Skip to main content
The Model catalog is where your organization decides which models it uses. It lists every model your provider credentials can serve, with its price and this month’s use. Once you have decided, one switch makes the AI Gateway refuse every model you have not approved. Open it from AI assets, Model catalog. Owners and Governance Leads decide. Admins can read it, and auditors see each model’s status and history without its requests or spend.

​
Read the catalog

Each row is one model from one provider, so gpt-4.1 through OpenAI and through Azure are two rows with their own status. A row shows:
  • the model, marked Deprecated when its provider has deprecated it;
  • its status: Approved, Under review, Blocked, or Not reviewed;
  • the price per million input and output tokens, in US dollars: your own price if you set one;
  • its context window;
  • this month’s requests, refusals, spend, and how many AI systems used it;
  • who last decided its status, and when.
Filter by provider, status or use. Choose a row to see which AI systems used it and every decision made about it, or View requests to open the request log for it.

​
Decide about models

1

Pick the models

Tick one or more rows, up to 200 at once.
2

Decide

Choose Approve, Review or Block, with a note for the record if you want one.
On its own, a status changes nothing about traffic: it records what your organization decided.

​
Serve approved models only

Turn on Approved models only at the top of the catalog. From then on the AI Gateway serves only approved models, on every key, including requests that bring their own provider key. Anything else is refused before it reaches the provider. Before it turns on, Altrum lists the models used this month that are not approved, because their requests will start failing. A refused request gets status 403:
Approve the model and the next request is served. Turn the switch off to go back to each key’s own model list. A model must be both approved here and allowed on the key (Models and routing). List models on a key returns only those. A request that names no model, such as an MCP tool call, is not affected. While the switch is on, each decision reaches the AI Gateway at once. If it cannot, the decision is not saved either, so the catalog never shows Approved for a model that is refused: “The AI Gateway could not be updated, so nothing changed. Try again shortly.”

​
Where else a status shows

  • Discover lists models in use that nobody has decided about.
  • An AI system’s bill of materials marks a model it used that is Under review or Blocked.
Each decision, and each change of the switch, is recorded in the audit trail. Each person can make up to 120 changes an hour.