Custom Model Catalog for Custom Providers
Bring any OpenAI-compatible model under full cost tracking. Define pricing, context limits, and capabilities per custom provider key — so requests through custom providers get billed, enforced, and reported just like a built-in model. Available on Enterprise.
Custom providers let you route to any OpenAI-compatible endpoint — a self-hosted model, an internal service, or a provider we don't support natively — using mycompany/my-model. Until now those requests had a blind spot: LLM Gateway held no catalog entry for the model, so it couldn't know the price, context window, or capabilities. That meant no cost attribution and no enforced limits. The Custom Model Catalog closes that gap.
Define a model
Open Organization → Custom Models, pick a custom provider key, and add a catalog entry. Only the model id is required — every other field is optional:
| Field | What it controls |
|---|---|
| Model id | The id after the provider prefix (gpt-5.5 in mycompany/gpt-5.5). |
| Display name | An optional human-readable label for dashboards and analytics. |
| Context size / max output | Token limits, enforced per request. |
| Token prices | Input, output, cached input, cache read/write (5m and 1h), per-request, web search, image, and audio input. |
| Capabilities | streaming, vision, tools, reasoning, jsonOutput, audio. |
| Supported parameters | An advisory list of accepted request parameters. |
Prices are in USD per token and accept decimal (0.000003) or exponent (3.0e-6) notation.
Cost attribution that matches reality
When a request matches a catalog entry, the gateway bills it at the prices you set and records the cost on the activity log — the same place your built-in model spend already shows up. No catalog entry, no change: the request stays unbilled, exactly as before.
One thing to know: a known model id (say gpt-5.5) routed through a custom provider is not billed at that model's public pricing. Only the prices you define in the catalog apply, so define the model to attribute its cost.
Limits and capabilities, enforced upstream
When an entry sets limits or capabilities, the gateway checks them before forwarding the request:
- Context size / max output — requests over the configured window or output budget are rejected with
400, before they ever reach your endpoint. - Capabilities — a flag set to disabled rejects requests that use it: images against a non-
visionmodel, tools against a non-toolsmodel, streaming against a non-streaming model. Flags you leave unset stay permissive and your upstream provider decides.
Lock it down with catalog-only mode
Every custom provider key has an Only allow catalog models switch. Turn it on and requests must reference a defined catalog model — anything else returns 400. That guarantees every request through the key has known pricing and enforced limits. Leave it off and undefined models keep working, unbilled, as they always have.
Availability
The Custom Model Catalog is available on the Enterprise plan for organization owners and admins. The gateway always honors entries that already exist, so your pricing and limits keep working even after a plan change.