Gemini has an unusually wide surface area, the Gemini app, the API through Google AI Studio, the same models via Vertex AI on Google Cloud, and Gemini features baked into Search, Workspace, and Android. Any one of those can wobble while the others are fine, so "is Gemini down?" almost always needs a "which Gemini?" before you can answer it.
Which "Gemini" are you using? ¶
- gemini.google.com, the consumer chat app.
- Google AI Studio / the Gemini API, the developer surface most integrations call.
- Vertex AI, the same models served through Google Cloud, with its own regional endpoints and status.
- Embedded Gemini, inside Search (AI Overviews), Gmail/Docs, or on-device, which depends on those products' own availability.
Because these ride partly on Google Cloud, a Gemini problem is sometimes actually a Cloud-region problem. That matters for where you look next.
The check ¶
- Probe via 4 probe methods. Check gemini.google.com (and your API endpoint if you use one) using 4 independent probe methods. All four failing is global; a mixed result points to a partial issue, sometimes a single degraded Cloud region behind the endpoint.
- Compare on the AI board. The AI provider status board puts Google next to OpenAI and Anthropic. If ChatGPT and Claude are healthy but Gemini isn't, it's a Google-specific event.
- Check the right status page. The Gemini app, the Gemini API, and Vertex AI report on different Google dashboards. If you're on Vertex, the Google Cloud status for your region is the source of truth, not the consumer app's status.
Reading the API errors ¶
| Status | Meaning | Down? |
|---|---|---|
| 500 / 503 | Server error / unavailable | Yes, retry with backoff. |
| 429 RESOURCE_EXHAUSTED | Quota / rate limit | No. You've hit a per-minute or daily quota. Handle it → |
| 403 PERMISSION_DENIED | API not enabled / key scope | No. Enable the API or fix the key. |
| 400 INVALID_ARGUMENT | Bad request | No. Your payload, often a safety block or context length. |
A common Gemini surprise is a response that returns successfully but empty because a safety filter blocked it, that's not an outage at all, it's the model declining. Check the response's finish reason before assuming the service is down.
If it's local, not Google ¶
All regions reach Gemini but you can't? The usual causes are a regional availability restriction (some Gemini surfaces aren't offered in every country), an ad blocker or extension interfering with the app, a signed-out or wrong Google account, stale DNS, or a corporate network blocking it. Try a private window, a different account, and a different network to isolate it.
FAQ ¶
Is Gemini down for everyone or just me?
Probe gemini.google.com with all four probe methods. If all fail, it's a global or wide outage. If all succeed but you can't reach it, it's local, your account, network, or a country-level availability restriction. Because Gemini spans the app, the API, and Vertex AI, also confirm you're checking the same surface you're actually using.
Why did Gemini return an empty answer instead of an error?
That's usually a safety filter blocking the output, not an outage, the request succeeded but the model declined to answer. Inspect the response's finish/block reason. A true outage returns a 5xx status or fails to connect, not a successful-but-empty response.
My Vertex AI calls fail but the Gemini app works, where do I look?
Vertex AI runs on Google Cloud with regional endpoints, so its availability can differ from the consumer Gemini app. Check the Google Cloud status for your specific region rather than the app's status page, and confirm the region you're calling is healthy.
The three Geminis, and why people check the wrong one ¶
"Gemini" names three products with separate availability, and most confused reports come from checking one and using another:
- The Gemini app (gemini.google.com), the consumer chat surface. Tied to your Google account and subject to country-level availability rules.
- The Gemini API via Google AI Studio, key-based, aimed at developers, with its own quotas.
- Vertex AI, the same models inside Google Cloud, with regional endpoints, IAM, and Cloud's own status reporting.
An incident in one frequently leaves the others untouched. Before concluding anything, confirm which surface your traffic actually uses, then check that one. The app being fine tells you nothing about a Vertex region in europe-west4.
Down, degraded, or rate-limited ¶
These need different responses and are easy to conflate:
- Down, connections fail or every request returns 5xx. Nothing to do but retry with backoff and wait.
- Degraded, requests succeed but far slower than usual, or a fraction fail. Often worse operationally than a clean outage, because retries pile on and your own timeouts start firing. If your p95 has tripled but nothing errors, you're here.
- Rate-limited, 429
RESOURCE_EXHAUSTED. Not an outage at all. The service is healthy and declining your volume; the fix is your side. We covered the general shape in AI API rate limits explained.
The distinction matters because only the first is worth waiting out. Treating a 429 as an outage means sitting idle while the actual fix, backoff, batching, or a quota increase, goes unmade. And a status page that reads green during a degradation is not lying; it's answering a coarser question than the one you asked, which is why measuring your own latency beats waiting for a provider to confirm what you already suspect.
What to do while it's down ¶
Retry with exponential backoff and jitter, not a fixed interval, synchronised retries from every client are how a recovering service gets knocked over again. Cap total retry time so a request fails cleanly rather than hanging a user's page. If Gemini is load-bearing for you, having a second provider configured behind a flag turns a provider outage into a degraded mode instead of a stoppage; the same reasoning applies to OpenAI and Anthropic, and it's why we probe all three.