Google ships Live Avatar for Gemini Enterprise agents before stating what the avatar costs

Asynchronous tool calling is the part to test first. Google says Live Avatar can trigger tool calls and fetch data in the background while the dialogue continues ^[3]. In a blocking design, the agent goes quiet until the lookup returns. On video, the customer then watches an idle face. The asynchronous path hides the lookup behind speech without making it any faster, and that is still good engineering. Google's hotel check-in demo shows the agent calling tools for a guest while the conversation carries on ^[4].

The feature sits on a dialogue system Google shipped nine days earlier ^[15]. Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking came out on September 15 as models for real-time voice agents and task completion ^[8]. Live Avatar layers video onto that system ^[10]. The launch post was written by research scientist Shuo-yiin Chang and software engineer CJ Zheng of Google's Gemini Audio Team ^[9].

Google Cloud's documentation specifies 24-frame-per-second avatar video and sub-second interaction latency. RuntimeWire notes these are Google's specifications, not independent results ^[14]. Google also says lip-sync and expressions hold across 97 languages, including switches mid-conversation, without visual drift ^[11]. Treat both as measurements of Google's workload. For the latency figure to transfer, it would have to hold on the US or EU endpoint your users actually reach ^[16]. It would also have to hold on a turn with a tool call in flight and your own back end behind the tool.

Preset avatars are open to Gemini Enterprise customers. A custom avatar generated from a high-quality reference image is a separate path ^[5]. Google wrote: "Custom avatar creation is currently available only through enterprise allowlisting." ^[6] Google says all generated audio and video carries a SynthID watermark so it stays detectable as AI output ^[7]. Rights and consent for the face and voice samples behind a custom avatar are the customer's responsibility, according to Google's documentation ^[17].

Google did not disclose a Live Avatar rate or usage limits; its launch materials link to a pricing page, according to RuntimeWire ^[12]. A pilot can still answer the product question. Google Cloud's named customer example is Cox Automotive, which it says built an Autotrader assistant that uses an avatar to help shoppers search, compare vehicles and discuss financing ^[13]. A pilot built the same way, with a bounded task and a few tools, can show whether customers finish the task, though not what a session costs. Google pitches web, mobile and kiosk deployments ^[18], and finance will want a per-session number before any of them goes into production.

In my view, a team already running Gemini 3.8 Live voice agents should pilot a preset avatar now, because the avatar adds video to a dialogue system the team already operates ^[10]. I'd hold custom-likeness work until both the allowlist approval and the consent records are in hand ^[6][17].