Google’s Live Avatar Puts a Face on Gemini Agents

Google opened Gemini 3.8 Live with Live Avatar in Gemini Enterprise on September 24, adding a lip-synced talking face to its live voice agents. The persona listens, sees, and speaks, and it is not in the consumer Gemini app.

The company is selling a human-looking agent while wrapping it in the same tools it uses to flag fake media: an allowlist for cloned faces, a silent watermark on every frame, and a session that lasts a few minutes.

Gemini Enterprise Gets the Talking Face First

Research scientist Shuo-yiin Chang and software engineer CJ Zheng, writing for the Gemini Audio Team, said Live Avatar pairs near real-time video with speech so an enterprise agent can hold a visible conversation. Google Cloud group product manager Fabien Blanc-paques said the feature is generally available in Gemini Enterprise after a preview at Google Cloud Next 2026, with US and EU endpoints, provisioned throughput, and enterprise data controls.

Gemini 3.8 Live itself shipped on September 15, nine days earlier, as a speech-to-speech model for voice agents. Live Avatar is the visual layer on that model, not a new brain. W. Ronny Huang, who works on realtime Gemini models at Google DeepMind, wrote that the work began as a bottom-up research project on video agents with a small team in Gemini Audio, backed first by private cloud customers.

Google Cloud Tech posted the enterprise launch the same day.

Power your agents: Gemini 3.8 Live with Live Avatar is now generally available in Gemini Enterprise.

Key capabilities include: video avatars, fluid dialogue, tool calling, and more.

Start building today ↓ https://t.co/MD9Smf29Hq

— Google Cloud Tech (@GoogleCloudTech) September 24, 2026

THE PATH TO A FACE

1. Google Cloud Next 2026: Google previews Live Avatar for enterprise production.
2. September 15, 2026: Gemini 3.8 Live and 3.8 Live Extended Thinking ship as live dialogue models. Extended Thinking stays in private preview.
3. September 24, 2026: Live Avatar becomes generally available in Gemini Enterprise, with custom faces still on an allowlist.

The Gemini app, Google AI Studio, and the public Gemini API still get the voice model. The talking head is a Cloud product for web, mobile, and kiosk agents.

Google Treats the New Face Like a Deepfake

Chang and Zheng wrote that Live Avatar was built with strict safeguards designed to respect identity and keep generated media transparent. Blanc-paques put the same lock in plainer terms: firms start from a library of curated, pre-built avatars, and a custom face has to clear a strict enterprise allowlisting and verification process.

Google’s own Live API overview names the intended customer for that custom path. Branded corporate ambassadors, it says, should be generated from authorized employee likenesses or paid talent with executed corporate releases. The face on the kiosk is supposed to be someone the company already paid, not a scraped celebrity still.

THE IDENTITY LOCK

• Preset library: Customers can deploy from a curated set of built-in avatars and pair them with prebuilt voices, including a stock face named Ben and a voice named Puck.
• Custom allowlist: A reference photo can become a live likeness only after Google allowlists the enterprise and runs a verification process.
• Likeness rights: Google’s use-case list calls for employee faces or paid talent with signed corporate releases, not random photos from the web.
• Hard bans: Docs tell builders not to upload images of minors, celebrities, or offensive content.
• Silent mark: Every audio and video stream carries an imperceptible SynthID watermark so the clip can be checked later.

That is a lot of process for a customer-service puppet. It is also an admission that a public talking-face generator would be a deepfake factory. Google already lets people watermark and identify AI-generated content in the Gemini app by uploading a file and asking whether Google AI made it. Live Avatar weaves the same mark into the stream as the mouth moves.

One Photo, an Allowlist, and No Celebrities

Custom avatars are configured per session. A developer uploads a still, and the model builds a talking head that is supposed to keep the reference likeness, brand styling, or character identity. Huang described the same loop in his own words: upload an avatar image, a voice sample, and a character description, then chat with it live. He also said an enterprise account is required for that custom path.

Google’s configure-live-avatars guide is picky about the still you feed it. The photo should be a bust shot, head and shoulders filling more than 60% of the frame, face to camera, expression at rest, no smile and no visible teeth. Backgrounds stay plain. Hands, mics, and extra people are cropped out. And the docs say not to use images of minors or celebrities.

CUSTOM FACE STILL SPECS

The API path is a WebSocket session on gemini-3.8-live with response_modalities set to VIDEO. The still goes in as base64 inside customized_avatar. Stock faces skip that upload and take an avatar_name instead. Either way, the video is supposed to come out of the model, not from a separate animation stack bolted on after the audio.

Autotrader Puts a Talking Agent on Vehicle Search

Blanc-paques listed the jobs Google wants this face to do: customer service, interactive walkthroughs, and conversational video agents on the web, on phones, and on kiosks. The Live API overview adds shopping assistants, in-game characters, healthcare wayfinding, bank triage, and AI tutors.

Cox Automotive is the named customer that actually put an Avatar on a shopping flow. It built an AI shopping assistant for Autotrader that highlights the screen and calls tools while it talks shoppers through search, comparison, and financing.

Shoppers increasingly expect to describe what they need in their own words rather than work through filters and menus. Autotrader’s new conversational AI Avatar brings that experience to vehicle discovery by matching natural conversation to the right inventory.

Marianne Johnson, EVP and Chief Product Officer, Cox Automotive

Equal AI’s chief executive, Akhilesh Damaraju, said his personal AI already handles over a million live calls a day across nine Indian languages, and that Gemini 3.8 Live improved interruption handling, multilingual talk, and tool-call reliability. That quote is about the voice model, not the new face. Salesforce is pairing 3.8 Live with Agentforce. Specs, a platform for AI assistants, praised voice-activity detection and latency. The talking head is optional. The call volume is not.

Google’s own demos are blunter about what the face is for. One clip has the avatar check a guest into a hotel while tool calls run in the background. Another is an insurance intake agent. You talk and show the damage on camera, and a claim notebook fills itself in while a backend team of agents checks the policy. A face that looks at a crumpled bumper is easier to trust than a waveform. It is also easier to mistake for a person, which is why the disclosure argument keeps coming up around that demo.

Can Live Avatar Switch Languages Mid-Sentence?

Yes. Google says Live Avatar uses native multilingual speech-to-speech, so the mouth and the expression change with the language, and the model can move across 97 languages without visual drift. A launch video has the same avatar speaking English and Japanese, with the lips tracking both. Firms do not have to tell it which language the caller used. Detection is automatic.

That is the part of the launch that is harder to copy than a pretty preset. A call-center bot that can greet in Malayalam, switch to English for a policy number, then finish in Hindi without the face melting is a staffing product. Equal AI is already running nine Indian languages at a million calls a day on the voice stack. The avatar is how that stack shows up on a screen.

The model also sees. Input can include live camera feeds and screen shares as 1 FPS JPEG frames, plus 16 kHz audio, while output is 24 kHz speech and 24 FPS MP4 avatar video. Google’s developer guide for gemini-3.8-live says the face and the voice are generated together, so there is no extra lip-sync renderer in the middle. Affective dialogue, which matches tone to the caller, is on by default. Tool calls are asynchronous, so the avatar can keep talking while a booking or a CRM write finishes behind it.

Video Burns 6,192 Tokens for Every Second of Speech

Google’s Agent Platform pricing page bills Gemini 3.8 Live by token, not by a flat minute. Text in is $0.75 per 1 million tokens. Image and video in is $1.00. Audio in is $3.00. Text out is $4.50. Audio out is $12.00. Avatar video out is $1.00 per 1 million tokens.

The expensive line is not the dollar on video. It is the burn rate. Google lists audio at 25 tokens per second and avatar video at 6,192 tokens per second of video. Video charges apply only while the avatar is speaking. Listening time is not billed.

WHAT GOOGLE BILLS FOR A TALKING FACE

One minute of the agent speaking is 371,520 video tokens and 1,500 audio tokens. At those rates that is about $0.39 in output for a minute of talk, before input. Audio alone for the same minute is about two cents. The face is the line item.

The Model Card Caps Chats at a Few Minutes

Google’s Gemini 3.8 Audio model card, posted with the launch, is blunter than the marketing clips. Live Avatar can support a few minutes of continuous interaction, rather than extended hours. With the face on, output is capped at 24K tokens, down from 64K for audio-and-text. The context window is up to 128K tokens. The knowledge cutoff is January 2025. Known limits include hallucinations, occasional slowness, and timeouts.

SESSION LIMITS ON THE CARD

• Talk time: A few minutes of continuous interaction, not a full shift.
• Output budget: 24K tokens with Live Avatar, 64K without it.
• Memory: Knowledge stops at January 2025.
• Failure modes: Hallucinations, slowness, and timeouts are listed as known limits.

A hotel check-in demo fits that window. A long claims call may not. The same model card that sells natural head movement also tells buyers the face will not sit on a kiosk all afternoon. For firms that already run a million voice calls a day, the talking head is a short, watermarked appearance, on an allowlisted likeness, in a product Google will not put on the public Gemini app.

The mouth can switch through 97 languages without drifting. The session still ends in a few minutes, and the clip is marked as machine-made before anyone hits stop.