Draft poster · the finished cut will appear here
What happened
Google announced Gemini 3.8 Live with Live Avatar as a visual layer for live dialogue. Google says the system produces a near-real-time video persona that can listen, see, and speak, and that the feature is available through Gemini Enterprise rather than the free consumer tier.
How it works
The avatar is intended to keep talking while tools run asynchronously in the background and return results mid-conversation. Google says its speech-to-speech system synchronizes voice, lip movement, and expressions across 97 languages. Organizations can choose preset characters or generate a custom avatar from a reference image; custom creation currently requires enterprise allowlisting, according to Google.
Google also says generated output carries an imperceptible SynthID watermark. That provides a detection signal after generation, but it does not turn the avatar into proof that a human is present or that every part of the conversation is real-time.
The combination is designed to hide waiting time without hiding the fact that a tool call happened. In principle, the avatar can keep a conversation moving while the underlying agent pauses to fetch information or complete an action.
Why it matters
An avatar is becoming part of the agent interface itself. Visual presence and tool latency now have to work together: the system needs a way to stay responsive while its workflow branches into search, actions, or other tools.
What we don’t know yet
The 97-language, lip-sync, expression, and near-real-time quality claims come from Google. Custom avatars are allowlisted, and the announcement does not establish consumer-wide availability, independent quality testing, or proof of human identity.
Source: Google Blog, Shuo-yiin Chang and CJ Zheng, September 24, 2026 - original announcement