Open AI realtime WS connections are cut off continuously

Yes, I can reproduce this.

With OpenAI-Beta: realtime=v1, the WebSocket upgrade succeeds, but OpenAI then sends an error event and closes the connection with:

code=4000
reason=invalid_request_error.beta_api_shape_disabled

Using the same endpoint and model without the beta header succeeds and returns session.created.

So the practical next step appears to be removing:

OpenAI-Beta: realtime=v1

And then updating any beta Realtime event or session shapes to the GA version.