Google launches Gemini 3.8 Live speech-to-speech models

Google's release of Gemini 3.8 Live and 3.8 Live Extended Thinking adds two speech-to-speech models to its lineup. The pair mirrors the shape of OpenAI's GPT-Live family, with one standard model and one extended-thinking variant, suggesting Google is positioning the two as alternatives for live voice applications rather than as distinct products.

Alongside the release, a browser-based demo UI was built against the documentation, letting users select a model and voice preset, enter an optional system prompt, and hold a voice conversation that can be interrupted while the model is talking. The implementation deliberately avoids libraries, connecting directly to a Google WebSocket endpoint and using a Web Audio API AudioContext for both capture and playback, which points to how lightweight a client for these models can be.

The practical test for these models is likely whether developers can get live voice conversations working without heavy dependencies, and whether Google's own tutorial makes that setup straightforward. For teams building voice interfaces, the extended-thinking variant may matter most where a live answer benefits from more deliberation, though that depends on how the models behave in practice.