CONFIRMED

Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two live dialogue models aimed at the biggest weakness of voice agents: conversation stops when the task becomes complicated. The new approach lets the system keep talking, acknowledge the request, and work in the background while it calls tools, processes visual context, or executes a sequence of actions.

Two speeds for the same conversation

Gemini 3.8 Live is designed for scale, cost efficiency, and low latency. It combines fluid dialogue with near real time visual understanding and can detect and switch between 97 languages during a conversation. Live Extended Thinking targets harder work, with multi step planning and parallel reasoning. That distinction matters for builders: one agent needs to answer immediately, while another needs to finish a complex task without leaving the user in silence.

The change is operational, not just vocal

The most important feature for product teams is non blocking tool use. An agent can acknowledge that it is checking something, provide progress cues, and keep the interaction alive while an asynchronous call completes. This fits bookings, support, onboarding, document search, customer service, and any voice experience that must query external systems without sounding stalled. Extended Thinking can also reason and speak at the same time, using short verbal cues to keep the turn active.

Where it is rolling out

The rollout began on September 15. Gemini 3.8 Live is arriving in the Gemini API and Google AI Studio for developers, in Search Live for the public, and in private preview in Gemini Enterprise. Extended Thinking is rolling out through the Gemini API, AI Studio, the Gemini app, and Google Workspace features including Docs, Gmail, and Keep, with availability varying by product and subscription. Google’s documentation and model card describe the family as based on Gemini 3 Pro, with multimodal audio, image, video, and text input, up to 128K tokens of context, and audio and text output up to 64K tokens.

Benchmarks are not a blank check

Google says Extended Thinking scored 82.6 on Artificial Analysis’ Speech to Speech Quality Index, 68.6% on τ-Voice, 35.1% on Sierra’s banking benchmark, and 97.7% on Big Bench Audio. The company also says the standard Live model placed second in Speech Agent Arena. These figures are relevant signals about product direction, but they remain vendor reported results with configurations and conditions that should be reproduced before guiding architecture or commercial promises.

The MaxAssistant reading

This launch matters because it makes tool latency part of the visible experience instead of leaving it as a backend problem. For voice agents, speaking while working may be more valuable than simply producing a nicer voice. The caveat is rollout: parts of the offering remain in preview or are limited to subscribers. Teams testing it should separate the fast endpoint from the reasoning endpoint and keep the integration ready for changes in availability, cost, and behavior.

Sources

Google, Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/ Google DeepMind, Gemini 3.8 Audio model card: https://deepmind.google/models/model-cards/gemini-3-8-audio/ Google AI for Developers, Thinking in the Live API: https://ai.google.dev/gemini-api/docs/live-api/thinking