Error messages

This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.

This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.

The Gemini API returns this as HTTP 503 when the model you asked for is too busy. The whole body is short. This one is from a CI job's report filed on September 25, 2026:

{
  "error": {
    "code": 503,
    "message": "This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.",
    "status": "UNAVAILABLE"
  }
}

Earlier in 2026 the same 503 carried a shorter message, The model is overloaded. Please try again later. You'll still find that in older logs and issues, but since June the new wording has all but replaced it on GitHub: 245 new issues quote "experiencing high demand", against 5 for "overloaded" (as of September 25, 2026).

What UNAVAILABLE means here

Google's error reference lists 503 as "The service is temporarily overloaded or down" and says to "Wait and retry with exponential backoff." That's the whole of what it says. Nothing in the message points at your key, your request or your quota. A quota problem looks different: that's a 429 with RESOURCE_EXHAUSTED, and we cover it on its own page.

Someone in the SDK's issue tracker checked with raw HTTP and got the same 503, so the SDK isn't to blame. We didn't find a Google staff reply in that thread, which has 46 comments going back to September 2025.

Here's one odd case. A developer saw a request of about 72,000 characters fail with this 503 over and over, while a tiny call to the same model went through at the same moment. That's one person's observation, and Google doesn't say request size plays a part, but it's a reason to test with a small prompt before deciding the model is down.

Is the SDK retrying for you?

Probably not. Google's troubleshooting page says the Python SDK "automatically retries transient errors up to four times." We read google-genai 2.25.0 and it doesn't. With no retry_options set, it makes one attempt and raises a ServerError. Set retry_options (an HttpRetryOptions) in the client's HTTP options and you get five attempts in total, waiting from about one second up to 60, and 503 is on its list of retried codes alongside 408, 429, 500, 502 and 504.

Gemini CLI puts its own dialog over this: "We are currently experiencing high demand," followed by an offer to keep trying or switch with /model.

When retries keep failing

Backoff handles a short spike. For the longer stretches people describe (the SDK thread has users reporting failures for days), the practical fix is a second model to fall back to: catch the 503, and once your retries run out, send the same request to another model. It's no guarantee (the person who opened that thread said switching models didn't help them). Google's troubleshooting page puts 5xx among the errors worth retrying, and also says to set a maximum number of attempts, so a long outage doesn't turn into an endless loop.