Realtime¶
The gateway doesn't serve OpenAI's Realtime API. It answers HTTPS requests, including streams; it doesn't carry
the WebSocket connection at /v1/realtime. For voice and live text, use the endpoints below: each call is governed,
charged and recorded.
What it is¶
OpenAI's Realtime API keeps one WebSocket (or WebRTC) connection open for speech in and speech out with low latency. It's built for live speech-to-speech conversations.
What to use instead¶
| You want | Use | Reaches |
|---|---|---|
| Live text, token by token | Chat Completions with stream=True |
Every provider |
| Live text with tools and stored state | Responses with stream=True |
OpenAI and Azure OpenAI |
| A voice assistant | Speech to text, then chat, then text to speech: Audio and Chat Completions | OpenAI-format providers for audio; every provider for the chat step |
| Audio in and out in one call | Chat Completions with an audio-capable model (modalities: ["text", "audio"]) |
That model's own provider |
Speech to text, chat and text to speech in sequence, each streamed, is fast enough for most voice assistants and gives you a recorded, priced call for every step.
Check it¶
curl -i https://<your-gateway>/v1/realtime -H "Authorization: Bearer $OPENAI_API_KEY"
The gateway refuses it with a 404 and x-vidai-error-code: endpoint_not_found: the path isn't served, as a plain request or as a WebSocket upgrade.
Good to know¶
- An app that connects to OpenAI's Realtime API directly, around the gateway, isn't seen by Vidai: its calls don't show in Request logs or Cost.