Brief
Gemini 3.8 Live ships two speech-to-speech models behind a single WebSocket endpoint
Google has released Gemini 3.8 Live and 3.8 Live Extended Thinking, two speech-to-speech models, with a working browser client built to try them. The evidence covers the client and the socket endpoint only — not latency, cost, session limits or recovery behaviour.
Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, two speech-to-speech models that the source describes as a similar shape to OpenAI's GPT-Live family. The demo client built to try them uses no libraries.
The integration detail worth noting is the surface itself: one WebSocket endpoint, wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent?key=..., with the Web Audio API AudioContext handling both microphone capture and playback. A model and voice preset are chosen, an optional system prompt is entered, and the model can be interrupted while it is talking — barge-in behaviour handled inside the stream rather than by restarting a request. The source also points to a tutorial for that WebSockets API.
Our reading
For anyone running voice steps in a workflow, the shift is from request-and-response calls to a long-lived duplex socket that owns the audio session, which moves retry, interruption and partial-output handling into your client rather than your orchestration layer. Teams building transcription, screening or phone-shaped flows should care because the WebSocket and the browser audio context are now…
What to do or watch
Before wiring either model into a workflow, build the socket client and test two things the evidence does not cover: how interruption interacts with your own downstream steps, and what happens when the socket drops mid-conversation. Latency, per-minute cost and session limits remain an open question the source does not answer.
Source details and supporting facts
Each line is stated by the page named above it.
Stated by simonwillison.net
- The demo web UI lets you select a model and voice preset, enter an optional system prompt and start a voice conversation through your browser, including the ability to interrupt the model while it is talking.
- The implementation uses no libraries and connects to the wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent?key=... WebSocket endpoint.
- The implementation uses a Web Audio API AudioContext for both capture and playback.
- The source links to a Gemini Live tutorial for getting started with that WebSockets API.
Sources
- Simon Willison's WeblogText stored 16 September 2026
How this story was checked. Written from the 1 page listed above, stored 16 September 2026; claims checked against that stored text on 16 September 2026.
What that means
- 4 of 5 reported statements were confirmed against the page that carries them; the rest were removed rather than published.
- Figures in the text were required to appear in the stored source text: yes. Identifiers: yes.
- The check reads stored text only: no claim rests on a fresh look that did not happen.
- Where the reporting was silent, the text says so instead of filling the gap.