Gemini 3.8 Live gains an avatar on 24 September

Slide graphic for Gemini 3.8 Live with Live Avatar from Google's announcement Caption: Hero slide published with Google's 24 Sep 2026 Live Avatar post · Source: Google Keyword, Gemini models blog · link

On 24 September 2026 Google published Gemini 3.8 Live with Live Avatar. Research scientist Shuo-yiin Chang and software engineer CJ Zheng wrote the post on behalf of the Gemini Audio Team. Google says Live Avatar couples its live dialogue models with low-latency streaming video. The result is an agent that listens, sees, and speaks through an animated persona.

The launch comes one week after Gemini 3.8 Live. Google says the feature is available in Gemini Enterprise from the day of the post. Google pitches it for customer service and interactive walkthroughs. A companion Google Cloud post by product manager Fabien Blanc-paques calls the feature generally available. That post says Google first previewed it at Google Cloud Next 2026.

Enterprise tenants opt in and an allowlist gates custom faces

Google ships and controls the feature. Gemini Enterprise customers decide whether to switch it on for their agents.

  • Product and customer-experience teams choose between Google's library of preset avatars and a custom avatar.
  • Brand and legal teams own likeness decisions for custom avatars. Google offers custom avatar creation only through enterprise allowlisting. The Cloud post adds a verification step.
  • Trust and safety teams should read Google's SynthID claim together with the linked Gemini 3.8 audio model card.
  • Developers can start from the Gemini Live API documentation. The Cloud post lists US and EU endpoints, and it routes provisioned throughput and custom avatar allowlisting through Google Cloud sales.

Background tool calls and lip-sync across 97 languages

Google describes five capabilities.

  • Video on the dialogue path. Live Avatar generates near real-time video with speech, so the agent shows a face while it talks.
  • Audio and visual input together. Google says the system processes what it sees and hears at the same time. It answers with expressive audio and video.
  • Asynchronous tool calls. Google says the avatar can call tools and fetch data in the background while the conversation continues. The post demonstrates this with a hotel guest check-in.
  • Language switching. Google claims lip-sync and expressions adapt across 97 languages when a speaker switches language mid-conversation. Google also claims the switch causes no loss of video fidelity and no visual drift. The post publishes no measurement for either claim.
  • Custom avatars from one image. Developers can animate an avatar from a high-quality reference image. Google says the avatar keeps the reference likeness, brand styling, or character identity.

Google says SynthID watermarks all output from its AI products. According to the post, the imperceptible watermark sits in both the audio and the video stream.

No price, latency figure, or consumer plan in the post

  • The Google blog post states no price, rate limit, session cap, or latency target. The Cloud post links to a pricing page but gives no rates in its text.
  • Both posts name Gemini Enterprise as the surface. Neither post mentions the consumer Gemini app.
  • Google does not publish the allowlist criteria or a date for self-serve custom avatars.
  • Bandwidth, device, and browser requirements for the video stream are not disclosed.
  • The SynthID claim points to the model card. The post gives no detection rate, false-positive rate, or result for watermark removal attempts.
  • Google publishes no latency or quality comparison with other avatar vendors.

Pilot one support flow and confirm allowlist status

  1. If you already buy Gemini Enterprise, pilot Live Avatar on one support or walkthrough flow. Measure audio and video latency and tool-call continuity on your own scripts before you market the feature to customers.
  2. If you need a brand-matched face, ask Google whether your tenant qualifies for the custom avatar allowlist. Get allowlisting and legal sign-off in writing before you promise a likeness-matched avatar.
  3. If you work in a regulated or high-misinformation market, treat SynthID as Google's detectability claim. Check detection tooling and disclosure rules with counsel before you cite the watermark in a compliance file.
  4. If voice alone serves your users, stay on Gemini 3.8 Live without the avatar until you have a clear reason to add video. Read Google's pricing page before you budget for it.