Persistent 503 on gemini-3.7-flash with Priority tier (Tier 2 paid) — 0% success over multiple retries

Gemini API
21

Thank you all for sharing your Id via DM

We are looking into this and working on making the experience better

22

@Mustan_lokhand

For roughly two days, every generateContent call on my account has returned 503 UNAVAILABLE — “This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.”

The message describes capacity, but the surrounding evidence doesn’t fit capacity, and I’ve been unable to find any remaining variable on my side. Posting the full diagnosis in case it’s useful, and to ask what these requests are actually being routed to.

The contrast that rules out load

call auth result
ListModels API key 200, 0.13s
generateContent — gemini-flash-latest same API key 503 UNAVAILABLE
generateContent — gemini-3-pro-image same API key 503 UNAVAILABLE
AI Studio web chat session, same account works, responds in seconds

ListModels returns 50 models and lists generateContent among the supported methods for all three models above, using the same key that then 503s.

Latency profile

From Cloud Console → APIs & Services → Gemini API, over 62 requests:

  • Error rate: 93%
  • Median latency: 1,310 ms
  • 95th-percentile latency: 131,980 ms (~132 seconds)

Locally, three consecutive identical calls took 0.9s, 9.4s, and one full 30s client timeout (TCP connect in 18ms, then nothing).

Load-shedding rejects fast and consistently to protect a backend. A 132-second p95 with outright hangs looks like requests being accepted and then stalling, which is why I don’t think this is demand.

Variables I’ve eliminated

  • 2 API keys on the primary project — both fail
  • 2 projects, including a brand-new one created only for this test — both fail
  • 3 networks: home broadband, mobile tethering, and Google Cloud Shell (inside Google’s own network, different IP and route) — all fail
  • Multiple model families, text and image — all fail
  • Auth path is healthy and evaluated before the failure: no key → 403 PERMISSION_DENIED; invalid key → 400 INVALID_ARGUMENT; my key → 503. The key is accepted, then generation fails.
  • API enabled — generativelanguage.googleapis.com, Status: Enabled
  • No quota pressure — Current usage > 90%: 0 and Seven-day peak > 90%: 0 across all 1,963 quotas. Every GenerateContent request limit per minute for a region row reads Unlimited, current usage 0, in all regions.
  • Billing tier confirmed as Tier 1 · Prepay in AI Studio, with a billing account linked to the project and a positive balance. A silent reversion to free tier was specifically checked and ruled out.
  • ~62 lifetime requests total on this account — nowhere near any limit, ever.
  • No incident posted on the AI Studio / Gemini API status page for this period.

Reproduction

curl -s -w "\nHTTP %{http_code} in %{time_total}s\n" \
  -X POST "https://generativelanguage.googleapis.com/v1beta/models/gemini-flash-latest:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"hi"}]}]}'

Returns 503 every time. Swapping to https://generativelanguage.googleapis.com/v1beta/models?key=$GEMINI_API_KEY returns 200 in ~0.13s with the same key.

Question

Given zero quota usage, a working web UI on the same account, an instant ListModels on the same key, correct Tier 1 billing, and a 132-second p95 latency — can someone check what this account’s generateContent requests are being routed to? Happy to provide the project ID and a request timestamp privately.

23

Same 503 NBP / It’s been going on for 3 days.

24

same thing, really, i will switch to another provider, it’s unacceptable.

25

I tend to agree quite unacceptable. We were coming over from Anthropic, primarily because of better latency here on the flash models. but if TTFT is variable from ~ms to 30+ seconds while we wait for retries we will have to abandon.

26

The most troubling part of this is that the Google AI status page ( Google AI Studio ) says everything is fine. It feels like a trust issue at this point that they won’t admit to these problems.

It feels like we’re just yelling into the wind and no one is listening. Because of this, I think it’s best for us to start looking at other companies’ models instead—specifically ones that seem to care more about their users and platform reliability.

27

In my case, no other models are capable of do the work as good as gemini does, so im hanged down.

28

Hello Guys, with the announcment for 3.8 flash i just tried and seems like now works…

29

Still same problem… 503 all day, for.. weeks now.. 3.7 / 3.8 flash