Cost & Performance

Understanding when to use which model—and how to minimize token usage—separates efficient Claude users from those burning through budgets.

Model Selection Guide

The Three Models

Model	Speed	Quality	Cost	Use When
Haiku	Fastest	Good	Lowest	Simple tasks, high volume
Sonnet	Medium	Great	Medium	Default for most work
Opus	Slowest	Best	Highest	Complex reasoning, architecture

Decision Framework

Is this task...

Simple/mechanical?
├─ Yes → Haiku
│   Examples: formatting, simple refactors, boilerplate
│
└─ No → Does it require deep reasoning?
    ├─ Yes → Opus
    │   Examples: architecture, complex debugging, security review
    │
    └─ No → Sonnet
        Examples: feature implementation, code review, most tasks

Switching Models

# Check current model
> /model

# Switch to Haiku for simple work
> /model claude-haiku-4-5

# Switch to Sonnet for standard work
> /model claude-sonnet-4-5

# Switch to Opus for complex work
> /model claude-opus-4-5

What Costs Tokens

Understanding token consumption helps you optimize:

Action	Token Cost	Notes
Your messages	Low	~1 token per 4 chars
Claude’s responses	Medium-High	Verbose by default
File reads	High	Full file content
MCP tool responses	Variable	Can be massive
Context accumulation	Compounds	Grows every message

The Hidden Costs

1. Every message includes full context - Message #50 is 50x more expensive than message #1
2. File reads stay in context - Reading 10 files = all 10 files in every subsequent request
3. MCP responses can be huge - Documentation tools often return 10k+ tokens

Cost Optimization Strategies

1. Start Fresh Frequently

# Bad: 100-message session
# Each message costs more as context grows

# Good: Multiple focused sessions
claude "implement the user model"
exit
claude "implement the API routes"
exit
claude "write the tests"

2. Use Subagents for Heavy Lifting

# Keep your context clean
> use a subagent to research authentication libraries
> just give me a summary of the recommendation

# Subagent burns tokens, your context stays lean

3. Compress Aggressively

# Before starting a new phase
> /compact

# If context is heavy
> /clear
> here's where we are: [summary]

4. Be Specific to Reduce Iterations

# Bad: Vague (causes back-and-forth)
> add authentication

# Good: Specific (one-shot)
> add JWT authentication
> - use python-jose
> - access tokens expire in 15 mins
> - refresh tokens expire in 7 days
> - store refresh tokens in Redis

5. Model Switching Mid-Session

# Start with Haiku for exploration
> /model claude-haiku-4-5
> explain the auth module

# Switch to Sonnet for implementation
> /model claude-sonnet-4-5
> implement the password reset flow

Speed Optimization

Factors Affecting Speed

Factor	Impact	Fix
Model choice	High	Use Haiku when possible
Context size	High	/compact frequently
MCP servers	Medium	Disable unused servers
Network	Variable	Check VPN/connection

Speed vs Quality Tradeoffs

# Maximum speed (simple tasks only)
> /model claude-haiku-4-5
> [task]

# Balanced (default for most work)
> /model claude-sonnet-4-5
> [task]

# Maximum quality (complex tasks)
> /model claude-opus-4-5
> [task]

Monitoring Usage

Track Spending

Use ccusage to monitor your usage:

# Check overall usage
npx ccusage

# Check last 7 days
npx ccusage --days 7

# Check this month
npx ccusage --month

Estimate Before Large Tasks

# Before a big task
> how many files will this touch?
> estimate: how complex is this?

# Then decide model appropriately

Anti-Patterns

❌ Opus for Everything

# Wasteful
> /model claude-opus-4-5
> fix this typo

❌ Never Compacting

# 200-message session with 50 file reads
# Every message now costs 100x what it should

❌ Massive MCP Responses

# Pulling entire docs into context
> get all React documentation
# 50k tokens just entered your context

❌ Redundant File Reads

# Reading the same file multiple times
> read src/main.py
> [work]
> read src/main.py again
# File is already in context!