Google DeepMind released two end-to-end voice models on September 15 — Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — bringing "think-while-you-speak" real-time dialogue to the center stage of production-grade voice agents.
Gemini 3.8 Live is a voice model built for scale and cost efficiency: it processes voice, camera feed, images and text, and replies in voice or text, with real-time turn-taking, interruption handling, and the ability to run functions or query APIs in the background while the conversation continues — essentially chatting and working at the same time. Its multilingual capability is a highlight: it detects and switches among 97 languages on the fly, even mid-sentence. For high-volume, cost-sensitive workloads, it is Google's default choice.
Gemini 3.8 Live Extended Thinking, meanwhile, is the "deep reasoning" tier: it maintains fluid dialogue while narrating its own reasoning through verbal cues like "let me check that," suited to complex multi-step tasks. It tops Artificial Analysis' Speech-to-Speech Quality Index at 82.6, leads the agentic voice benchmark Tau3 at 68.6%, and scores 35.1% on Sierra's Tau3-banking — a clear enterprise-grade leader.
Both models run over WebSocket via the Gemini Live API, taking audio, video, images and text as input. Pricing is competitive: $0.005 per minute of audio input and $0.018 per minute of audio output — roughly $1.38 for an hour of continuous conversation. Access spans the Gemini app, Search Live, Google Workspace, and the Live API and AI Studio for developers.
The launch timing is notable: it lands in the middle of a heated global debate among AI giants over whether to slow down. OpenAI had just opened GPT-Live-1 to its API, while Anthropic's CEO urged "pacing the frontier" — Google chose this moment to ship stronger voice models, a clear statement of intent. As "thinking while talking" becomes table stakes, voice AI competition has moved from "understanding" to "reasoning," and falling costs are pushing real-time voice agents from niche demos toward large-scale commercialization.