GPT-Live (gpt-live-1, France Central): output_creation_failed on session start, only between 07:00–10:15 UTC

Question

GPT-Live (gpt-live-1, France Central): output_creation_failed on session start, only between 07:00–10:15 UTC

Charan A 0

We run an outbound voice agent on Azure GPT-Live and get intermittent session-start failures that only happen in a fixed morning time window. We'd like to confirm whether this is a known capacity/service issue in France Central / the EU region.

Setup

Model: gpt-live-1, deployed in France Central (WebSocket, /openai/v1/live/sessions)
Client: LiveKit Agents 1.8.4 (livekit-plugins-openai 1.8.4, GPTLiveModel.with_azure), Responses delegation enabled
Agents run on a VM in Central India (4 worker containers)
Usage: outbound phone calls in batches, ~20–30 new sessions per minute at batch start, up to ~90 concurrent sessions

Flow

We dial the patient. When they answer, we open a GPT-Live session (session.start).
We ask the model to speak the greeting.
On affected calls, no audio comes back. After ~10 seconds the server sends the error below and closes the connection. session.started is never received.
Retries (new sessions) often fail the same way for several minutes.


{"type": "error", "error": {"type": "server_error",

 "code": "output_creation_failed",

 "message": "Failed to create Live WebSocket session"}}

The WebSocket connects successfully every time. There are no 429/rate-limit errors and no auth or validation errors.

When it happens

| Date (2026) | Failure windows (UTC) | Calls affected |

|---|---|---|

| 6 Oct | 07:39–07:48, 10:06–10:10 | 12% |

| 7 Oct | 07:09–07:13, 07:45–07:56, 09:49–09:57 | 49% |

| 8 Oct | 07:08–07:13, 07:48–07:54, 09:53–09:56 | 28% |

| 9 Oct | none | 1% |

| 12 Oct | 08:33–08:44, 09:31–09:47 | 17% |

| 13 Oct | none | 0% |

Why we believe this isn't on our side

All failures across all days fell between 07:00 and 10:15 UTC. After 10:30 UTC we had 0 failures in ~415 calls, at the same or higher load (up to 89 concurrent sessions).
The same code, configuration and call rate were used on every day. Some days were clean (9 and 13 Oct), others were not.
On bad days, failures start in the first minute of a batch, when only a few sessions are open.
All 4 worker containers fail at the same moments.

Questions

Were there capacity or service issues for GPT-Live in France Central / the EU region on these dates and times?
Is output_creation_failed a capacity-related error? Is there a recommended way to detect or handle it?
Could we be hitting a concurrent-session or session-creation-rate limit (shared across our subscription)?
Would a different deployment type (Data Zone / Global), another region, or Provisioned Throughput avoid this for GPT-Live?

We can share exact failure timestamps if needed. Thanks!

0 comments

1 answer

Your answer

Answer 1

Hello Charan, Welcome to MS Q&A

Based on the details shared, this looks more like a possible Azure service-side or regional capacity issue than a client-side problem, especially since all four workers are affected simultaneously and failures occur only during specific morning windows.

However, we can’t confirm the root cause without Microsoft suppot investigating the deployment and service telemetry. The output_creation_failed error alone doesn’t establish that this is a capacity issue or that a concurrency limit has been reached.

I’d recommend asking you to create Azure Support ticket to:

Confirm any service incidents or capacity constraints in France Central during the reported UTC windows.

Check for session-creation rate or concurrent-session limits on the subscription and deployment.

Confirm whether Data Zone/Global deployments or Provisioned Throughput are supported for this model and would help mitigate the issue.

Advise on recommended retry handling and backoff for failed session starts.

The exact failure timestamps, deployment details, and request IDs (where available) should help Microsoft narrow this down.

Pls check and let me know if any further questions

Thanks
Deepanshu

Charan A 0 Reputation points

2026-10-13T10:08:36.2266667+00:00

Thanks Deepanshu, that matches what we're seeing. We'll raise an Azure Support ticket with the exact failure timestamps and deployment details, and will update this thread with Microsoft's findings so others in France Central can benefit.