Official concurrent session / RPS limits for Gemini Live API — where are they documented?

Gemini API
, , , ,
1

I can’t find anything official about the concurrent session limit or the RPS limit for the Live API models (gemini-flash-live 3 preview and Gemini 2.0 Flash Live native audio). Gemini and ChatGPT claim “3 in AI Studio, ~1000 in Vertex,” but neither can give me an official source.

I run telephony infra that handles 10k+ concurrent calls and I’m connecting them to the Live API, so I need real numbers before scaling.

I did run a small stress test on AI Studio (RPS 2, 10 concurrent sessions) and got no 429 errors — but honestly, why should I have to discover these limits by burning my own quota? This should just be documented somewhere.

Could someone please point me to:

  1. The official concurrent session limit (AI Studio vs Vertex)?
  2. The official RPS limit?
  3. Whether these are hard caps or quotas that can be raised?

Thanks!

2

Ok finally found something:

«Bidi generate content concurrent requests per project per region per base model» = 5,000
(for gemini-live-2.5-flash and gemini-live-2.5-flash-native-audio). No info for gemini-3-flash-live