Gemini 3.8 Live review 2026. Honest scoring on Google's real-time voice models — visual grounding, 97 languages, Extended Thinking, pricing, and who it fits.
Gemini 3.8 Live is one of the strongest real-time voice models yet — natural full-flow dialogue, background tool calls, visual grounding, and mid-conversation switching across 97 languages, at a reported price well below rival live models. As a voice interface it earns high marks. The honest limit is scope: it produces an interaction, not an artifact. It exports no post, video, or graphic and publishes nowhere. Score it as the excellent voice model it is, not the content tool it isn't.
Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026 as live audio models built for real-time spoken interaction. The design goal is a voice agent that feels like a conversation rather than a turn-based assistant: the models can run tool and API calls in the background while they keep talking, take in a live camera or screen feed and reason about it in near real time ("visual grounding"), and automatically switch between 97 languages mid-conversation.
This review scores Gemini 3.8 Live for what it is: a real-time voice model. I run a competing content engine, so the disclosure is upfront — Kompozy is a generation + publishing tool — and I won't understate how good the conversation is, because it's among the most natural real-time voice yet, nor overstate its usefulness for making content, because that isn't the job it does. Two variants shipped: the cost-efficient 3.8 Live and the reasoning-heavy Extended Thinking, which Google says reasons and speaks at the same time and narrates progress out loud.
The genuinely new thing beyond the conversation feel is the price-to-performance story. Google reports Extended Thinking topping Artificial Analysis' Speech-to-Speech Quality Index at 82.6, ahead of rival live models, with strong agentic scores on the τ-Voice benchmark, while third-party coverage put input-audio pricing well under competitors. Everything below reflects Gemini 3.8 Live at its launch state on 2026-09-15, verified against Google's own announcement; benchmark figures are vendor-reported until independently tested, and pricing wasn't headlined by Google, so confirm current API rates.
Gemini 3.8 Live is a set of live audio models that power real-time voice experiences across Google's surfaces. Because they're built for the flow of dialogue, they can execute tool and API calls in the background while continuing to speak, react to a live visual feed in near real time, and detect and transition between 97 languages without being told to. The 3.8 Live tier is tuned for scale and cost efficiency; the Extended Thinking tier targets high-complexity, multi-step tasks and adds a "reasons and speaks simultaneously" behavior with early verbal cues like "Let me check that…" and spoken progress narration. All AI-generated audio carries SynthID watermarking. It is a voice interface, not a content product. Gemini 3.8 Live converses, reasons, sees, and runs tools, but it produces no exportable deliverable: no captioned video, no carousel, no blog or newsletter, no image, and no scheduled posts. The output is a spoken exchange. It rolled out to developers via the Gemini API and Google AI Studio, to enterprises via a private preview in Gemini Enterprise, and to general users via Search Live, Gemini Live, and Extended Thinking inside Google Workspace (Docs, Gmail, Keep) for Pro and Ultra subscribers.
The clearest fit is anyone who wants a fast, natural, hands-free way to talk to an AI — to think out loud, run visual walkthroughs, and get spoken answers with background tool calls handled mid-flow. It's especially strong for multilingual creators who want to dictate or brainstorm in more than one language, and for developers building voice agents where cost per hour matters. For creators specifically, it's an excellent ideation, dictation, and research partner — the front of the workflow. Where it fits poorly is the back of the workflow: producing and publishing content. Gemini 3.8 Live writes no shippable copy, makes no video or graphics, governs no brand voice for an audience, and posts to nothing. If your bottleneck is turning an idea into on-brand posts across platforms, a voice model — however good the conversation — leaves that entire job undone, and you'll want a content engine like Kompozy for it.
| Dimension | Score | Why |
|---|---|---|
| Real-time conversational voice | 4.5 / 5 | Full-flow dialogue with background tool calls means it keeps talking while it works — among the most natural real-time voice yet. |
| Visual grounding | 4.2 / 5 | Reasoning about a live camera or screen feed in near real time is a genuine multimodal upgrade over audio-only assistants. |
| Multilingual (97 languages) | 4.3 / 5 | Automatic detection and mid-conversation switching across 97 languages is a real edge for multilingual creators. |
| Extended Thinking reasoning | 4.1 / 5 | Reasoning and speaking simultaneously, with progress narration, keeps complex multi-step tasks from going silent — Google reports category-leading agentic scores. |
| Price-to-performance | 4.4 / 5 | Reported input-audio pricing well below rival live models makes real-time voice practical at volume, if the figures hold up. |
| Availability & rollout | 4.0 / 5 | Developers, enterprise preview, and consumer surfaces at once — broad reach across the Gemini API, Workspace, and Search Live. |
| Safety / provenance | 3.8 / 5 | SynthID watermarking on all generated audio is a sensible provenance measure at launch. |
| Usefulness for content production | 1.5 / 5 | Not a content tool — it produces no exportable copy, video, or graphics and publishes nowhere. |
Google didn't headline a consumer price in its announcement, so the numbers to reason about are the API rates that surfaced in third-party coverage: roughly $0.84 per hour of input audio for Gemini 3.8 Live and around $3.50 per hour for Extended Thinking. Measured against the rival live models it was benchmarked against — reported in the same coverage at several dollars more per hour — that's an aggressive position, and it's the most interesting part of the launch. Confirm current rates in the Gemini API docs before you build a cost model, as pricing changes.
For a real-time voice model, that pricing is strong value: category-leading reported quality on the reasoning tier and a base tier cheap enough to run voice agents at volume. Judged inside the voice-model category, it's a compelling option, particularly for developers and enterprises building customer-experience agents.
The framing only breaks if you try to price Gemini 3.8 Live as a content tool. It produces nothing you can export or publish, so the spend buys you a better way to talk to and build with an AI — not a caption, a video, or a scheduled post. Turning a Gemini 3.8 Live conversation into finished, on-brand content across platforms still costs you, in time or in tools, for the writing, the formats, the brand-voice layer, and the distribution.
| Use case | Fit | Why |
|---|---|---|
| Hands-free brainstorming and thinking out loud | Strong | Full-flow, low-latency conversation is exactly built for fast, natural back-and-forth while your hands are busy. |
| Multilingual dictation and ideation | Strong | Automatic mid-conversation switching across 97 languages makes it a genuine tool for creators working more than one audience. |
| Live visual walkthroughs of a product or edit | Strong | Visual grounding lets it reason about a camera or screen feed in near real time as you narrate. |
| Building a voice agent for customer experience | Strong | Background tool calls, reasoning, and low reported cost per hour are aimed squarely at this — enterprise access is in preview. |
| Researching a content angle before you make it | OK | It sharpens the idea and can run tools to check facts, but the research still has to be written up and produced elsewhere. |
| Producing captions, scripts, or posts | Weak | It speaks answers; it drafts no exportable, publishable copy and makes no graphics or video. |
| Building a consistent brand voice across platforms | Weak | There is no Persona Brief or governance layer — nothing it says is held to a brand voice for an audience. |
| Scheduling and publishing content | Weak | It publishes nowhere and has no scheduler; distribution is entirely outside its scope. |
Scored on its own terms, Gemini 3.8 Live is a very good real-time voice model, and Kompozy isn't trying to be one — the two sit at opposite ends of the same workflow. The cleanest way to see the boundary is the word "artifact." Gemini 3.8 Live produces an interaction: a spoken exchange that's fast and natural precisely because it never stops to render anything you can keep. Kompozy produces artifacts — a carousel, a blog, a newsletter, text posts, and persona or avatar video — each one a file you can post. A conversation, however good, isn't a deliverable; the moment you need something an audience can see, you've crossed from Gemini's job into Kompozy's.
The second boundary is governance. Gemini 3.8 Live has no concept of your brand voice; it answers however the model answers. Kompozy runs everything through a [Persona Brief](/glossary/persona-brief) and banned-word filters, so a batch of output reads as one consistent brand rather than raw model text — then schedules and publishes it across nine platforms plus blog and email. The honest read is that they compose rather than compete: dictate and interrogate the idea with Gemini 3.8 Live, then run the transcript through Kompozy to make it on-brand, turn it into video, and ship it. Where Gemini's job ends at a spoken answer, Kompozy's begins.
As a real-time voice model, yes — it offers natural full-flow conversation, visual grounding, 97-language switching, and Google-reported category-leading reasoning on the Extended Thinking tier, at a reported price below rival live models. It's not worth judging as a content tool, because it produces nothing you can export or publish; it only holds a better conversation.
3.8 Live is the cost-efficient tier for fluid dialogue and visual grounding. Extended Thinking is the higher-intelligence tier for multi-step reasoning — Google says it reasons and speaks at the same time, gives early cues like "Let me check that…," and narrates progress out loud. Extended Thinking is reported to cost more per hour of input audio.
Google didn't headline a price, but third-party coverage put Gemini 3.8 Live at roughly $0.84 per hour of input audio and Extended Thinking around $3.50 per hour — below the competing live voice models it was benchmarked against. Confirm current rates in the Gemini API pricing docs.
They're close competitors in the real-time voice category. Google benchmarked Extended Thinking as topping the Artificial Analysis Speech-to-Speech Quality Index ahead of GPT-Live, at a lower reported price per hour of input audio. Both are conversational voice models, not content tools — neither produces captioned video, carousels, blogs, or scheduled posts.
No. It is a conversational voice interface — it talks, reasons, sees, and runs tools, but it produces no captions, carousels, images, blogs, newsletters, or video, and it publishes nowhere. To turn a spoken idea into finished, on-brand posts across platforms, you need a content engine like Kompozy.
Visual grounding is the ability to take in a live camera or screen feed and reason about what it sees in near real time — so you can point at a product, a rough edit, or a whiteboard and talk through it, and the model responds to the visual context rather than audio alone.
Yes. Google says all AI-generated audio from these models carries SynthID watermarking, its provenance system for marking AI-generated media, aimed at reducing misuse and making synthetic audio identifiable.
They solve different halves of the workflow. Use Gemini 3.8 Live to talk through, dictate, and research an idea by voice; use Kompozy to turn that idea into a carousel, blog, newsletter, video, and text posts in your brand voice, then schedule and publish across nine platforms. Many creators brainstorm by voice and produce and ship in Kompozy.