Configure the Model Router

The Model Router controls model: "auto" for the organization. Its ordered rulebook chooses an eligible public model for each request, person, or App.

Configure the rulebook

1. Open Organization Administration → Model Router → Rules.
2. Create rules in the order they should be evaluated.
3. Define who is making the request: everyone, selected people, groups, Apps, App type, environment, or credential kind.
4. Define what the request requires: request kind, modality, token estimates, message or tool counts, streaming, context, structured output, or workload type.
5. Select which AI Model should handle a match, or choose automatic selection from eligible public models.
6. Test the rulebook and review the catch-all rule before enabling changes.

The first enabled rule that matches wins. Rules can narrow or prioritize eligible models, but they cannot grant model access, bypass a credential restriction, or override an allowance.

Keep a catch-all rule

An enabled rulebook has one catch-all rule. Evaluated last, it selects either a specific public model or automatic selection from eligible public models. You cannot disable or move it while the router is enabled.

If the selected model has no eligible route, Gateway returns a no-route denial. The router does not switch to a different model because a route is unavailable.

Simple and Advanced modes

Simple mode uses authenticated subject context and Gateway-derived request facts. It does not interpret prompt content before routing.

Advanced mode can use interpreted request attributes such as intent, domain, complexity, reasoning need, and expected output. It adds a separate interpretation request with its own latency and cost. Treat Advanced mode as Preview until the organization approves its data boundary and operating behavior. Callers can opt down to Simple mode, but an API request cannot enable Advanced mode for the organization.

Keep model and provider decisions separate

The Model Router selects a public AI Model. Configure preferred and fallback provider routes on that model’s operational page. Keep provider names, provider-native IDs, provider credentials, and route identifiers out of router rules.

Use the rule tester and diagnostics to review matched, skipped, disabled, unavailable, and winning rules. Confirm that the selected model remains available under current Model Access policies and allowances.