Introducing Gemini 3.8 Live with Live Avatar

Conversation is multimodal. DeepMind’s Live Avatar couples near-real-time video generation with live speech so enterprise agents can listen, see, and speak with a visual persona.

It is available in Gemini Enterprise starting September 24, 2026, with preset characters, custom avatars from a reference image (allowlisted), and background tool calls that keep dialogue moving.

This is the enterprise face of the live-model wave: customer service and walkthroughs that look like a person, not a typing bubble.

I’m watching the product bet, not the demo reel. Live Avatar claims precise lip-sync, natural expressions, and fluid turn-taking, plus native multilingual speech-to-speech across 97 languages without visual drift. For global support desks, that is the difference between a novelty and a staffing plan.

The useful-AI case is straightforward: hotels, retail, and education want agents that stay present while tools fetch data. Asynchronous tool execution while the avatar keeps talking is the right architecture for that.

The harm path is equally clear. A branded, lip-synced face lowers the bar for trust—and for deception. DeepMind puts SynthID watermarks in audio and video and points to a model card. That is necessary. It is not sufficient if enterprises deploy these faces without clear disclosure to the human on the other end.

Custom avatars from a single reference image, even behind allowlisting, raise likeness and brand-abuse questions the moment the allowlist leaks or a partner ships carelessly. We’re for richer, more accessible digital service. We’re not for pretending the avatar is a person.

The post builds on the prior week’s Gemini 3.8 Live launch. Authors Shuo-yiin Chang and CJ Zheng wrote on behalf of the Gemini Audio Team.