Google Rolls Out Gemini Voice Models That Keep Working After They Reply

Google has begun rolling out two Gemini voice models designed to keep a spoken exchange moving while the system handles harder work. Gemini 3.8 Live Extended Thinking can stream audio while it performs background reasoning or asynchronous tool calls, promising less interrupted dialogue but making task completion harder to read from the conversation alone.

The release has two versions. Google positions Gemini 3.8 Live as a scalable, cost-efficient model for fluid dialogue and visual grounding, meaning it can use visual material as context. Extended Thinking is the higher-reasoning option for complex, multi-step work while maintaining a live conversation.

Both models can process visual inputs in near real time and execute tools or API calls in the background while a conversation continues. Google also says they can detect and transition between 97 supported languages during a conversation.

Extended Thinking’s operating limits

• It accepts text, images, audio and video, and produces text and audio.
• Only asynchronous, non-blocking function execution is supported; synchronous calls return an error.
• Developers can set low, medium or high reasoning levels; a minimal setting is not supported.

Both models are rolling out through the Gemini API and Google AI Studio. Google says 3.8 Live is also coming to Search Live, while Extended Thinking is available in Gemini Live and selected Workspace offerings depending on subscription tier. Enterprise access begins in private preview, with more Gemini Enterprise and customer-experience availability described as coming soon.

Google reports that Extended Thinking scored 82.6 on Artificial Analysis’ Speech to Speech Quality Index, alongside results on voice-agent and audio-reasoning benchmarks. Those company-reported figures do not establish performance in a particular workflow. Google also says audio generated by its AI products carries an imperceptible SynthID watermark intended to make generated audio detectable.

The release makes an assistant’s first response potentially different from the result behind a request. That can preserve the flow of a live exchange, but applications need to distinguish an acknowledgment from completed work. Google’s status requirement turns that distinction into a product decision as well as an engineering one.