← Back to feed

One week after Gemini 3.8 Live launched as a real-time dialogue system, Google added a visual layer. Gemini 3.8 Live with Live Avatar is generally available in Gemini Enterprise as of September 24, pairing near real-time video generation with the existing speech-to-speech dialogue model.

The feature ships three distinct capabilities on top of the base Live model:

Multimodal presence. The avatar processes both visual and audio input simultaneously, generating synchronized facial animation — precise lip-sync, natural expressions, fluid turn-taking — in near real time. Google describes this as enabling enterprises to expand virtual offerings with a richer, more accessible interface than audio alone.

Asynchronous tool execution. When a task requires an external data fetch, the avatar continues active dialogue rather than going silent. A hotel check-in demo shows the model calling a property management system in the background while maintaining conversation; the visual presence does not drop during the tool round-trip.

97-language lip-sync. The avatar dynamically adapts its lip-sync and expressions when transitioning between languages mid-conversation, without video drift or quality degradation. Google targets multinational customer service deployments where language switches happen within a single session.

Customisation and availability

Beyond a preset avatar library, organizations can generate custom Live Avatars from a high-quality reference image. The output preserves reference likeness and brand styling while being fully animated and responsive. Custom avatar creation is currently available only through enterprise allowlisting, not the standard Gemini Enterprise tier.

All audio and video output is watermarked with SynthID, Google’s imperceptible content provenance signal, designed to keep AI-generated video detectable in downstream distribution.

Live Avatar is positioned explicitly at enterprise deployments: customer service agents, interactive product walkthroughs, and virtual staff interfaces. Access is through Gemini Enterprise on Google Cloud. The underlying model card, covering safety and responsible deployment, is published at DeepMind’s model hub.

Context

Gemini 3.8 Live launched September 15 alongside an Extended Thinking variant that took the top-ranked speech-to-speech quality position on Artificial Analysis with an 82.6 score; the standard Live model scored 76.0. Audio input is priced at approximately $0.005 per minute ($3.00 per million tokens); audio output at approximately $0.018 per minute ($12.00 per million tokens). The Live Avatar add-on extends that foundation into the video domain without requiring a separate model call — the same dialogue model generates both speech and avatar frames.

Google’s approach of building avatar generation natively into the dialogue model, rather than layering an animation system on top of a separate TTS output, is the key architectural point. The result is lower latency than composited pipelines and preserves the turn-taking behavior the base model already handles. How that holds under production load at scale has not been independently verified.