# Google adds avatars to Gemini agents, with Cox building Autotrader's car-shopping assistant > Source: > Published: 2026-09-25 13:39:08+00:00 # Google adds avatars to Gemini agents, with Cox building Autotrader's car-shopping assistant **Gemini 3.8 Live with Live Avatar is generally available in Gemini Enterprise. Google Cloud says Cox Automotive built an AI shopping assistant for Autotrader using the feature.** By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff) ยท Published Primary source: [Engadget](https://www.engadget.com/2268587/google-video-avatars-gemini-3-8-live-agent/) ## Why it matters Google is packaging live speech, video and background tool use as one enterprise agent interface. Google Cloud names Cox Automotive's Autotrader assistant as a customer use case, but provides no deployment scale or independent comparison with voice or text agents. Custom likenesses remain gated, and the launch materials link to a pricing page without stating Live Avatar-specific pricing. Google announced Gemini 3.8 Live with Live Avatar on September 24th, adding generated, speaking video personas to its real-time conversational AI for enterprise use. The launch gives companies a way to put a face on customer-service and sales agents, while keeping custom likenesses behind an access gate. Google says the feature is generally available in Gemini Enterprise; custom-avatar creation requires allowlisting. ([Google's announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/?ref=runtimewire); [Google Cloud's availability details](https://cloud.google.com/blog/products/ai-machine-learning/gemini-3-8-live-with-live-avatar-is-now-generally-available?ref=runtimewire)) The product is the work of Google's Gemini Audio Team. Research scientist Shuo-yiin Chang and software engineer CJ Zheng authored the launch post, describing a system meant to listen, see and speak while presenting a visual persona. That framing puts the bet on the quality of a conversation, not simply on whether an avatar can move its mouth: the agent is supposed to keep talking while it processes what a customer says and handles work in the background. ([Engadget's report](https://www.engadget.com/2268587/google-video-avatars-gemini-3-8-live-agent/?ref=runtimewire)) ### The avatar is the interface; the agent does the work Google's demos show realistic and cartoon-style characters, with synchronized lip movements, facial expressions and turn-taking. The more consequential capability is asynchronous tool calling: Gemini can fetch information or run a task while the avatar continues the conversation. Google's example is a hotel check-in, where an agent can retrieve information without leaving the customer in silence. Its Cloud team describes possible uses across web, mobile and kiosk experiences. Google Cloud also names Cox Automotive as a customer example: it says the company built an AI-powered shopping assistant for Autotrader using a conversational [AI Avatar](https://runtimewire.com/models/fal/Ai-Avatar) to help shoppers search, compare vehicles and discuss financing. The post does not provide deployment numbers or independently measured outcomes. ([Google's product announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/?ref=runtimewire); [Google Cloud's customer examples and product details](https://cloud.google.com/blog/products/ai-machine-learning/gemini-3-8-live-with-live-avatar-is-now-generally-available?ref=runtimewire)) The split between general availability and allowlisted customization is important for developers planning a product around the feature. Google says enterprise customers can choose from preset [avatars](https://runtimewire.com/models/fal/veed-avatars-1), while organizations seeking an avatar generated from a reference image must go through an approval process. Google's documentation says customers using custom avatars are responsible for obtaining the necessary rights and consent for face and voice samples. That places some of the identity risk on the deploying organization, even as Google says the product has safeguards and SynthID watermarks in generated audio and video. ([Google Cloud documentation on configuring avatars](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/live-api/configure-live-avatars?ref=runtimewire); [Google's announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/?ref=runtimewire)) Google has not published Live Avatar-specific pricing or usage limits in the launch materials, which link to a pricing page. Nor do those materials offer independent measurements of latency, accuracy or likeness preservation. Google Cloud's post names Cox Automotive's Autotrader assistant and describes it as a conversational AI Avatar for vehicle discovery, but gives no deployment numbers. The post says Live Avatar is generally available in Gemini Enterprise and lists US and EU endpoints; it does not establish that every customer has identical access or that custom-avatar creation is open to all. Developers will need to check the deployment and approval terms for their own use case. ([Google Cloud's availability details](https://cloud.google.com/blog/products/ai-machine-learning/gemini-3-8-live-with-live-avatar-is-now-generally-available?ref=runtimewire)) ### Google's voice-agent push gets a face Live Avatar follows Google's September 15th release of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, which Google positioned as models for real-time voice agents and task completion. The avatar feature layers video onto that live dialogue system. Google's Cloud documentation describes 24-frame-per-second avatar video and sub-second interaction latency; those are Google's technical specifications, not independent performance results. ([Google's September 15th model announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/?ref=runtimewire); [Gemini 3.8 Live developer guide](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-3-8-live?ref=runtimewire)) Google has also been adding controls for generated speech. RuntimeWire reported on September 23rd that Gemini 3.8's text-to-speech release credited Hume AI founder Alan Cowen, with tools for custom voice design and line-by-line performance control. Live Avatar extends the same broad product direction into a combined voice-and-video interface, although the announcements do not establish that the releases share the same underlying work or team. [RuntimeWire's report on the TTS release](https://runtimewire.com/article/google-gemini-3-8-flash-tts-alan-cowen) For companies building customer-facing agents, the attraction is straightforward: a single interface that can talk, interpret live visual input and call business systems. But a face also changes the expectations placed on the agent. A fluent voice can already make a scripted interaction feel personal; a customized likeness adds identity and consent questions, while a visibly human-like persona may encourage customers to infer capabilities the system does not have. Google says its watermarking is intended to make generated content detectable, but it has not supplied independent evidence here on how well the safeguard works in real deployments. The launch gives enterprise developers another way to package Gemini's live agent capabilities. Whether customers want an avatar in a support flow will depend on access terms, cost, reliability and whether the visual presence improves the interaction. Google has named customer examples, including Cox Automotive's Autotrader shopping assistant, but has not disclosed deployment scale or provided evidence that an avatar resolves support requests better than a voice or text agent.