Skip to main content
Model routing lets Max automatically choose the most appropriate model for each message based on task complexity. Simple questions go to a fast, lightweight model. Complex design or architecture work goes to a more capable premium model. You pay for premium compute only when you actually need it. Routing is disabled by default. When disabled, Max uses whatever model is set as the default for every request.

​
The three tiers

Fast

Simple questions, quick lookups, short follow-ups. Uses a lightweight model for low latency.Default: gpt-4.1

Standard

Most coding tasks, general reasoning, multi-step problems. The everyday workhorse.Default: claude-sonnet-4.6

Premium

Complex design, architecture decisions, UI/UX work, and other tasks that benefit from maximum capability.Default: claude-opus-4.6
The default model set during max setup (stored as COPILOT_MODEL in ~/.max/.env) is used as the standard tier model.

​
How routing works

When a message arrives and routing is enabled, Max runs it through a three-step pipeline:
1

Check keyword overrides

Overrides are checked first and bypass everything else — including the cooldown. If the message matches a keyword rule, the specified model is used immediately.The default override rule:
Rule nameKeywordsTarget model
designdesign, ui, ux, css, layout, styling, visual, mockup, wireframe, frontend design, tailwind, responsiveclaude-opus-4.6
Keyword matching uses word boundaries, so "ui" won’t match "fruit".
2

Classify with an LLM

If no override matches, Max classifies the message as fast, standard, or premium using GPT-4.1. If the LLM is unavailable, it falls back to standard.Short follow-up phrases (yes, no, do it, sure, looks good, etc.) inherit the tier of the previous message rather than being re-classified.Background task completion notifications always resolve to standard.
3

Look up the tier model

The classified tier is mapped to a model using the tierModels configuration:
{
  "fast": "gpt-4.1",
  "standard": "claude-sonnet-4.6",
  "premium": "claude-opus-4.6"
}
If the target model differs from the current model, Max destroys the existing orchestrator session and recreates it with the new model before processing the message.

​
Cooldown

To prevent rapid back-and-forth model switching, Max enforces a cooldown of 2 messages after any tier change. If the classified tier would require a switch but the cooldown is still active, the current model is kept for that message.

​
Enabling and disabling routing

/auto
Toggles routing on or off. When on, replies include a small indicator showing the model used: ⚡ auto · claude-sonnet-4.6.

​
Manual model override

To use a specific model regardless of routing state: 
/model claude-sonnet-4.6
Switching the model this way also turns off auto routing and persists the selection to ~/.max/.env.

​
Viewing and modifying router config

The router configuration is stored in the router_config key of Max’s SQLite state table (~/.max/max.db). You can inspect or update it by asking Max directly: 
Show me the current router config.

Update the fast tier to use gpt-4.1-mini.

Add a keyword override: if the message mentions "security audit", use claude-opus-4.6.
The full config structure: 
{
  "enabled": false,
  "tierModels": {
    "fast": "gpt-4.1",
    "standard": "claude-sonnet-4.6",
    "premium": "claude-opus-4.6"
  },
  "overrides": [
    {
      "name": "design",
      "keywords": ["design", "ui", "ux", "css", "layout", "styling", "visual", "mockup", "wireframe", "frontend design", "tailwind", "responsive"],
      "model": "claude-opus-4.6"
    }
  ],
  "cooldownMessages": 2
}

​
Routing examples

MessageClassified tierModel used
”What’s 2 + 2?“fastgpt-4.1
”Yes, do it” (follow-up)inherits previoussame as previous
”Fix the bug in src/auth.ts”standardclaude-sonnet-4.6
”Design a new dashboard layout”premium (override)claude-opus-4.6
”Refactor the entire payments module”premiumclaude-opus-4.6
”Add a tailwind class to the button”premium (override)claude-opus-4.6

​
When routing is not used

Routing is skipped entirely when:
  • Routing is disabled (enabled: false — the default). Max uses COPILOT_MODEL for every message.
  • A manual model has been set via /model <name>. This overrides the router and sets enabled: false.
  • All tiers map to the same model. If fast, standard, and premium all point to the same model ID, no switch ever occurs, but classification still runs.

Build docs developers (and LLMs) love

Get started for freeTalk to us