Google releases two new Gemini models built for live voice

Google has announced two new live dialogue models aimed at voice interaction: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are available from today through the Gemini API, Google Workspace and the Gemini app.

Gemini 3.8 Live processes visual input in near real time, automatically detects the 97 languages it supports and moves between them mid-conversation. It runs tools and API calls in the background, so it can acknowledge a request and keep talking while the task finishes.

The Extended Thinking variant reasons and speaks at the same time. According to Google, it uses early verbal cues such as "Let me check that..." to acknowledge prompts and narrates progress through multi-step background tasks.

Benchmark figures

Google says Gemini 3.8 Live Extended Thinking took the top overall spot on Artificial Analysis' Speech to Speech Quality Index with 82.6. On agentic task completion it scored 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark, plus 97.7% on Big Bench Audio.

Gemini 3.8 Live came second in the Speech Agent Arena. Google describes it as a cost-effective option built for scale. ServiceNow's EVA-Bench, a benchmark for voice agents, was run on the Live API on the Gemini Enterprise Agent Platform.

Integrations

Developer platforms using the Gemini Live API include Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents. Google also said it is working with Salesforce, Genspark and Lumeris, which pointed to the models' latency, fluidity and tool-calling.

In Workspace, Extended Thinking can be tried in Docs Live, Gmail Live and Keep Live. All audio generated by the models is watermarked with SynthID.

Full details are in the Google DeepMind announcement.