Google ships Gemini 3.8 Live: voice agents meet production

On 15 September 2026, Google announced two live voice dialogue models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are available through the Gemini API from the moment of the announcement, and already wired into the Gemini app, Google Workspace and Search. One caveat belongs up front: the only source for this article is Google’s own blog post. Every figure below is the vendor’s, and not one of them has been reproduced by an independent third party.

One model for volume, one model to think

The split is clean. Gemini 3.8 Live is pitched at scale and cost efficiency: smooth dialogue, conversational intelligence and visual grounding — meaning the model also reasons about what it can see while you talk to it. Google puts it second on the Speech Agent Arena, a leaderboard built on user preference.

Gemini 3.8 Live Extended Thinking takes the other end of the problem: high-complexity tasks and multi-step reasoning, with what the post calls near real-time reasoning. That phrasing is the format confessing its own weakness — in voice, thinking time is audible. Three extra seconds of compute is a spinner on screen; out loud, it’s the pause where you check whether the call dropped.

The scores, and what they actually say

Extended Thinking takes first place on Artificial Analysis’s Speech to Speech Quality Index with 82.6, and posts 97.7% on Big Bench Audio. So far, the curve only goes up.

Then come the two numbers that will mean something to anyone who has ever tried to put a voice agent in front of real customers: 68.6% of tasks completed on τ-Voice, and 35.1% on τ-Voice-banking, Sierra’s banking test suite. Google notes that both are the best on the market for agentic task execution. Read that again: on a realistic banking scenario, the best voice dialogue model money can buy today gets it wrong roughly two times out of three. The post talks about building blocks for reliable, production-ready voice agents. It never says whose production.

What speech-to-speech changes in your code

The technical interest isn’t in the leaderboards. A voice assistant is traditionally built on three floors: speech to text, a text model, then text to speech. Each floor adds its own latency, and both ends drop everything text can’t carry — hesitation, intonation, the fact that the user just cut you off mid-sentence. A model that takes audio in and puts audio out removes those seams, and with them a good share of the plumbing you currently babysit.

Google also says these models run background tasks, tool calls included, without interrupting the conversation in progress. On paper, that’s the end of the dead air while an agent goes off to query an API. On paper only: the post documents neither measured latency nor test conditions.

A “highly competitive” price, and nothing else

Both models are sold on value for money against what Google calls other frontier models — naming none of them, and quoting no price whatsoever. For Gemini 3.8 Live, cost is even the opening line of the pitch.

A model whose headline feature is never leaving an awkward silence, announced by a post that goes quiet the second money comes up. On that point at least, the dead air comes through loud and clear.