Claude Code Token Economy: Understanding Costs

Claude Code charges based on token usage. Understanding how tokens work helps you predict costs and use Claude Code efficiently.

What Are Tokens?

Tokens are pieces of text. Roughly:

1 token = 4 characters in English
100 tokens = about 75 words
1,000 tokens = about 750 words

Code is less predictable:

Short variable names = fewer tokens
Long class names = more tokens
Comments add tokens too

How Costs Are Calculated

You pay for:

Input tokens - What you send (prompts, file contents)
Output tokens - What Claude generates (responses, code)

Pricing varies by model (approximate as of 2025):

Model	Input	Output
Haiku	$0.25/1M tokens	$1.25/1M tokens
Sonnet	$3/1M tokens	$15/1M tokens
Opus	$15/1M tokens	$75/1M tokens

What Consumes Tokens

System prompt

Every request includes a system prompt (~2,000-5,000 tokens). This is sent with every interaction.

Conversation history

As your session grows, more history is sent:

Start of session: ~5k tokens
After 30 minutes: ~50k tokens
After 2 hours: ~150k+ tokens

File reading

Reading files adds tokens:

Small file (100 lines): ~500 tokens
Medium file (500 lines): ~2,500 tokens
Large file (2000 lines): ~10,000 tokens

Tool calls

Each tool call (read, write, bash) adds overhead:

Tool definition: ~100 tokens
Tool input: varies
Tool output: varies

Estimating Session Costs

Quick estimate

A typical development session:

Light (30 min, simple tasks): $0.10-0.50
Medium (1-2 hours, moderate tasks): $0.50-2.00
Heavy (3+ hours, complex tasks): $2.00-10.00

Detailed calculation

Example session:

System prompt:           3,000 tokens
Reading 5 files:        15,000 tokens
Conversation (1 hour):  30,000 tokens
Code generation:        10,000 tokens
───────────────────────────────────────
Total input:           ~58,000 tokens
Total output:          ~10,000 tokens

Cost (Sonnet):
  Input:  58k × $3/1M = $0.17
  Output: 10k × $15/1M = $0.15
  Total: ~$0.32

Monitoring Usage

In-session monitoring

Claude Code shows token usage in the status bar:

[tokens: 45k/200k]

Watch this as you work.

API dashboard

Check console.anthropic.com for:

Daily usage
Cost breakdown
Usage trends

Set up alerts

Configure alerts before hitting budget limits.

Cost Optimization Strategies

1. Start fresh sessions frequently

Long sessions accumulate history. Fresh sessions reset:

# Instead of continuing a 2-hour session
# Start fresh for new tasks
claude

2. Use /compact

Reduce context without losing essential info:

/compact

3. Choose the right model

Task	Model	Why
Quick questions	Haiku	12x cheaper than Sonnet
Code generation	Sonnet	Good balance
Complex architecture	Opus	Better reasoning

Switch with:

/model haiku

4. Be specific in prompts

Vague prompts lead to more back-and-forth:

# Expensive: Multiple rounds needed
Fix the bug

# Cheaper: One-shot completion
Fix the null pointer error on line 45 of auth.ts
by adding a null check before accessing user.email

5. Read files selectively

# Expensive: Full file read
Read utils.ts

# Cheaper: Targeted read
Read the validateEmail function in utils.ts

# Expensive: 5 separate sessions
Session 1: Fix typo in file1
Session 2: Fix typo in file2
...

# Cheaper: 1 session
Fix all typos in the auth module

Daily/Monthly Budgeting

Solo developer estimates

Usage Level	Daily	Monthly
Light	$0.50	$15
Moderate	$2	$60
Heavy	$5	$150

Team estimates

Multiply by team size, but expect overlap savings from shared context in CLAUDE.md.

Hidden Token Costs

Repeated context

Every message resends conversation history. A 50-message conversation resends earlier messages 50 times.

Tool overhead

Each tool call adds fixed overhead. Many small reads cost more than one large read.

Failed attempts

When Claude tries something that fails, you still pay for the attempt.

Tracking ROI

Consider time saved:

Developer hourly rate: $75
Time saved per day with Claude: 2 hours
Value created: $150/day

Claude Code cost: $5/day
ROI: 30x

Even heavy usage often pays for itself quickly.

Quick Reference

Question	Answer
1 token =	~4 characters
Reading 500 line file =	~2,500 tokens
1 hour session =	$0.50-2.00 (Sonnet)
Cheapest model	Haiku (12x cheaper)
Most capable	Opus (5x more expensive)
Reduce costs	/compact, fresh sessions, specific prompts