Gemini 3.8 Live and 3.8 Live Extended Thinking

TL;DR Highlight

Gemini 3.8 Live models run background tools while still talking in voice

Who Should Read

A backend developer building a voice agent or phone support bot who is choosing a real-time voice model

Core Mechanics

Two Live models launched
- 3.8 Live targets scale and low cost and accepts visual input.
- 3.8 Live Extended Thinking targets complex multi-step reasoning.
Tools run while the model keeps talking
- Tool and API calls execute in the background during the conversation.
- The model acknowledges the request and keeps chatting until the task finishes.
Extended Thinking reasons and speaks at once
- It starts with early verbal cues like "Let me check that…".
- It narrates live progress on multi-step background tasks.
Voice benchmark numbers were published
- Extended Thinking ranks #1 on Artificial Analysis Speech to Speech Quality Index at 82.6.
- It scores 68.6% on τ-Voice, 35.1% on Sierra τ-Voice-banking and 97.7% on Big Bench Audio.
3.8 Live ranks second in Speech Agent Arena

Community Reactions

Niche languages work well, per one user
- A US-based user chatting in Afrikaans while driving said it was good enough to shock family members.
Works on Workspace accounts with low latency, per one user
- They said thick accents were handled well and voices were pleasant.
- They could use it on a Workspace account, unlike recent releases.
Warning against agentic use
- A voice-assistant builder switched to Gemini 3.8 Flash and reverted the same day because it invented requirements and implementation details.
- They advise caution using Live for agentic scenarios.
Telephony and stability issues reported
- One commenter said missing SIP support forces WebSocket forwarding that adds noticeable latency.
- Another reported an infinite self-reply loop and random language switching.
Missing MCP tool support in voice
- One commenter complained voice mode lacks the tool parity of text mode.

How to Apply

Test voice chat in AI Studio today
- Pick 3.8 Live in Google AI Studio and speak with accents, language switches and interruptions.
- Compare latency against your current voice model using the same scenario.
Build a background tool-call scenario
- Pick one slow function call like booking or lookup and register it as async.
- Check that the conversation stays smooth while it runs.
Run an eval set before agentic voice use
- Write delegation test cases that check whether the model invents unrequested requirements.
- Attach it to production voice agents only after it passes.
Check SIP support for phone integration
- For phone bots, measure the latency of WebSocket forwarding.
- If it is high, compare platform paths like LiveKit or Pipecat.

Try It

copy & run

# Paste into Claude or ChatGPT — test scenario generator before adopting a voice agent

I am building a voice agent for {{service description}}. A real-time voice model must keep talking while making background tool calls.

For each item below, create 3 scenarios I can test by speaking.
1. Acknowledge the request, then run a background task
2. The user interrupts or changes topic mid-task
3. The language changes mid-conversation
4. Check whether the model invents requirements I did not specify

Put a sample user utterance and pass criteria for each scenario in a table.

Next step: run the scenarios by voice in AI Studio and record pass or fail.

Terminology

Live APIA Gemini interface that streams audio and video in real time. It is the channel for talking like a phone call without breaks.

Pareto FrontierThe boundary where improving one metric, like accuracy, means losing another, like conversational quality. Pushing it means improving both.

τ-VoiceA benchmark that measures whether a voice agent completes tasks end to end. It is a work exam for voice agents.

SynthIDAn imperceptible watermark embedded in AI-generated audio. It acts as a stamp to detect AI output later.