Gemini-3.1-flash-live-preview: first audio 16–26 s since 2026-09-05 ~09:00 UTC, reproduced on an isolated project; same model on 2026-09-03

Gemini API
, ,
1

Confirming the degradation reported earlier today, with measurements from an independent setup.

Setup:

- Production telephony voice agent, Live API (BidiGenerateContent), audio in/out, tools, input/output transcription. Python google-genai 2.21.0, uvloop.

- Three environments (prod, staging, dev), each with its own GCP project and its own API key since 2026-09-03. Dev project runs 4 idle agents, zero open Live sessions before each test, no load anywhere.

- Client in Europe; traffic to generativelanguage.googleapis.com.

Symptom:

Session connects, setupComplete arrives, we send the greeting text. No audio, no transcription, no goAway, no error, socket stays open. First audio eventually arrives 16–26 s later. Our 8 s guard treats the session as dead, so every call fails silently for the caller.

Measurements today (2026-09-05, ~15:00 UTC), minimal standalone probe, one session at a time, 25 s wait window:

  • gemini-3.1-flash-live-preview, default API version: first audio at 16.1 s, 25.7 s and 24.8 s.

- Same model via v1alpha: 17.9 s and 21.1 s, so the API version makes no difference.

- gemini-2.5-flash-native-audio-preview-12-2025, same key, same minute: 8.0 s, 3.4 s and 3.1 s.

All sessions eventually answered. The same probe on the same key two days ago (2026-09-03 evening) gave 1.6–2.1 s for 3.1.

Previous episode:

We saw identical behaviour on 2026-09-03 from roughly 13:15 UTC for several hours, at that time on a shared project across all three environments. It cleared without any change on our side. We then split projects and keys per environment. Today it recurred on the isolated dev project with no load, so project quota or session churn on our side is not the cause.

Vertex AI:

Not a workaround: gemini-3.1-flash-live-preview is still not found on Vertex in europe-west4, us-central1, or global (checked today). Only gemini-live-2.5-flash-native-audio is available there (connects in 1.8 s from europe-west4).

Questions

1. Is this a known incident? Nothing is listed on the Cloud status page.

2. Is there a regional or alternative endpoint for the Developer API Live route?

3. Any timeline for 3.1 Live on Vertex so we can move off the shared Developer API fleet?

Happy to share session IDs from the affected windows.

2

Same here on gemini-3.1-flash-live-preview.

The first response in a session arrives only after 20-30 seconds. And after that, once the dialogue has started, every following turn still has a big delay - roughly 10 seconds per response.

Separately, for the last several days the model itself has been talking noticeably worse in conversation - constant mistakes and poor replies, on a setup that used to work fine.

Is this a known incident, and is there any ETA for a fix? I’m building a large project on this and right now it is not usable in production.

Because of these bugs, I wasted a lot of dollars during my tests.

3

Update:

This behavior is just from the EU. I just made a quick test from a tmp US server, and the timing is perfect.

4

Does this problem not bother other people, why is there no explanation or timeframe for solving the problem?

5

It’s the weekend… and also US region is unaffected