Error messages

The server had an error while processing your request. Sorry about that!

The server had an error while processing your request. Sorry about that!

This is the OpenAI API's catch-all for a failure on its own side. The opening sentence is always the same, and what follows it varies:

The server had an error while processing your request. Sorry about that!

The full body, as a developer posted it in April 2026:

{
  "error": {
    "message": "The server had an error while processing your request. Sorry about that! You can retry your request, or contact us through our help center at help.openai.com if the error persists. (Please include the request ID req_f9c2fc5926104bd286d264d57ea56a78 in your message.)",
    "type": "server_error",
    "param": null,
    "code": null
  }
}

It's HTTP 500, and code is null, so the type and the request ID are all you get. There are 244 GitHub issues quoting the sentence as of September 25, 2026.

Is it you or OpenAI?

OpenAI's error-code page puts the cause as "Issue on our servers" and says to retry after a brief wait, contact them if it continues, and check the status page. The forum threads mostly bear that out. The April poster was back the same day to say it was working again. In June 2026 OpenAI Support told another developer "There was an outage yesterday that may have caused this, but that issue has since been resolved."

The Realtime API uses a longer version that asks for a session ID in place of a request ID. In one December 2025 case an OpenAI engineer traced it to "an assert on our side" that fired right after a conversation.item.truncate event. So a server error can still depend on what you sent.

What retrying does, and what the SDK already did

We read openai-python 3.19.2. Any status of 500 or above becomes InternalServerError and is retried twice by default, backing off from half a second. A forum leader's advice on one thread was the same: retry "with exponential backoff" for this "(hopefully) transient" kind of error. If it reaches your code, three attempts have already failed, so a short wait and a check of status.openai.com is the next step.

(A 2023 report from before version 1.0 shows the old library raising this text as RateLimitError, so older threads can look confusing.)

When the same request fails every time

Retries only help when the failure is random. One September 2026 Codex report found every GPT-6 Astra turn on Amazon Bedrock that ran past about 180 seconds ended in this exact server_error. Codex treats it as retryable, so each attempt ran the full three minutes again before failing. We didn't expect a 500 that dependable, and it's a good reason to cap your own retries.

If your failures line up with something (a long request, one model, one input), stop retrying and contact OpenAI support with the request ID from the message. That ID is what support asks for: in one thread OpenAI Support asked for "the error message with the new request id" so it could check the backend.

A 503 with server_is_overloaded is a different error. OpenAI documents it as the model lacking capacity for the moment, and says to follow the Retry-After header when it's there.