Gemini 3.8 Live: what Google’s new voice AI changes

On September 15, Google introduced Gemini 3.8 Live and 3.8 Live Extended Thinking, two new models for voice interaction. The useful question is how an assistant handles an actual conversation: whether it understands corrections, keeps track of constraints and helps you reach a usable result.

What changes in a conversation

According to Google’s developer announcement, the models can run tool calls in the background while continuing the conversation. They also use visual context and support switching between 97 languages. Extended Thinking targets more complex tasks with multiple steps, while standard 3.8 Live focuses on responsive, cost-efficient voice interaction.

The practical appeal is fewer pauses when a task needs clarification. For example, you describe a presentation, change the intended audience and then ask for a shorter outline. A useful result is a coherent outline that reflects the latest constraints. A polished voice alone does not establish that.

Where the rollout has started

The September 15 announcement puts 3.8 Live in Search Live and Extended Thinking in Gemini Live. In Docs, the latter is rolling out to Google AI Pro and Ultra subscribers; in Gmail and Keep, to all Google AI subscribers. Access for Workspace business customers is listed as coming later.

Gemini Live updates arrive gradually. Open Live in the mobile app to start a conversation; camera and screen sharing require a separate action. Seeing the Live button does not, by itself, confirm that your account has received every capability in the new announcement.

What language support means

The advertised language count describes the new models. App availability and individual actions have their own restrictions: Google maintains the full language and country lists and flags languages where some voice commands and actions are unavailable. Check your intended language and feature together, rather than assuming identical behavior everywhere.

A short check for your own workflow

• Give it one small task, such as outlining a talk or explaining a public set of instructions.
• Add a new constraint midway through the conversation. Check whether the result actually changes to reflect it.
• Ask for a recap of the decisions and a list of what remains unknown.
• Verify names, numbers and links before using the answer.

This is a suggested evaluation method, not the result of our own hands-on testing. For choosing a service for regular work, see our guide to comparing ChatGPT, Gemini and Claude.

Before sharing your camera or screen

Google’s Gemini privacy guidance distinguishes recordings from transcripts. Audio, video and screen shares are not used to improve services by default, but transcripts may be used when Keep Activity is on. Conversations may still be retained for up to 72 hours when activity is off. Review your settings and close windows containing passwords, payment information or private messages before a session.