server rejected WebSocket connection: HTTP 401

[ibrhimAli]

Describe the bug

~2 months ago I faced an issue with http requests to openai agents, why? because my requests was taking too long ~100k token per request.
Suddenly those http requests stopped and return no results.

I migrated to use websocket which helped fixing the issue and improved the performance with 10-40%.

Today the websocket isn't available and every request returning below:

{ "error": "server rejected WebSocket connection: HTTP 401" }

Debug information

• Agents SDK version: latest
• Python version (e.g. Python 3.13)

Repro steps

Just try any request using websocket setup.

Expected behavior

It should return answer/response instead of errors.

PS:
I'm not sure if i'm the only one, but I faced many reliability issues with openai platform lately.

this one was days ago: https://community.openai.com/t/chat-completion-tab-on-platform-isnt-loading/1380271?u=budy

[seratch]

Thanks for raising this, and we're sorry for the reliability issues you've been running into. We understand why this makes it hard to feel confident about a production setup, especially after you already moved from HTTP to WebSocket to work around long-running requests.

I checked the SDK-side WebSocket runtime path and the current public OpenAI status page. The SDK's Responses WebSocket transport still builds the connection to the /responses WebSocket endpoint and forwards the OpenAI client's auth headers during the handshake, and the related local runtime tests are passing. I also don't see an active platform-wide incident right now on status.openai.com. There was a previous resolved incident for elevated 401s on API endpoints on Apr 14, 2026.

A 401 during the WebSocket handshake usually points to authentication or access configuration rather than the model request itself. Since confirming the exact cause may require looking at your specific app, project, organization, key configuration, or network/IP allowlist settings, please contact our customer support team with the relevant request details and redacted logs. They can investigate account- or org-specific configuration safely.

From the SDK side, I don't currently see evidence that this is caused by the Agents SDK runtime itself. If support identifies an SDK-level issue, we're happy to follow up here with a fix.

[ibrhimAli]

That's what I suspected when the issue happened.

But I didn't change any api keys, the fix was just reverting back to http requests instead of websocket.