Google shipped two new live dialogue models today: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The first is built for scale and cost, the second for complex multi-step work. Both aim at the same thing — voice agents you can actually put into production.
The numbers Google puts forward
Extended Thinking takes the top spot on Artificial Analysis’ Speech to Speech Quality Index with 82.6. On task completion it reaches 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking benchmark. Big Bench Audio: 97.7%.
The smaller 3.8 Live lands second in the Speech Agent Arena. Google stresses price on both, and with voice models that is not a footnote: run a call centre on this and you pay per minute, not per milestone.
What is actually different in the conversation
3.8 Live processes visual input in near real time, handles 97 languages and switches between them mid-conversation without being asked. Tools and API calls run in the background while the model keeps talking — it can acknowledge a request and carry on chatting while the task finishes.
Extended Thinking reasons and speaks at the same time. Instead of going quiet, it drops in verbal cues (“Let me check that…”) and narrates its progress through multi-step background tasks. Google demos it reading a chess board in near real time, and turning rough sketches into working React components off spoken feedback.
Where it lands
For developers: Gemini API and AI Studio, starting today. For enterprises: private preview in Gemini Enterprise. For everyone: 3.8 Live in Search Live, Extended Thinking in Gemini Live and for Google AI Pro and Ultra.
On the platform side Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents are on board; Salesforce and Lumeris are among the named customers. All generated audio carries a SynthID watermark.
Talking while computing is the real change here
The part I care about is not the number one ranking. It is the pause.
Anyone who has worked with a voice assistant knows the moment: you ask something, and then nothing happens for three seconds. People fill that gap on their own — with “hang on”, with “hmm”, with a glance at the screen. A model that fills it too, and says what it is working on while it does, does not just feel faster. It is the first thing a voice agent needs the moment a task takes longer than five seconds.
Whether that holds up outside a demo is a separate question. But it is a more honest answer to the problem than trying to force every response under one second.