Model Routing Recommendations
Tell us what your agent does. Get a recommended model per task, with the reasoning behind each choice. Then grab the config snippet and paste it in.
What does your agent do? Select all that apply.
Recommended models
Email triage
Sonnet handles structured triage well at 5x less than Opus. Use Haiku if you only need simple categorization with no drafting.
Code review
Code review is where Opus earns its cost. Complex reasoning, security analysis, and subtle bug detection benefit from the quality gap.
Background monitoring
Monitoring is high-frequency and low-complexity. Haiku is 19x cheaper than Opus and handles check-in tasks perfectly. This is your biggest cost lever.
Savings vs. running everything on Opus
All on Opus
$180/mo
With routing
$26.8/mo
You save
$153/mo
85% less
Costs assume typical token counts per task at the frequencies shown. Your actual savings depend on real usage. See the task estimator for adjustable per-task calculations.
Suggested openclaw.json config
{
"model": "anthropic/claude-opus-4-5",
"heartbeat": {
"model": "anthropic/claude-haiku-3-5"
},
"routing": {
"subagent": "anthropic/claude-sonnet-4-6"
}
}This is a starting point. Review the OpenClaw docs for the full config schema. Not all routing keys map directly to openclaw.json fields.
Want to see what this configuration costs per month? Open the calculator and set your models to match these recommendations.
Read more: Model routing saves 60% -- here is how or The real cost of an AI agent heartbeat
Why routing matters
Running all tasks on the same model is like using a torque wrench for everything. It works. It is also expensive and overkill for most of what you are actually doing.
19x
cost difference between Opus and Haiku at identical token counts
90%
of OpenClaw heartbeat tasks that do not require Opus-level intelligence
$143
average monthly savings from switching heartbeats from Opus to Haiku alone
The routing principle: use the cheapest model that produces acceptable quality for each task. For monitoring and classification, that is Haiku. For content and sub-agent work, that is Sonnet. Reserve Opus for tasks where reasoning depth and quality margins actually matter.