Confirmed in a primary source 24 Sept 2026 Models

Google made a talking synthetic face generally available to any business, in 97 languages.

On September 24 Google announced that Gemini 3.8 Live with Live Avatar is generally available to enterprise customers, with United States and European endpoints, guaranteed capacity and enterprise data governance. It puts a video avatar with synchronised lip movement on top of a speech-to-speech model that detects and answers in 97 languages, can watch a camera feed or a shared screen while it talks, and calls tools in the background without pausing the conversation.

Named customers include Cox Automotive on Autotrader and Equal AI, which Google says handles more than a million live calls a day across nine Indian languages. The caveat that matters is the disclosure question: every piece of audio and video it generates carries a SynthID watermark, which a machine can read and a person on a call cannot.

Until now

A synthetic person who looks at you and answers in your language was a demo, a preview or a consumer novelty. Enterprise voice agents had no face and no eyes.

Now

Treat a lip-synced synthetic person answering in 97 languages and watching your screen as generally available business software, and treat an invisible watermark as the only thing separating it from a video call with a human.

The capability list is the part to read slowly. This is not a chatbot with a picture attached. It hears and answers in audio directly, detects which of 97 languages you are speaking, moves a face in time with what it says, recovers when you talk over it, looks at your camera or your screen while the conversation continues, and runs tools in the background without going quiet. Google's own example is a car buying assistant that highlights things on your screen while it explains them. Its other example is a company already putting more than a million calls a day through the underlying stack in nine Indian languages. That is not a pilot.

Nothing here is a breakthrough, and that is why it matters. Each piece has existed in preview for a year; what changed on September 24 is the service agreement, the European endpoint, the reserved capacity and the compliance paperwork, which is the exact set of things a bank or an insurer waits for before putting a synthetic face in front of a customer. The one safeguard Google names is SynthID, a watermark in the audio and video that a detector can read and a person cannot hear or see. No jurisdiction currently requires that a caller be told. Treat a lip-synced synthetic person answering in 97 languages and watching your screen as generally available business software, and treat an invisible watermark as the only thing separating it from a video call with a human.

How it got here

Sep 15
Meta begins charging for its AI, from 2.99 dollars a month to 499. Carded in Edition 023.
Sep 22
OpenAI and Anthropic cut prices against each other on the same day. Carded in Edition 030.
Sep 23
Google releases Gemini 3.8 Flash and Flash-Lite text-to-speech, more than 100 languages, over 2,000 voices, voice replication from a 30 second sample.
Sep 24
Gemini 3.8 Live with Live Avatar reaches general availability with United States and European endpoints and SynthID marking on all generated audio and video.
Who has to react
The contact centre industry and the several million people employed in it, particularly in India and the Philippines; ElevenLabs and Synthesia, whose products this bundles into a hyperscaler's enterprise agreement; OpenAI and Anthropic, neither of which sells a face; Salesforce, which is named as an integration partner and also sells the seats this replaces; European data protection regulators, since a European endpoint and enterprise data governance are precisely the conditions under which this gets deployed on European consumers; the drafters of the AI Act's transparency obligations, which turn on telling a person they are interacting with a machine; banks, insurers and health providers evaluating synthetic agents for regulated conversations; fraud and identity teams, for whom a lip-synced face with live camera access is a new attack surface; and voice actors and dubbing houses in 97 languages.

The evidence

ObservedThursday September 24, 2026. On the Google Cloud blog, Google announced general availability of Gemini 3.8 Live with Live Avatar for enterprise customers.

The whole observation, as recorded

What is included: video avatars with lip movement synchronised to generated speech; a native speech-to-speech foundation rather than a chain of transcription and text-to-speech; automatic language detection and response across 97 languages; improved voice activity detection, lower latency and recovery from interruption without losing the thread of the conversation; tool calling that runs in the background while the avatar keeps talking; and live visual understanding of camera feeds and shared screens. What comes with it for buyers: endpoints in the United States and the European Union, provisioned throughput, enterprise compliance and data governance terms. What does not ship: Gemini 3.8 Live Extended Thinking stays in private preview. On marking: Google states that SynthID watermarks are applied to all generated audio and video. Named deployments in Google's own post: Cox Automotive, using it on Autotrader for a shopping assistant that highlights items on the user's screen while speaking; Equal AI, which Google says handles more than one million live calls a day across nine Indian languages; Salesforce, integrating it with Agentforce; and Specs, quoted on the voice activity detection and latency work. Pricing is published on Google Cloud's generative AI pricing page and access runs through the Gemini Enterprise console. The release follows Google's September 23 launch of Gemini 3.8 Flash and Flash-Lite text-to-speech models covering more than 100 languages and over 2,000 voices, with voice replication from a 30 second sample.

Established

Google announced on September 24, 2026 that Gemini 3.8 Live with Live Avatar is generally available for enterprise customers, with United States and European Union endpoints, provisioned throughput and enterprise data governance.
The product supports automatic detection and response across 97 languages, lip-synchronised video avatars, native speech-to-speech, background tool calling, and live understanding of camera feeds and shared screens.
Google states that SynthID watermarks are applied to all audio and video the system generates.
Google names Cox Automotive, Equal AI, Salesforce and Specs as deployments, and states that Equal AI handles more than one million live calls a day across nine Indian languages.

Not established

That anyone is told. Google states that generated output is watermarked; it does not state that any customer discloses to the person on the call that the face is synthetic, and no watermark is perceptible.
That the deployments are running on the avatar. The Equal AI figure is a call volume on the underlying stack, and Google does not say how many of those calls carry a face.
That this replaces workers. No customer named a headcount, a cost saving or a role, and general availability is a purchase option rather than a deployment.
That the watermark survives. Google publishes no figure here for SynthID's robustness to re-encoding, screen capture or compression, all of which are how a video actually travels.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityMedium

PersistenceYears

We are scoring this modestly and deliberately. Cost is medium: shipping general availability with European endpoints, provisioned throughput, compliance terms and named reference customers is real engineering and legal work, but it is a product launch by a company that ships several a month, and the underlying models already existed. Commitment is medium because general availability carries service agreements Google has to honour and named customers who have deployed against it, which is more binding than a preview and far less binding than a capital commitment. Reversibility is medium: features get deprecated, and Google has retired products at this stage before, though the customers listed have built against it. Persistence is years, and this is the part that earns the card, because synthetic faces entering regulated customer conversations creates disclosure questions that outlast any particular product. A modest signal, scored modestly.

What to watch

strengthens A regulator or legislature requires that a person be told when a synthetic avatar is on the call, or a named enterprise deploys Live Avatar to consumers in a regulated sector such as banking, insurance or health, by June 30, 2027.
weakens No named customer deploys the avatar to end consumers, or Google returns Live Avatar to preview, deprecates it, or restricts it to internal and agent-assist use, by June 30, 2027.

The words

Generally available: Out of testing and sold with a service agreement, support and compliance terms. It is the point at which a large company will put a product in front of customers.
Speech to speech: A model that hears audio and produces audio directly, rather than transcribing to text, thinking in text and reading the text aloud. It is faster and keeps tone and interruption.
SynthID: Google's watermark, hidden in the audio and video it generates, that its own detector can read. It does not appear or sound like anything to the person watching.
Provisioned throughput: Capacity a customer pays to reserve in advance, so the service does not slow down under load. It is what a company buys when it intends to run something at scale.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.