Gemini 3.8 Live: Should a Voice Agent Think While Talking?

On September 15, 2026 Google released two live voice models instead of one. Gemini 3.8 Live answers without a reasoning pause and is built, in Google's words, for scale and cost efficiency. Gemini 3.8 Live Extended Thinking reasons and speaks at the same time, saying things like "Let me check that…" while it works. The split turns a model choice into a product choice: should your voice agent sound instant, or should it be allowed to think out loud?

This post is for the builder choosing a live voice model this month. It reads Google's announcement and the Gemini API pricing page as read on September 16, sets out the decision, and prints the benchmark figures with their sources. We have run no latency tests of our own and quote none.

Both models are audio-to-audio, meaning speech goes in and speech comes out without a separate transcription step, over Google's Live API. Google describes 3.8 Live as combining conversational intelligence with fluid dialogue and visual grounding, and 3.8 Live Extended Thinking as built for high-complexity tasks with multi-step reasoning. The model IDs in the API are gemini-3.8-live and gemini-3.8-live-extended-thinking, and both are listed as stable, not preview. The earlier 3.1 Flash Live preview is now marked legacy with a recommendation to move.

Three capabilities Google attributes to 3.8 Live matter for a business agent. It processes visual input in near real time. It "automatically detects and transitions between 97 supported languages mid-conversation". And it executes tools and API calls in the background while continuing to talk, so it can acknowledge a request and keep the conversation going while the task finishes. Extended Thinking adds reasoning on top of that, and Google's description of how it fills the gap is the reason this post exists.

A voice agent has one budget that a chat agent does not: the caller's patience with silence. Every second a model spends reasoning before it speaks is a second the caller hears nothing. The industry's first answer was to make models fast enough that the pause disappeared. Google's second answer is to let the model talk through the pause. That is a different product, and it suits different calls.

An instant model suits short, high-volume interactions where the answer is usually retrieval: opening hours, order status, a booking change, a routed transfer. A narrating model suits calls where the caller expects the agent to go and do something that takes several steps: check three systems, apply a rule, come back with a decision. In that setting a spoken "Let me check that" is what a competent human would say, and silence is what a broken one would produce.

The cost of narration is that it is speech the caller has to sit through. If the task takes two seconds, a preamble is padding. If it takes twenty, it is the difference between a caller staying on the line and hanging up. Google's post gives no figures for how long the background tasks take, so the only way to know which side of that line your calls fall on is to measure your own. Our voice agent latency measures reference defines the numbers to collect; this post does not repeat them.

The routing below is ours. It follows from Google's own descriptions of the two models and from the shape of the call, not from any measurement of the models. Because Google's pricing page lists a single Live rate for both, price does not enter the decision at the per-token level.

One earlier decision still applies whichever model you pick. OpenAI's live voice API took a different route to the same problem, which we described in our business guide to GPT Live 1, and the choice between vendors is not settled by either vendor's benchmarks. The Grok Voice think-while-talking release is the third design in the same space. Three vendors, three different answers to the silence problem, and no cross-vendor test any of them has published.

Google lists availability by audience. The table reproduces it, with one caution: "private preview" is an invitation, not availability, and "coming soon" carries no date.

On price, the Gemini API pricing page as read on September 16 lists a single paid tier for "Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, and Gemini 3.1 Flash Live Preview" together. Per million tokens, input is $0.75 for text, $3.00 for audio and $1.00 for image or video; output, including thinking tokens, is $4.50 for text and $12.00 for audio. Google also states the audio rates per minute: $0.005 in and $0.018 out. The page lists a free tier for both models and says free-tier content is used to improve Google's products while paid-tier content is not.

Google cites five results. Each is either a leaderboard position on a third-party site, dated the day of the announcement and free to move, or a score Google reports for its own model. None is a test of a business call, and none compares the two Gemini models against each other on the same task.

Google also says the models sit on the accuracy-versus-quality frontier of ServiceNow's EVA-Bench for complex workflows, with a note that the run used the Live API on Google's enterprise agent platform. Read all of these as a vendor's selection of favourable results. That is normal, and it is why the routing in section three rests on call shape rather than on these numbers. If you want a figure you can act on, record the silence your current agent produces on your ten most common intents and test both models against it. Our AI transformation practice runs that comparison as the first week of a voice-agent build.