Google Cloud announced on September 25 that Gemini 3.8 Live with Live Avatar is now generally available on Gemini Enterprise. The release adds a talking "face" to the existing real-time voice model: a generated video avatar with lip movements synced to speech, aimed at customer service, retail assistance, and product walkthroughs — scenarios that benefit from a human-like presence. Google first previewed it at Google Cloud Next 2026.
What this version can do
According to Google Cloud's blog post, Gemini 3.8 Live is built around five capabilities this time:
- Video avatar: generates a talking avatar with lip-synced speech;
- Speech-to-speech: native voice conversation that lets users interrupt mid-sentence and pick back up without losing prior context or in-progress background actions;
- Asynchronous tool calling: calls APIs or looks things up in a CRM or ERP in the background without interrupting the conversation;
- Multilingual support: understands and speaks 97 languages with automatic language detection;
- Real-time vision: processes camera feed, screen share, and audio simultaneously.
On the deployment side, it is available through US and EU endpoints, supports provisioned throughput, and Google emphasizes enterprise compliance and data governance.
Two limitations are spelled out in the announcement. Custom avatars are currently allowlist-only — enterprises that want to use their own likeness or spokesperson have to clear Google's review first. Every generated audio and video stream embeds an invisible SynthID watermark. Separately, Gemini 3.8 Live Extended Thinking, which adds longer reasoning, remains in private preview and was not promoted to GA alongside this release.
What customers are doing with it
The announcement names four customers. Cox Automotive's Autotrader uses it as a car-shopping assistant: the avatar talks while highlighting listings, comparing models, and walking through financing on the page. Cox Automotive Chief Product Officer Marianne Johnson put it this way:
"Shoppers increasingly expect to describe what they need in their own words rather than work through filters and menus."
India's Equal AI runs a personal-assistant service handling more than one million live phone calls a day across nine Indian languages; the company's CEO said the new version improved reliability around handling interruptions, multilingual conversation, and tool calling. Salesforce says it is combining the model with Agentforce, and Specs, an AI assistant platform, cited improvements in voice activity detection and overall latency. These are all self-reported by the companies — the announcement does not provide comparable figures like latency in milliseconds or conversion rates.
Is this usable for developers in China
Gemini Enterprise and the Live API run on Google Cloud, so developers in mainland China cannot call them directly; teams building for overseas markets can go through the US or EU endpoints. Chinese is among the 97 supported languages, and the most direct use case is customer service and sales assistance for multilingual markets like Southeast Asia and the Middle East.
Domestic digital-human customer service and livestreaming avatars have been running for a few years now, typically by chaining together separate components — speech recognition, a large language model, speech synthesis, and an avatar engine that drives the lip movements. Each stage adds its own latency, and handling interruptions is often the weak point. Google's bet here is that folding listening, speaking, seeing, tool-calling, and video output into a single real-time model gives an edge on latency and interruption handling end to end. Which approach is more cost-effective depends on avatar output pricing and real-world concurrency costs, and there are no public numbers yet to compare.
Google had already shipped Gemini 3.8 Flash TTS and Flash-Lite TTS speech synthesis models, and this announcement also folds Gemini 3.5 Transcribe and Gemini 3.5 Live Translate into the enterprise audio model lineup. With real-time conversation, transcription, translation, synthesis, and now a video avatar, Google's enterprise voice product line is essentially complete.
What is still unclear
The blog post only links to the pricing page for Live Avatar without breaking out a rate for avatar video output, and it does not disclose the allowlist's application requirements or review timeline. It also does not say which third parties, if any, get access to a SynthID detection tool. These are the details that will directly shape cost and compliance decisions for enterprises evaluating whether to deploy it.
Sources: Google Cloud official blog, Android Headlines, CocoLoop, Unite.AI; still to verify: supported language count, allowlist and watermark restrictions, Equal AI's daily call volume, and customer quotes.