ai-toolsgeminitutorials

How to Use Gemini 3.8 Live for Real-Time Voice Conversations

A practical guide to Gemini 3.8 Live and Live Extended Thinking: when to use each model, how to try Search Live and Gemini Live, and how developers start with the Live API.

September 17, 2026

Typing a prompt is still the default way most people use AI. Gemini 3.8 Live is built for the other mode: you talk, it talks back, and it can keep working in the background while the conversation continues.

Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September 2026. They are native speech-to-speech models for low-latency dialogue. This guide explains what each model is for, how to try them as a user, and how developers connect through the Live API—using Google’s own product and developer documentation.

What Gemini 3.8 Live actually is

These are not “text models that also play audio.” They take audio, images, video, and text and return audio and text in a live session.

Official model pages list:

Context window: 131,072 input tokens and 65,536 output tokens (about 128K / 64K)
Function calling (Live Extended Thinking: async only)
Search grounding
Interleaved or background thinking, depending on the variant
No image generation, file search, code execution, caching, or structured outputs on these Live endpoints

The pair is based on Gemini 3 Pro and is optimized for latency-sensitive dialogue rather than long offline reports.

Live vs Live Extended Thinking

Pick the model the same way you pick a teammate for a call.

Gemini 3.8 Live (gemini-3.8-live)

Default for most voice agents
Fluid dialogue and visual grounding (it can use what the camera or screen is showing)
Interleaved reasoning without the extra delay of a long think-then-speak loop
Built for scale and cost

Gemini 3.8 Live Extended Thinking (gemini-3.8-live-extended-thinking)

For multi-step work during a live call
Reasons in the background while it keeps speaking
Google cites #1 on Artificial Analysis’ Speech to Speech Quality Index (82.6), 68.6% on τ-Voice, 35.1% on Sierra’s τ-Voice-banking, and 97.7% on Big Bench Audio
Used for heavier product surfaces such as conversational Gmail search, Docs Live drafting, Keep Live notes, and the Gemini Live rollout

If the user is asking “what am I looking at?” or “keep chatting while you look that up,” start with 3.8 Live. If the task is “plan this, call tools, then keep talking,” use Extended Thinking.

Try it as a user: Search Live

Google is rolling 3.8 Live into Search Live in the Google app.

Open the Google app on a supported phone.
Tap the Live control.
Ask the question out loud. You can switch languages mid-conversation.
Listen to the spoken answer. Links appear on screen so you can open sources.
Open the transcript if you want the same conversation as text later.

Useful prompts for Search Live:

“Compare the official Gemini 3.8 Live and 3.8 Live Extended Thinking docs and tell me which one fits a customer-support voice bot.”
“Walk me through setting up a passkey on this Android phone while I follow the screen.”
“Explain this error message I’m pointing the camera at.”

Treat on-screen links as the source of truth. Live answers can still be wrong; the links are how you check.

Try it as a user: Gemini Live

Gemini 3.8 Live Extended Thinking is rolling out to Gemini Live and related Workspace live surfaces (Gmail Live, Docs Live, Keep Live), according to Google’s announcement coverage and product posts.

Practical habits:

Start the live session only when you can speak clearly and grant microphone (and camera, if you need visual context).
Say the goal in one sentence first: “Help me draft a reply to this email, then tighten the subject line.”
Interrupt when the model goes off-track. These models are built for turn-taking, not monologues.
Switch to text or export when you need a document you can edit.

Do not read passwords, one-time codes, or confidential customer data into a live session you would not type into Gemini Apps.

Developer path: Live API in AI Studio

Google made both models available to developers in the Gemini API and Google AI Studio on launch day. Pricing published with the developer blog is $0.005 per minute of audio input and $0.018 per minute of audio output.

1. Open the Live surface

Go to Google AI Studio.
Sign in with a Google account that can use the Gemini API.
Open a Live / realtime session and select gemini-3.8-live or gemini-3.8-live-extended-thinking.

2. Choose the model string in code

Everyday voice agent: gemini-3.8-live
Complex agent that must think while speaking: gemini-3.8-live-extended-thinking

If you are migrating from gemini-3.1-flash-live-preview, change the model string to gemini-3.8-live. Google’s model page says thinking_level / thinking_config is not supported on 3.8 Live—omit those fields from session setup.

3. Design for async work

Key capabilities from the developer announcement:

Asynchronous function calling: tools can run while audio still streams to the user
Visual context: live image or video frames can ground the reply
Search grounding: answers can be tied to web results when you enable it

For Extended Thinking, turnComplete: true does not mean the server is idle. Background reasoning or tool calls can still arrive. Keep the WebSocket (or Live session) open and keep handling server messages.

4. A concrete agent sketch

A support voice bot that should stay on the line:

Session model: gemini-3.8-live-extended-thinking
Tools: lookup_order, create_ticket, search_help_center
System instruction: speak in short turns, confirm the order ID, never invent refund policy
Enable search grounding only for public help articles
Stream audio back immediately; run lookup_order in the background
When the tool returns, speak the result instead of restarting the call

A camera helper that names parts on a desk can stay on gemini-3.8-live with video input and no heavy tool graph.

What these models do not do

Do not expect the Live endpoints to replace every Gemini feature.

They do not currently support, per the official model tables:

Image generation
Code execution
File search
Context caching
Structured outputs
URL context
Grounding with Google Maps
Batch API

For cited long reports, use Deep Research or a text Gemini model. For offline transcripts of existing audio files, Google also launched Gemini 3.5 Transcribe (85+ languages) alongside the Live models—that is a different product than a live conversation.

Limits worth planning around

Availability is a rollout. Search Live, Gemini Live, and Gemini Enterprise preview do not all flip on for every account on the same hour.
Enterprise access was announced as private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience to follow.
Safety and accuracy still apply. Google’s model card describes intended use as real-time spoken help for users, developers, and enterprises—not unsupervised legal, medical, or financial decisions.
Session design matters more than prompt poetry. Interruptibility, tool timeouts, and “keep listening after turnComplete” are the difference between a demo and a product.

Conclusion

Gemini 3.8 Live is the everyday real-time voice model. Gemini 3.8 Live Extended Thinking is the same idea with background reasoning for harder tasks. Users can try the first in Search Live and the second as it rolls through Gemini Live and Workspace live features. Developers should start in AI Studio, pick the correct model string, drop unsupported thinking configs, and treat async tool calls as part of the conversation—not a side queue.

If you only need a written brief with citations, stay in text or Deep Research. If you need a voice that can see, search, and keep talking, these September 2026 Live models are the official path.