---
title: "Gemini 3.8 Live can keep talking while it runs tools"
url: https://blog.vomini.com/blog/gemini-3-8-live-voice-agents
date: 2026-09-16T13:30:00+00:00
updated: 2026-10-27T22:42:16.491469+00:00
author: "Roberto R."
section: "News"
language: en
site: "Vomini"
license: free to quote with attribution
---

# Gemini 3.8 Live can keep talking while it runs tools

> Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, two voice models that answer speech with speech and can call tools in the background while the conversation goes on. Google says all their audio carries a…

On 15 September, Google released two voice models for developers: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are speech-to-speech models, meaning they listen to a person and answer out loud directly, without a separate step that writes the words down first. They are the kind of model that sits behind a voice agent on a phone line or a help desk.

Google pitches 3.8 Live for scale and lower cost, and the Extended Thinking version for harder, multi-step tasks.

## What changes in a call

For anyone building phone agents, background tool use is the change to look at. Google says 3.8 Live can call tools and APIs while it keeps talking, so it can tell a caller it is checking something and continue the conversation while a booking or a lookup finishes. The Extended Thinking version reasons while it speaks, with short cues to acknowledge a request and spoken updates on longer tasks.

Google also says the models can take in visual input in near real time, and can detect and switch between 97 languages in the middle of a conversation.

## The numbers Google gives

Google says Extended Thinking takes first place on Artificial Analysis' Speech to Speech Quality Index, with a score of 82.6, and scores 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking, two tests of whether a voice agent completes tasks. The banking number is a reminder of how hard those tasks still are. The standard 3.8 Live model placed second in the Speech Agent Arena, a ranking based on user preference. Google calls the pricing competitive but does not list prices in the post.

All audio from Google's AI products is watermarked with SynthID, an invisible mark that lets software detect that the audio was generated by AI.

## Where to try it

Both models are rolling out in the Gemini API and Google AI Studio. For companies, they are in private preview in Gemini Enterprise. Regular users meet 3.8 Live in Search Live.

## What to do if you run a voice agent

Pick the one call type where callers wait the longest in silence today, such as checking an appointment slot, and rebuild just that flow in AI Studio with a background tool call. Listen to whether the pause gets shorter and whether the agent says clearly what it is doing. Keep the first line of the call telling people they are speaking with a virtual assistant, which in the EU has been a [legal requirement since August](/blog/eu-ai-act-transparency-rules-chatbots).

## What we think

In our view, silence while a tool runs is the moment a caller is most likely to give up on a voice agent, so talking through the wait is a practical fix. A 35.1% score on a banking task test also says these agents still need a clear handoff to a person for anything that touches money.

## Sources

- [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/), 15 Sep 2026


---

Originally published at https://blog.vomini.com/blog/gemini-3-8-live-voice-agents