Module 9: From Toy to System•Lesson 4 of 6

Cost Control

AI agents can get expensive. Here's how to manage costs.

Understanding Costs

API Costs

Input tokens: What you send (prompts, context)
Output tokens: What the model generates
Tool calls: Often billed separately

Cost by Model (per 1M tokens)

Model	Input	Output
Claude Opus	$15	$75
Claude Sonnet	$3	$15
Claude Haiku	$0.25	$1.25
GPT-4o	$2.50	$10
GPT-4o-mini	$0.15	$0.60

Cost Reduction Strategies

1. Right-Size Your Models

Don't use Opus for everything:

# Coordinator: Smart (but expensive)
default: claude-opus-4

# Background tasks: Good enough (cheaper)
cron:
  model: claude-sonnet-4

# Simple tasks: Fast (very cheap)
simple:
  model: claude-haiku

Potential savings: 60-80%

2. Minimize Context

Every token counts:

# Bad: Loading everything
Load: SOUL.md, USER.md, MEMORY.md, all daily logs, 
      all project files, all people files...

# Good: Load on demand
Always: SOUL.md, USER.md, MEMORY.md (index only)
On demand: Specific detail files when needed

3. Cache When Possible

If you're doing the same lookups:

Store results in files
Check cache before calling APIs
Set reasonable TTLs

4. Set Budgets

Implement spending limits:

# Example budget config
limits:
  daily: $10
  monthly: $200
  perSession: $1

Get alerts before hitting limits.

5. Monitor Usage

Track where your money goes:

Which sessions cost most?
What tasks are expensive?
Are there runaway processes?

Cost Monitoring Example

My setup tracking:

📊 Daily Usage Report
- Main sessions: $2.40 (Opus)
- Cron jobs: $0.80 (Sonnet)
- Sub-agents: $0.30 (Sonnet)
Total: $3.50/day

Monthly estimate: ~$100

Red Flags

Watch for:

Sudden cost spikes
Infinite loops (keeps calling APIs)
Unnecessary tool calls
Oversized contexts
Wrong model for task

Budget Template

Category	Model	Daily Budget
Main chat	Opus	$3
Cron jobs	Sonnet	$1
Research	Sonnet	$1
Quick tasks	Haiku	$0.50
Total		$5.50/day

Adjust based on your actual usage patterns.

Previous: Security Basics Next: Next Paths