What You'll Build

A model routing strategy that uses cheap models for simple tasks and expensive models only when needed. Your AI agent stays just as capable but costs 90-97% less.

Why It Works

Most people set their AI agent to use Claude Opus (the most powerful, most expensive model) for everything. Every heartbeat check, every simple question, every status ping costs premium tokens. That's like hiring a brain surgeon to take your temperature.

Results: Monthly bill dropped from $90 to $6. Same agent. Same capabilities. 97% cost reduction.

Prerequisites

The Problem

Here's what's burning your money:

Task Frequency Tokens Per Call Model Used Monthly Cost
Heartbeat check Every 30 min 5,000-15,000 Opus $30-45
Simple replies 20/day 2,000-5,000 Opus $15-25
Complex tasks 2-3/day 10,000-30,000 Opus $10-15
Background jobs 5-10/day 5,000-15,000 Opus $10-20
Total $65-105

The heartbeat alone (a simple "anything new?" check) eats 30-50% of your budget.

The Solution: Tiered Model Routing

┌─────────────────────────────────────────────┐
│              Task Incoming                   │
│                                              │
│  Simple? ──────▶ Haiku ($0.25/M tokens)     │
│  (heartbeats,    - Heartbeat checks          │
│   status,        - Yes/no questions          │
│   basic)         - Simple lookups            │
│                                              │
│  Medium? ──────▶ Sonnet ($3/M tokens)       │
│  (writing,       - Email drafts              │
│   analysis,      - Content creation          │
│   planning)      - Research summaries        │
│                                              │
│  Complex? ─────▶ Opus ($15/M tokens)        │
│  (strategy,      - Complex analysis          │
│   coding,        - Multi-step reasoning      │
│   critical)      - Code generation           │
│                                              │
│  Local? ───────▶ Ollama (FREE)              │
│  (private,       - Personal data processing  │
│   offline,       - Simple classification     │
│   simple)        - Drafts nobody sees        │
└─────────────────────────────────────────────┘

Step-by-Step Setup

Step 1: Set Your Default Model to Sonnet

In your OpenClaw config, change the default model from Opus to Sonnet:

{
  "model": "anthropic/claude-sonnet-4-20250514"
}

This single change cuts your baseline costs by ~80%. Sonnet handles 90% of daily tasks perfectly well.

Step 2: Configure Heartbeats to Use Haiku

Heartbeats are the biggest cost driver. They run every 30 minutes and usually just check if anything needs attention. Haiku is perfect for this.

In your OpenClaw config, set the heartbeat model:

{
  "heartbeat": {
    "model": "anthropic/claude-haiku"
  }
}

Cost impact: Heartbeat cost drops from $30-45/month to under $1.

Step 3: Use Opus Only for Complex Tasks

Reserve Opus for tasks that genuinely need it:

You can override the model per-session or per-task:

/model opus

Or set specific cron jobs to use Opus:

{
  "payload": {
    "kind": "agentTurn",
    "message": "Complex overnight task...",
    "model": "anthropic/claude-opus-4-6"
  }
}

Step 4: Add Local Ollama (Optional, Saves More)

For tasks that don't need cloud AI at all, run a local model:

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Pull a small, fast model
ollama pull llama3.2:3b

Use cases for local models:

Cost: $0. Runs on your hardware.

Step 5: Monitor and Adjust

After a week, check your usage:

/status

Look at:

The Math

Task Before (All Opus) After (Routed)
Heartbeats (48/day) $35/month $0.50/month
Simple tasks (20/day) $20/month $2/month
Medium tasks (10/day) $20/month $3/month
Complex tasks (2/day) $15/month $3/month
Total $90/month $8.50/month

That's a $81.50/month savings. $978/year. For changing a config file.

Customization Ideas

Gotchas & Tips