Google made Live Avatar generally available with Gemini 3.8 Live on September 24. It adds a speaking, animated face to the company's live voice model, which can also receive camera or screen input and call external tools during a conversation. For developers building customer-facing agents, the practical change is an integrated video output and a documented API path. The announcement does not establish that an avatar improves service outcomes or works reliably in every setting.

The big change

  • What changed: Google's developer guide adds synchronized avatar video to the speech output available in its earlier Live model. Developers can now obtain a talking face and voice from one streaming API.
  • Why it matters: Teams building customer-service agents can add that visual presence without a separate avatar-rendering pipeline. Their applications still execute tool calls and handle failures; the animation does not establish that a transaction succeeded.
  • What to watch: The immediate design choice is between a preset avatar and a custom likeness. Google's configuration guide restricts custom avatars to selected customers and requires rights and consent for face or voice samples. A branded likeness therefore needs a separate access and consent decision.

The avatar is available; a custom likeness has a separate gate

In its launch post, Google shows avatars responding to live audio and visual input, changing languages, and continuing a hotel check-in conversation while a tool call runs. Those are company demonstrations. They show the intended interaction, but provide no measured error rate, accessibility result or evidence of better customer service.

Google's Cloud announcement says the feature is generally available in Gemini Enterprise and through the API, with US and EU endpoints. Developers can choose from preset avatars. Creating one from a person's reference image remains limited to selected enterprise customers, who must request access through Google Cloud. The configuration guide says customers are responsible for obtaining the rights and consents needed to process face or voice samples.

The model page lists gemini-3.8-live and availability in the us and eu multi-regions and us-central1. It lists both standard pay-as-you-go and provisioned throughput as supported consumption options. Teams should check the endpoint and commercial terms for their own deployment before treating general availability as universal access.

The conversation still depends on the application

Google describes Gemini 3.8 Live as a speech-to-speech system that can take live camera frames or a screen share and send back speech and video. Its developer guide distinguishes calls that run while the agent keeps speaking from blocking calls that wait for a response. The application still has to connect and authorize its own tools, handle their results and decide what action the agent may take. A face that stays on screen during a lookup does not establish that the lookup succeeded.

The guide says tool responses should report a status and whether retrying makes sense; they must match the original call ID; and developers should set a limit on repeated tool calls. Those requirements affect how an agent handles an empty account lookup or an interruption. Google also says Gemini 3.8 Live can switch among 97 languages. The announcement's claim that lip movements stay aligned across those languages has not been independently measured here.

Google says generated audio and video carry SynthID watermarks and restricts custom likeness creation. Those are stated safeguards, not evidence that every viewer will recognize the avatar as synthetic or that misuse is prevented. A deployment assessment would need to check consent, disclosure, the handling of live camera and voice data, failed tool calls, interruption behavior and observed performance with its actual users. The published launch material and documentation do not supply those results.

Sources & further reading

  • Google's Live Avatar announcement (September 24, 2026) describes and shows the company's intended interactions, multilingual avatar claims and SynthID claim. Its demos are not independent performance tests.
  • Google Cloud's general-availability post (September 24, 2026) specifies enterprise availability, US and EU endpoints, preset and gated custom avatars, and example applications. Customer comments there are promotional accounts, not outcome measurements used in this story.
  • Gemini 3.8 Live developer guide (updated September 24, 2026) documents the model comparison, streaming format and tool-call behavior. It describes API behavior rather than measured performance in a particular deployment.
  • Gemini 3.8 Live model page (updated September 24, 2026) lists the model ID, GA release date, regions and consumption options. Availability should be checked again for a specific project and endpoint.
  • Configure live avatars (checked September 25, 2026) gives the stock-avatar API setup and custom-likeness access and consent requirements.