Google ships Gemini 3.8 Live to keep voice agents talking while tools work

Why it matters

Google is moving live AI beyond rapid answers toward voice agents that can reason and operate software without leaving users in silence, a key requirement for support, tutoring and task automation.

Google DeepMind introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15th, giving developers two audio-to-audio models for voice agents that can process visual input, reason and call external tools during a conversation.

The release extends the Gemini push led by Google DeepMind co-founder and CEO Demis Hassabis into a problem that text models can largely avoid: users notice every second an agent spends silently reasoning or waiting for software to return an answer. DeepMind, founded by Hassabis and partners in 2010, previously developed WaveNet, the speech synthesis technology used in Google Assistant. The new models pair that audio lineage with the lab's current work on reasoning and agents.

Gemini 3.8 Live is designed for low-latency conversations and direct tasks. Gemini 3.8 Live Extended Thinking is the more deliberate option, intended for multi-step work such as diagnosing a technical problem, comparing travel options or tutoring a user through code.

Google DeepMind says both models offer upgraded reasoning, near real-time visual understanding and automatic language detection across 97 languages. They accept text, images, audio and video, and return text or audio. The published model specifications give each a 131,072-token input limit and a 65,536-token output limit.

Talking through the tool call

Extended Thinking's defining feature is its treatment of latency. The model can continue speaking while it reasons or waits for an asynchronous tool, using short progress updates such as saying it is checking flight options before delivering the completed answer.

That changes how developers must manage a conversation. A standard Gemini 3.8 Live session returns to an idle state when a turn is complete. With Gemini 3.8 Live Extended Thinking, a completed spoken utterance may only be an intermediate update. Applications must monitor a separate interaction_status field to determine whether the model remains IN_PROGRESS or has reached IDLE.

Extended Thinking also requires non-blocking tool declarations. Synchronous function calls return an error, according to Google's Live API thinking guide. Developers can configure the model's reasoning at low, medium or high levels, trading response time and computation for deeper planning. The standard Live model uses a fixed interleaved-reasoning profile and does not accept that setting.

The distinction gives Google two serving paths for the same voice interface. High-volume support triage, language practice and simple device commands can use the lower-latency model. Workflows where an incorrect answer carries a higher cost can route requests to Extended Thinking, while the agent narrates enough of its progress to prevent the session from feeling stalled.

Google demonstrated the reasoning model as a programming tutor. Its broader Gemini Audio product page also shows Extended Thinking coordinating background agents while maintaining a spoken conversation, and shows the standard model interpreting visual input during dialogue.

A Pro-based audio model

Despite the 3.8 branding, Google's model card says both audio models are based on Gemini 3 Pro. They carry a January 2025 knowledge cutoff and may still hallucinate, slow down or time out, according to the same document.

Google lists the model endpoints as stable, although its broader Live API documentation continues to label the interface as preview. That split matters for teams deciding whether to build customer-facing voice products around the release: the model identifiers have moved beyond a preview suffix, while the surrounding real-time service still carries preview status.

Distribution is broad from launch. The standard model is available through the Gemini API, Gemini app, Google AI Studio, Google Cloud and Google Search Live. Extended Thinking is listed for the Gemini app, API, AI Studio, Google Cloud and parts of Google Workspace, including Gmail, Docs and Keep.

All generated audio carries Google's SynthID watermark, according to DeepMind. The watermark provides a mechanism for detecting speech created or edited by Google AI, an increasingly important control as voice models become capable of handling longer, less scripted interactions.

The practical bet is that voice agents need to remain conversational while becoming more agentic. Tool use, visual input and multi-step reasoning make the software more useful, while each capability introduces delays and more complicated session state. Gemini 3.8 Live Extended Thinking turns that processing gap into another spoken turn, asking developers to build for a conversation that can continue even after the model sounds finished.