku	$	Quick questions, simple tasks
Sonnet	$$	Daily coding, most tasks
Opus	$$$$	Complex architecture, hard problems

Switch models based on task complexity:

> /model haiku
> what's the syntax for a Python list comprehension?

> /model sonnet
> refactor this authentication system

5. Limit File Scanning

Instead of:

> explain all the code in this project

Try:

> explain the main entry point in app.py

6. Use Print Mode for Quick Queries

# Interactive mode loads context
claude
> quick question

# Print mode is lighter
claude --print "quick question"

Setting Budget Alerts

Track spending with a simple script:

#!/usr/bin/env python3
"""Check Claude Code spending and alert if over budget."""
import subprocess
import json

DAILY_BUDGET = 10.00  # dollars

def check_usage():
    result = subprocess.run(
        ["npx", "ccusage", "--json"],
        capture_output=True,
        text=True
    )
    data = json.loads(result.stdout)
    cost = data.get("estimated_cost", 0)

    if cost > DAILY_BUDGET:
        print(f"⚠️  Over budget! ${cost:.2f} / ${DAILY_BUDGET:.2f}")
    else:
        print(f"✓ ${cost:.2f} / ${DAILY_BUDGET:.2f}")

if __name__ == "__main__":
    check_usage()

Context Window Limits

Claude has a maximum context window. When you hit it:

1. Oldest messages get dropped
2. You lose important context
3. Claude may forget earlier instructions

Signs you’re hitting limits:

• Claude forgets what you discussed earlier
• Responses become less coherent
• You see warnings about context length

Solution: Use /compact proactively or start a new session.