Google Adds Live Avatar to Gemini Enterprise

Google made Gemini 3.8 Live's avatar feature generally available in Gemini Enterprise on September 24, 2026. It adds a real time video face to the model's voice conversations: lip sync, facial expression and turn taking, generated live rather than pulled from a canned clip library, with US and EU endpoints and enterprise data governance from day one.

What Actually Shipped

Live Avatar sits on top of the speech to speech model Google shipped as Gemini 3.8 Live back on September 16. That earlier release was audio only: low latency voice, background tool calling, strong scores on conversational benchmarks. This update adds a face. Google's own announcement describes precise lip syncing, natural expressions and fluid turn taking, with enterprises able to pick a human-presenting or animated character appearance for the avatar.

Every frame of generated video and every second of generated audio carries a SynthID watermark, imperceptible to a viewer but detectable by Google's tools. That matters more here than in most launches: a live, lip synced face saying things on a company's behalf is a much easier thing to fake or misuse than a text chatbot, and Google is clearly aware customers will ask what stops a competitor from claiming a deepfake was a real support call.

Why This Is Aimed At Support Desks, Not Demos

The stated use cases are customer service and interactive walkthroughs, the kind of interaction that today runs through a phone tree, a text chatbot or a human agent watching a queue. A lip synced avatar does not make the underlying model smarter. What it changes is the interface: a returning customer talking to a face that tracks their eyes and responds in real time reads as a different kind of interaction than typing into a box, for better or worse depending on the audience.

That is also where the risk concentrates. A support avatar that mishandles a refund policy or invents a shipping date is the same failure mode a text-based customer support chatbot already has, just delivered with a face that makes the answer feel more authoritative than the underlying model actually is.

What Builders Should Take From This

• Availability is US and EU endpoints only at launch, through Gemini Enterprise, with provisioned throughput rather than a self-serve API key. This is not yet something a solo builder wires up over a weekend.

• 97 languages is the headline number Google is using for the underlying voice model, not a new claim specific to the avatar layer.

• If a competitor's product starts leading with a talking avatar in the next few weeks, this is almost certainly the stack underneath it.

For anyone already running a voice interface through Gemini, the practical question is whether a face actually improves task completion for your users or just adds latency and an extra thing that can look wrong. Google has not published task completion data for the avatar specifically, only the underlying voice benchmarks from the September 16 release. Treat the video layer as unproven until someone publishes numbers on it.

For the other side of this, what changes for builders when a lab cuts prices instead of adding features, see AI model price cuts: what changes for builders.

FAQ

Is Gemini Live Avatar available through the public Gemini API?

Not at launch. Google is shipping it through Gemini Enterprise with provisioned throughput and enterprise compliance controls, not as a self-serve API product.

Does the avatar change what the underlying model knows or how it reasons?

No. It is a rendering layer on top of the existing Gemini 3.8 Live voice model. Model quality and tool calling behavior are unchanged from the September 16 release.

How is Google preventing the avatar from being used to fake a real person?

Every generated audio and video stream carries a SynthID watermark, which is detectable with Google's tools even though it is not visible to a viewer.