Model Routing: Right Model for the Job
3 min read
If you're running OpenClaw, you're probably overpaying. By a lot.
By default, everything goes to your primary model. Heartbeats? Opus. Calendar lookups? Opus. Sub-agents? All Opus. That's like hiring a lawyer to check your mailbox.
The Hidden Cost Multiplier
Heartbeats fire 48 times per day. Sub-agents spawn for parallel work. Simple queries hit the same $30/M-token model you use for complex coding.
Result: premium prices for simple tasks, and no resilience when rate limits hit.
The Fix: Model Tiering
- Complex reasoning — architecture, refactoring → frontier model (Opus, GPT-5.2). Worth it.
- Daily work — code gen, research → mid-tier (Sonnet, DeepSeek R1). 90% cheaper.
- Simple tasks — heartbeats, lookups → cheapest that works. Gemini Flash-Lite is 60x cheaper than Opus.
One Config Change
Before (everything goes to Opus):
{
"agents": {
"defaults": {
"model": "anthropic/claude-opus-4-5"
}
}
}After (smart routing):
{
"agents": {
"defaults": {
"model": {
"primary": "anthropic/claude-opus-4-5",
"fallbacks": [
"openai/gpt-5.2",
"deepseek/deepseek-reasoner"
]
},
"heartbeat": {
"every": "30m",
"model": "google/gemini-2.5-flash-lite"
},
"subagents": {
"model": "deepseek/deepseek-reasoner"
}
}
}
}Heartbeats → Gemini Flash-Lite at $0.50/M instead of $30/M. Sub-agents → DeepSeek R1 at $2.74/M — 10x cheaper with solid reasoning.
Main tasks still hit Opus. Cheap models handle everything that doesn't need maximum intelligence.
What You'll Save
- Light user — $200 → $70/month (65% savings)
- Power user — $943 → $347/month (~$600 saved)
- Heavy user — $2,750 → $1,000/month ($1,700+ saved)
The fallback chain adds resilience too — if Anthropic is rate-limited, GPT-5.2 picks up instead of your agent just stopping.
Bottom Line
Edit ~/.openclaw/openclaw.json, save, restart. You could be saving hundreds per month by the time you finish your coffee.