Multi-Model Routing

The Key to Affordable AI

Claude comes in three tiers. Using the right model for each task is the difference between a $200/month bill and a $20/month bill.

To put real numbers on it: Haiku is roughly 60x cheaper per token than Opus. See anthropic.com/pricing for current per-token costs.

How Multi-Model Routing Works

OpenClaw can automatically pick the right model based on the type of request. You configure "profiles" that map to different models:

profiles:
  default:
    provider: anthropic
    model: claude-sonnet-4-6

  reasoning:
    provider: anthropic
    model: claude-opus-4-6

  quick:
    provider: anthropic
    model: claude-haiku-4-5

Recommended Routing Strategy

Here's a practical split that balances capability and cost:

Use Haiku (cheapest) for:

• Simple questions with straightforward answers
• Reformatting or summarizing short text
• Quick classifications ("is this spam?")
• Cron job health checks
• Translation of short phrases

Use Sonnet (default) for:

• General conversation
• Email drafting
• Research summaries
• Most daily assistant tasks
• Code review and simple coding
• Document analysis

Use Opus (most capable) for:

• Complex multi-step reasoning
• Long document analysis and synthesis
• Creative writing that needs nuance
• Architectural decisions for code
• Tasks where getting it wrong is costly
• Your daily briefing generation (worth the quality)

Setting It Up

Edit your OpenClaw config:

ssh claw@YOUR_SERVER_IP
nano ~/openclaw/config/default.yaml

Configure the profiles and set Sonnet as default. OpenClaw's routing can be triggered by keywords, channel, or explicit commands:

routing:
  default: sonnet
  rules:
    - match: "think deeply|analyze carefully|plan this"
      profile: reasoning
    - match: "quick|tldr|summarize briefly"
      profile: quick

Restart OpenClaw to pick up the changes:

docker restart openclaw

Testing the Routes

In Telegram, try sending messages that should trigger each model tier:

• "What's the weather like?" — should use Sonnet (default)
• "Think deeply about the pros and cons of moving to Portugal" — should use Opus
• "Quick: what's 15% of 340?" — should use Haiku

Verify which model actually handled each message by checking the logs:

docker logs openclaw --tail 50

Look for lines that mention the model name (e.g., claude-sonnet-4-6, claude-opus-4-6). You can also simply ask the bot: "What model are you?" — most models will identify themselves in the response.

Cost Expectations

With good routing:

• Light use (5–10 messages/day, mostly Sonnet): ~$5–10/month
• Moderate use (20–30 messages/day, mixed models): ~$15–30/month
• Heavy use (50+ messages/day, frequent Opus): ~$30–60/month

When You're Done

• Three model profiles configured (Opus, Sonnet, Haiku)
• Sonnet set as default
• Routing rules configured for model selection
• Tested each model tier via Telegram
• Verified correct model in logs