OutageRadar

Checking… Checking…

Gemini API error 503: UNAVAILABLE

What this Gemini API error means, whether it is on Google’s side or yours, and what to do now. All times are UTC.

Is it them or you?

Server-side: Google’s capacity

This error comes from Google’s servers, not from your code. Google publishes no machine-readable status feed, so the live signal below is our own check of its API server. Google’s own status page may not show it yet either.

Checking…

Gemini

No public status feed from the provider. Reachability checked by OutageRadar.

Our check: checking…

What it means

Google’s docs describe a 503 UNAVAILABLE as the service being temporarily overloaded or down, because it is temporarily running out of capacity.

It is a server-side error: your API key and request are not the cause.

Shape of the error body, from Google’s docs (the message text varies):

{
  "error": {
    "code": 503,
    "message": "...",
    "status": "UNAVAILABLE",
    "details": [
      "..."
    ]
  }
}

What to do

  1. Wait a little and retry with exponential backoff and jitter. The official Gemini SDKs already retry 5xx errors automatically.
  2. Check Gemini’s status above and Google’s status page for an ongoing incident.
  3. Google suggests temporarily switching to another Gemini model to see if it works.
  4. Or switch to a backup provider (below) until it clears.

Retry it safely in your own code

You may not need this: the official SDKs retry on their own (see above). If you call the API directly, this helper waits and retries, using the server’s requested delay when you pass one. Google’s rule: retry 429, 408 and 5xx errors; never retry 400, 402 or 403.

Python

import random
import time


def with_retries(call, is_retryable, retry_after=lambda err: None, attempts=5):
    """Run call(). On a retryable error, wait and try again, up to attempts tries.
    retry_after(err) returns the server's requested wait in seconds, or None."""
    for attempt in range(attempts):
        try:
            return call()
        except Exception as err:
            if attempt == attempts - 1 or not is_retryable(err):
                raise
            wait = retry_after(err)
            if wait is None:
                wait = min(60, 2 ** attempt) + random.random()  # backoff + jitter
            time.sleep(wait)

JavaScript

async function withRetries(call, isRetryable, retryAfter = () => null, attempts = 5) {
  // retryAfter(err) returns the server's requested wait in seconds, or null.
  for (let attempt = 0; ; attempt++) {
    try {
      return await call();
    } catch (err) {
      if (attempt >= attempts - 1 || !isRetryable(err)) throw err;
      let wait = retryAfter(err);
      if (wait == null) wait = Math.min(60, 2 ** attempt) + Math.random(); // backoff + jitter
      await new Promise((resolve) => setTimeout(resolve, wait * 1000));
    }
  }
}

If Gemini is down: switch to a backup

These AI APIs work with the official OpenAI library. Change the address and the key, pick a model from that provider’s list, and you’re running again. Each shows its live status, so you don’t switch to one that’s also down.

Test a backup in development before routing production traffic. Pick a model from a different company than the one that’s down. Your code talks to the provider directly; OutageRadar never sees your key. For a full automatic fallback setup, see TokenMargin.

Recent Gemini incidents

History logged by OutageRadar.

Full history: Is Gemini down?

Source: Gemini API errors (Google docs) · Gemini API troubleshooting: retry strategy (Google docs). Last checked against official docs: Nov 10, 2026. If the docs change, they win.