Gateway access and allowed models
Updated on 3 November 2026.
Control who may send requests through the gateway and which models they may call. Denied requests are rejected before routing or scanning: no provider call, no tokens.
Configure this control
Linked providers and enabled models define an app's model catalog in both modes. The Policy Engine adds caller and model-access rules on top of that catalog. Gateway Access has no classic equivalent. See What happens to classic settings.
- Classic settings
- Policy Engine
Set models for an app
In app settings, an app's usable models are the models of the providers linked to it. To limit an app to certain models, link only the providers (and enable only the models) it should use:
See Link providers to an app and Providers and models. There is no app setting to deny a request by user, group or prompt text. That needs the Policy Engine.
Going further with the Policy Engine
Open Govern > Policy Engine > LLM Gateway > Controls and choose Gateway Access or Allowed Models. Edits join a shared draft and apply once you publish a revision. Gateway Access is policy-only, so nothing freezes when you turn it on; see What happens to classic settings for what the engine takes over.
Gateway Access
Start from a recipe such as Block an app, group or person, or use the section's add menu. Choose Applies to, the decision and any exclusions, check Reads as, then Save to draft. Use Test for matching callers and exceptions before Review & publish.
Allow or deny an entire request by who sent it, which app or model it targets, or what its prompt says.
Phrase rules run at the request stage on chat, responses, bedrock and vertex.
Matching is case-insensitive (except is exactly) and reads the first 8,192 characters. There are no regular expressions; wildcards use * and ?.
Highest priority wins. When an allow and a deny share a priority, deny wins. An allow never bypasses identity, source IP, model or tool checks.
Allowed Models
Use Search rules to find configurations by app, group, person or model. Start from a recipe or a section's add menu, select the scope and model lists, then Save to draft.
Choose which models matching traffic may call, and which models nobody in scope may call.
- Allowed lists combine. Every matching allowed list is merged, and the request may use any model in the union.
- Rejected models always win, including models reached through routing. Priority is ignored on this card.
- No matching allowed list means any configured model may be requested unless it is rejected.
- The card runs before routing: it filters what may be used, then routing picks where the request goes.
The model picker is grouped by provider credential and lists only enabled credentials. Use Search models, Select all and Clear all, or All / None on one credential's group, to maintain long lists without ticking models one by one. Leave Allowed models empty to add only rejections.
Both sections apply on the assistants, bedrock, chat, copilot, embeddings, models, realtime, rerank, responses, sdk_check, stt, text, tts and vertex API surfaces.
On models, GET /v1/models lists only the models a caller may use, and GET /v1/models/{id} returns 404 for a model this card rejects. To take a model off those lists without blocking it, use Model visibility.
An allowed list scoped to Everyone restricts every application and user to that list. Choose Application or another scope to limit who it affects.
Scenarios
- Keep frontier models away from a group. Deny requests when the Smart group is Interns and the requested model is one of a list:
runs on requestpriority 800
- Restrict an app to one team. Deny requests to the Finance Copilot app unless the caller is in the Finance Platform Smart group.
- Pin an app to approved models. An allowed list scoped to the Support Copilot application limits it to
gpt-4.1andgpt-4.1-mini. A request forgpt-4ois rejected. - Ban preview models everywhere. A rejected list scoped to Everyone blocks
o1-previeweven through an app whose allowed list includes it or a routing group that targets it.
For limiting models, an allowed list is usually easier to maintain than a deny rule on Gateway Access.