Error messages

Our servers are currently overloaded. Please try again later.

Our servers are currently overloaded. Please try again later.

The text is short:

Our servers are currently overloaded. Please try again later.

It comes back with HTTP status 503. A developer posted the full body in May 2026, found inside a Batch API results file:

{
  "error": {
    "message": "Our servers are currently overloaded. Please try again later.",
    "type": "service_unavailable_error",
    "param": null,
    "code": "server_is_overloaded"
  }
}

What OpenAI says it means

OpenAI's error code guide calls it "Model temporarily overloaded" and says it "indicates that the requested model does not have enough capacity to process your request at the moment." Your key and quota aren't part of that definition, only the model's spare room at that moment.

A support reply from OpenAI on the forum this month says the same about a real case. A Pro subscriber running GPT-5.6 Sol through OpenCode kept getting it, and OpenAI_Support wrote on September 22, 2026: "I traced one of your OpenCode requests to a server_is_overloaded error before it generated any output. It passed the usage checks for your Pro plan, so the failure likely came from server capacity on our end."

How it differs from a 429

This changed recently enough that old error handlers get it wrong. The rate limit guide says some endpoints used to return a 503 with the code slow_down for two separate situations. Now a sudden jump in your own traffic gets a 429 with slow_down, and a model without room gets this 503 with server_is_overloaded. Video requests moved the other way: overload there used to come back as a 429 and now comes back as this 503.

The split matters for what you do next. A 429 slow_down is about how fast your traffic grew, so the guide says to ramp up more gradually (it can happen even inside your per-minute limits). For a 503 overload the guide asks for less: wait, then retry. In the Python library the two also raise different classes: RateLimitError for 429 and InternalServerError for 503, so a handler that only catches one of them lets the other through.

How long to wait

The guide's instruction: if the response has a Retry-After header, "wait at least as long as it specifies before retrying." If it doesn't, increase the delay between retries. If it goes on, check status.openai.com for an incident.

The official SDKs retry 503s for you. We read openai-python 3.19.2: it retries twice by default, starting at half a second and doubling up to 8 seconds, and it obeys Retry-After up to two minutes. A longer Retry-After makes it give up and raise the error to you, which is what OpenAI's guide wants too: defer the request rather than try again early. Raise max_retries on the client if two tries aren't enough for a batch job.

When it arrives mid-stream

With streaming, the HTTP status is sent before any output. If overload hits after the stream has started, the response is already a 200 and the error comes as an event inside the stream, like this one from an OpenCode bug report:

{"type":"error","error":{"type":"service_unavailable_error","code":"server_is_overloaded","message":"Our servers are currently overloaded. Please try again later."}}

The SDK's retry check looks at HTTP responses, so it never sees this one. That's why OpenCode and several Codex proxy projects have had bug reports this year about turns that stop dead on it. OpenAI's guide adds a caution for writing your own retry: "don't automatically replay a request after consuming output". (Your app has already shown part of an answer by then, so a silent replay would start it over from scratch.)

For the 500 that apologises instead, see “The server had an error while processing your request”. On older openai-python versions (before 1.0) the same overload showed up as ServiceUnavailableError: The server is overloaded or not ready yet., which is still all over the 2023 forum threads.