Google’s Gemini 3.8 Live gives enterprise agents a lip-synced face in 97 languages

Google on September 24, 2026 made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise, giving its real-time voice model a lip-synced video face. The avatars speak in 97 languages, keep talking while they call APIs in the background, and every generated frame and sound carries Google’s SynthID watermark.

Why it matters: until now, conversational video avatars were mostly sold by specialists such as Synthesia and Soul Machines. Google is bundling one into its enterprise agent platform, which turns the avatar from a product into a feature of a model subscription.

What is Gemini Live Avatar?

It is a video layer on Google’s Gemini 3.8 Live speech-to-speech model. Google’s announcement describes near real-time visual presence with “precise lip-syncing, natural expressions, and fluid turn-taking.” The Google Cloud blog lists what sits underneath: native speech-to-speech dialogue, live understanding of camera feeds and screen shares, and recovery from interruptions without dropping context or a back-end transaction in progress.

The tool-calling detail is the commercial point. An avatar that falls silent while it looks up an order is a worse experience than a text bot. Google says the model fetches data in the background while the conversation continues, and MediaPost reports it can make live API, CRM and ERP calls during a session. Customer service, product walkthroughs and hotel check-in are the use cases Google cites.

Who can create a custom Gemini avatar?

Only allowlisted enterprise customers. Every Gemini Enterprise customer can use a curated library of preset avatars, each with its own look and voice. Generating a custom avatar from reference images, for example a brand character or a real employee’s likeness, requires what Google calls strict enterprise allowlisting and verification, arranged through its sales team. Google says the safeguards are designed to respect identity and keep generated content transparent.

The watermark matters for European deployments. Under Article 50 of the EU AI Act, applicable since August 2, 2026, AI systems must disclose themselves and synthetic content resembling real people must be marked. SynthID handles machine-readable marking; telling users on screen that they are talking to an AI is still the deployer’s job.

How does it compare with Synthesia’s interactive avatars?

The two products start from opposite ends. Google begins with an open-ended conversational model and adds a face. Synthesia begins with studio-quality likeness and adds interaction. In a TechCrunch hands-on published September 26, reporter Dominic-Madori Davis had Synthesia build avatars of herself from photos and a two-minute voice recording at its New York office, a process that took a couple of days. Her interactive version answered questions about one article she had written, and redirected anything off-topic. She described it as deterministic: it only says what it was trained to answer.

That constraint is a feature for regulated corporate training, where Synthesia’s business sits. The company raised a $200 million Series E at a $4 billion valuation in January 2026 and passed $100 million in annual recurring revenue in April 2025, TechCrunch reported. The round was led by GV, Alphabet’s venture arm, so Google is now both an investor in and a competitor to the category leader. We covered Synthesia’s position in our profile of the synthetic presenter business.

What does this mean for avatar companies?

Pressure on the middle of the market. Named launch customers include Cox Automotive’s Autotrader, Salesforce for Agentforce, and Equal AI, which Google says handles more than a million calls a day. Specialists that resell a face on top of someone else’s language model now compete with the model vendor itself. The defensible ground is where the avatar layer needs things Google restricts: custom likenesses, celebrity and brand IP, and tight scripting.