Endpoints

The gateway offers the OpenAI-compatible endpoints most applications need.

What’s available

• Chat completions for conversational and instruction-following calls.
• Responses for the newer response-style interface.
• Embeddings for turning text into vectors, for search and similarity.
• Models to list the models available to you.

Use the model names your administrators have approved for your team (the models endpoint lists them).

Calling them

Because the surface is OpenAI-compatible, you call these exactly as you would OpenAI’s, using your key and the base URL from Get an API key. Streaming works where the corresponding OpenAI endpoint streams.

Note

If a model you expect isn’t listed, it isn’t approved for your team. Ask your administrator to add it to the allow-list (see Approved models & quotas).

Sampling parameters a model doesn’t accept

Some models reject temperature or top_p outright, including OpenAI’s reasoning models and current Claude models. Rather than return the provider’s error to you, the gateway leaves out a sampling parameter the model is known not to accept and sends the rest of the request unchanged.

If you send temperature to a model whose support isn’t known yet and the provider rejects it, the gateway retries your request once without it. You get a normal response, and later requests to that model leave it out from the start.

Your administrator can see which settings each model accepts on the Models page, and can override the platform’s finding for a deployment if it’s wrong.