Breaking

Google Rolls Out Gemini 3.8 Live Voice Models for Real-Time AI Agents

September 15, 2026 at 15:28 EDT

Foundation Models
AI Agents

Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, making two real-time voice models generally available for developers and Gemini users. The release is aimed at voice agents that can keep a natural conversation moving while reasoning, visual understanding and tool calls happen in the background.

Sep 15, 2026 | Google

Gemini voice agents move to GA

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking bring real-time dialogue, visual context, background reasoning and asynchronous tool calls to production voice experiences.

82.6

overall speech-to-speech index

68.6%

agentic task completion

97+

languages in live dialogue

128K

input context window

Audio minute pricing, drawn to scale

$0.005/minaudio input

$0.018/minaudio output

Each block represents $0.001 per audio minute.

The live-agent loop

Listen

Audio, text, images and video enter the same live session.

Reason

Extended Thinking can continue background reasoning after a turn.

Act

Asynchronous tool calls can run while audio keeps streaming.

Continue

The conversation stays active while work finishes.

Why teams will test it

Real-time visual context, multilingual switching and non-blocking tool calls target production voice agents.

What still needs care

Clients must handle async reasoning states; caching, code execution, file search and structured outputs are not supported.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.