Google ships a talking video avatar on Gemini 3.8 Live, generally available to enterprise customers

The mechanism is a pairing, not a new base model. Gemini 3.8 Live already did native speech-to-speech dialogue — audio in, audio out, with no intermediate text transcription. Live Avatar bolts near-real-time video generation onto that same loop, so the model generates the frames of a face in step with the speech it is producing. The lip-sync adapts per language across 97 supported languages, which is the harder part of the problem: mouth shapes for Mandarin and mouth shapes for German are not the same animation.

Two additions matter more than the face. Google says the model can execute tools asynchronously in the background while the dialogue keeps flowing, which is the difference between an avatar that goes silent while it looks something up and one that keeps talking. And Live Avatar carries live visual understanding — it can take a camera feed or a screen share as input. Together that is a video-call-shaped agent that can see what you are showing it and act while it talks.

The rollout is deliberately fenced. GA covers US and EU endpoints inside Gemini Enterprise. Creating a custom avatar is not open — it requires enterprise allowlisting. And Gemini 3.8 Live Extended Thinking, the slower and more deliberative variant, remains in private preview, so the generally available product is the fast conversational one rather than the reasoning one.

What Google did not publish is the commercial and performance detail. The Cloud post points readers at a pricing page but states no figures itself, and gives no latency numbers at all. The only performance signal offered is a customer testimonial about improvements to voice activity detection and overall latency — an impression, not a measurement, which is why no latency figure appears here.

• Confirmed Gemini 3.8 Live with Live Avatar is generally available in Gemini Enterprise as of 24 September, on US and EU endpoints. Google Cloud
• Confirmed Live Avatar lip-syncs across 97 languages with automatic language detection, and outputs carry SynthID watermarks. Google
• Confirmed Custom avatar creation is restricted to enterprise allowlisting, and Extended Thinking remains in private preview. Google Cloud