Google gives Gemini 3.8 Live a lip-synced face for business use

Google has given its real-time voice model a face. Gemini 3.8 Live with Live Avatar, announced on Thursday, September 24, generates a video persona that lip-syncs and changes expression as it talks, and it is available to business customers through Gemini Enterprise.

It arrives a week after the launch of Gemini 3.8 Live itself, a model that takes in streams of audio, video and text and answers aloud. The avatar is the new layer. In Google's words, it "creates an experience that listens, sees, and speaks with a dynamic visual persona".

What the avatar adds

Google says Live Avatar handles lip-sync, facial expressions and turn-taking, and adapts across 97 languages. The underlying model keeps talking while it runs tools and API calls in the background, and it can take in live camera feeds and screen shares alongside audio. The avatar version is generally available with US and EU endpoints, while a variant with longer reasoning, Gemini 3.8 Live Extended Thinking, stays in private preview.

The early customers point to sales and service. Cox Automotive built a shopping assistant for Autotrader that highlights parts of the screen and calls tools while it guides car buyers through search, comparison and financing. Equal AI, whose assistant handles over a million live calls a day across nine Indian languages according to chief executive Akhilesh Damaraju, credits Gemini 3.8 Live with better interruption handling and tool-call reliability. That endorsement is for the voice model; Google did not say whether Equal AI uses the avatar.

The limit sitting in the model card

The detail most coverage missed is in Google's model card, not the launch posts. It says the avatar versions "can support a few minutes of continuous interaction, rather than extended hours". That suggests the product fits a showroom greeter or a short support call better than a long tutoring session or an all-day kiosk.

The same card says that, to assess the audio models for frontier safety, Google relied on its evaluations of Gemini 3.7 Flash. The launch posts give no latency figures and no benchmark scores. Pricing sits on Google Cloud's pricing page rather than in the announcement.

Guardrails for a face that can say anything

A convincing, lip-synced face that speaks 97 languages is also what a fraudster would want. Google's answer comes in two parts. Customers pick from a library of pre-built avatars, and "custom avatar creation is gated behind a strict enterprise allowlisting and verification process". And, Google says, "all generated audio and video streams carry imperceptible SynthID watermarks", SynthID being its hidden marker for AI-generated media.

Those are real controls, but they work behind the screen. A watermark helps only if someone checks for it, and a shopper talking to an avatar on a car website has no way to do that. The launch posts do not say whether businesses must tell people they are talking to an AI.

Who this is for

Engadget's Steve Dent suggested that, given "people's distaste for AI slop", the avatars "may not be universally popular". The business case is clearer than the consumer one: a company that already runs voice agents at scale can now add a face without building its own video model.

What to watch: whether Google stretches the session limit beyond a few minutes, whether custom avatars open up beyond the allowlist, and whether any regulator asks for on-screen disclosure when the face on the screen is generated.