Gemini 3.8 Live Explained: What Changes With the New Real-Time AI
Talking to an AI assistant has always meant tolerating an awkward pause while it fetches an answer or runs a tool. Google's Gemini 3.8 Live, launched September 15, 2026 alongside a heavier Extended Thinking variant, is built specifically to close that gap — letting Gemini reason, call tools, and keep talking in the same breath instead of going silent to think. Here's what actually shipped, and where it's already live.
Gemini 3.8 Live is already powering Google Search Live's voice conversations as of its September 15 rollout.
Two Models Built for Talking, Not Just Answering
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are both voice-first models, but built for different jobs. The standard 3.8 Live model is tuned for efficiency and scale — fast, cheap, fluent conversation with visual grounding, meant to run at high volume for everyday voice interactions. Extended Thinking is the heavier sibling, built for multi-step reasoning tasks where the model needs to actually work through a problem while still talking, narrating its progress out loud instead of going quiet.
Both models can process visual input in near real time, meaning a user can point a camera at something and keep talking while Gemini reasons about what it's looking at, and both can detect and switch between 97 supported languages mid-conversation without the user having to restart or pick a language up front.
Thinking and Speaking at the Same Time
The core technical claim behind Extended Thinking is parallel reasoning and speech: instead of pausing to "think" and then delivering a finished answer, the model can narrate what it's doing while a background process runs — calling an API, checking a tool, working through a multi-step calculation — using natural filler like "let me check that" to keep the conversation from stalling out. Google's own demonstrations reportedly show the model converting a rough paper sketch into a working React component and managing multi-step tasks while continuing to talk the user through what's happening.
Background Tool Use
API calls and tool actions run while the conversation continues, instead of freezing the dialogue until they finish.
Mid-Conversation Language Switching
The model detects a language change automatically rather than requiring a manual reset.
Visual Grounding
Camera input is processed close to real time, so a spoken question about something visible doesn't lag noticeably behind the image.
Progress Narration
Extended Thinking speaks its intermediate steps aloud on complex tasks instead of leaving a silent gap.
Search Live, Workspace, and the Gemini API
Google Search Live — its voice-and-video conversational search feature — moved onto Gemini 3.8 Live on the same day it launched, meaning the upgrade was already reaching regular users rather than staying developer-only. The models are also rolling into Google Workspace's voice features and are available through the Gemini API and Google AI Studio for developers building their own voice agents, at a published rate of $0.75 per million input tokens and $4.50 per million output tokens for text, with separate audio pricing.
Bottom Line
Gemini 3.8 Live is a meaningful step for voice AI specifically because it targets the part of the experience people actually notice: the dead air while a model works. Whether it's a big leap depends on how it holds up outside a curated demo — parallel reasoning-while-speaking is a genuinely hard problem, and how naturally it handles a messy real conversation, not a scripted one, is the real test.
How This Compares to Other Real-Time Voice AI
Real-time voice AI has been a competitive front all through 2026, with OpenAI's voice mode and various Alexa-successor products all chasing the same goal: a conversation that feels like talking to a person rather than issuing commands to a device. Gemini 3.8 Live's specific claim — reasoning and speaking in parallel rather than pausing to think — targets the single most noticeable flaw in earlier voice assistants, the dead air while a system works. Whether Google's approach genuinely solves that problem more thoroughly than competitors, or simply narrows the gap in the specific demos Google chose to publish, is something that will only become clear once more people use it in ordinary, unscripted conversations rather than curated showcases.
What Developers Should Test Before Relying on It
For a developer building on the Gemini API specifically, the headline "reasoning and speaking simultaneously" claim is worth testing directly against a real use case rather than assumed from a demo. Background tool calls that complete quickly are one thing; a genuinely long-running lookup or a complex multi-step API chain is a harder test of whether the parallel-speech design actually holds up without the model either going quiet anyway or narrating in a way that feels repetitive rather than genuinely informative. Given that pricing is usage-based and separate for audio versus text, it's also worth modeling real conversation-length costs before committing a production voice product to the new models over the previous generation.
The Core Facts, in One Place
In short: Gemini 3.8 Live and its Extended Thinking variant launched September 15, 2026, built specifically to let Gemini reason, use tools, and keep speaking in the same breath rather than pausing silently, while supporting 97 languages and near real-time visual input. The standard model targets fast, efficient everyday conversation; Extended Thinking targets complex, multi-step tasks. Both are already live in Google Search Live and rolling into Workspace and the developer-facing Gemini API.
More AI Guides, Written Clearly
Browse our full library of AI tool guides, updated regularly as models and features change.
Common Questions
What is Gemini 3.8 Live?
It's a real-time voice AI model Google launched on September 15, 2026, built to hold fluid spoken conversations, process visual input near real time, and switch between 97 languages mid-conversation.
What is Gemini 3.8 Live Extended Thinking?
It's a heavier variant released alongside the standard model, designed for complex, multi-step tasks. It can reason and speak simultaneously, narrating progress aloud instead of pausing silently to think.
Where can I actually use Gemini 3.8 Live?
It's already powering Google Search Live's voice conversations as of its launch day, and it's rolling into Google Workspace's voice features as well as being available to developers through the Gemini API and Google AI Studio.
How much does Gemini 3.8 Live cost via the API?
Published pricing lists $0.75 per million input tokens and $4.50 per million output tokens for text, with separate rates for audio input and output.
Is Gemini 3.8 Live available to regular consumers, or only developers?
Both. It's live for consumers immediately through Google Search Live and is rolling into Google Workspace, while also being available to developers through the Gemini API and Google AI Studio.
Does Gemini 3.8 Live replace the standard Gemini 3 models?
No. It's a separate, voice-focused model line built specifically for real-time spoken conversation, existing alongside Gemini 3 Pro, Flash, and Deep Think rather than replacing them.
Will Gemini 3.8 Live replace the previous Gemini Live model entirely?
Google has not announced a hard cutoff for the previous Gemini Live model as of this writing; the practical rollout has focused on bringing 3.8 Live into existing surfaces like Search Live rather than an abrupt full replacement.
Can Gemini 3.8 Live be used offline?
No. Like other Gemini Live models, it requires an active internet connection to process audio, visual input, and tool calls in real time.
Does Gemini 3.8 Live support group conversations with multiple speakers?
Google's announcement materials focus on one-on-one conversational use; multi-speaker handling wasn't a headline feature detailed in the initial September 15, 2026 launch.
Does Gemini 3.8 Live work on all devices?
It's available through Google's own apps like Search and Workspace as well as through the Gemini API for developers, so practical availability depends on which surface a person or developer is using rather than being tied to one specific device type.
What happens to conversations after they end?
Standard Gemini data handling and privacy controls apply to Live conversations the same way they do to other Gemini interactions, and users can review Google's account privacy settings for specifics on retention.
More from the creator
Explore two other projects from the same creator.