Google introduces Gemini 3.8 Live models for voice agents

Google has launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two AI models designed for real-time spoken conversations, voice agents, and tasks that combine speech, visual input, and tool use. Tom Ouyang and Malini Jaganathan announce the models for Google’s official blog.

Gemini 3.8 Live targets applications that need to serve many users at a lower cost. It can process visual information during a conversation, switch between 97 languages, and continue talking while it runs tools or API calls in the background, Google says. That could allow an assistant to confirm a request verbally while it checks a calendar, retrieves information, or completes another task.

The Extended Thinking version is intended for more complicated, multi-step work. Google says the model can reason while speaking, using short verbal acknowledgments and progress updates rather than leaving users in silence while it works. The company presents this as useful for tasks such as booking coordination, troubleshooting, and turning spoken feedback into software components.

Performance and availability

Google reports that Extended Thinking leads several voice-model benchmarks, including Artificial Analysis’ Speech to Speech Quality Index. These results are based on benchmarks cited by Google and have not been independently assessed in the announcement.

Both models are available to developers through the Gemini API and Google AI Studio. Gemini 3.8 Live is also rolling out in Search Live and in private preview for Gemini Enterprise. Extended Thinking is coming to Gemini Live, selected Google Workspace features, and enterprise products, with access depending on subscription tier and rollout status.

Google says all audio produced by its AI products carries a SynthID watermark, an inaudible marker intended to help identify AI-generated sound.