Google DeepMind unveils Gemini 3.8 Live and an extended-thinking version

Google DeepMind has announced two new voice AI agents: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The company calls them its most advanced models for live dialogue and is betting on natural conversation: they handle interruptions, switch languages mid-conversation and run background tasks in parallel without pausing the exchange.

Two models for different jobs

Gemini 3.8 Live is positioned as a solution for scale and cost-efficiency. It processes visual context in near real time, automatically detects and switches among 97 supported languages as the conversation unfolds, and performs tool and API calls in the background — the model can acknowledge a request and keep talking while the task finishes. It came in second on the Speech Agent Arena leaderboard.

Gemini 3.8 Live Extended Thinking is designed for more complex, multi-step scenarios: it reasons and speaks at the same time. To keep the conversation from lapsing into silence, the model uses lines like “Let me check…” and narrates the progress of background operations aloud as it goes. According to the company, it took first place in Artificial Analysis's Speech to Speech Quality Index with a score of 82.6, scored 68.6% on the τ-Voice benchmark, 35.1% on Sierra's τ-Voice-banking, and 97.7% on Big Bench Audio.

Where it works

In the demos, the models run employee onboarding based on what the camera “sees,” play chess, turn sketches and voice comments into ready-made React components, and coordinate multi-step bookings with asynchronous function calls.

For developers, both models are already available in the Gemini API and Google AI Studio. For enterprise customers they are launching in a closed preview in Gemini Enterprise. Regular users will get 3.8 Live in Search Live, and 3.8 Live Extended Thinking in Gemini Live, as well as in Docs, Gmail and Keep for Google AI subscribers. The voice-interface ecosystem is supported by the Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents platforms; partners named include Salesforce, Genspark and Lumeris.

All audio generated by these models is tagged with an imperceptible SynthID watermark — which, in Google's view, should help distinguish AI content and counter disinformation.

Source: Google DeepMind