openclaw - 💡(How to fix) Fix Feature: Per-task model routing (route simple tasks to cheaper models) [1 participants]

Summary

Allow OpenClaw to route messages to different models based on task complexity, skill, or explicit configuration — so simple tasks (weather, reminders, smart home, basic Q&A) use a cheaper/faster model like Haiku while complex tasks (research, coding, financial analysis) use Sonnet or Opus.

Use Case

I run an always-on OpenClaw instance on a Mac mini managing 3 live SaaS products, an investment platform, and a custom research bot (Sentinel). My primary interface is Telegram with ~100+ interactions/day.

Current state: Every message hits Claude Sonnet 4.6 ($3/$15 per MTok) regardless of complexity — "turn on the lights" costs the same as "analyze this SEC filing."

Desired state: Simple tasks route to Haiku 4.5 ($1/$5) automatically, with Sonnet as default and Opus for explicitly complex work.

Estimated savings: 60-70% of my Telegram traffic is simple Q&A/commands. Routing those to Haiku would save ~$60/month on a ~$300/month bill.

Proposed Config

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-6",
        "fallbacks": ["anthropic/claude-opus-4-6"],
        "routing": {
          "rules": [
            {
              "match": {"skills": ["weather", "openhue", "sonoscli", "blucli", "eightctl", "apple-reminders", "apple-notes"]},
              "model": "anthropic/claude-haiku-4-5"
            },
            {
              "match": {"thinking": "max"},
              "model": "anthropic/claude-opus-4-6"
            }
          ]
        }
      }
    }
  }
}

Alternative approaches that would also work:

• A /model slash command to set per-session model (like /think sets thinking level)
• Skill-level model override in skill config
• A lightweight Haiku classifier that picks the model before the main call

Context

• OpenClaw version: 2026.3.23-2
• 38 skills active, many are simple command-response patterns
• Already using task-specific MCP configs (mcp-coding.json, mcp-financial.json, mcp-research.json) for tool routing — model routing is the natural next step
• Anthropic pricing: Opus $5/$25, Sonnet $3/$15, Haiku $1/$5 per MTok

Impact

This would be valuable for any OpenClaw user running always-on with high message volume. The cost difference between routing everything through Sonnet vs intelligently splitting across 3 models is 40-60%.