Google made Gemini 3.8 Live Avatar generally available on September 24, 2026, adding lip-synced video faces to its real-time voice agent across 97 languages. Introductory pricing runs $3.00 per million audio input tokens and $12.00 per million audio output tokens through December 31, 2026 — roughly $0.39 per minute of agent speech. Standard rates jump on January 1, 2027.
Google shipped the feature into Gemini Enterprise only, on US and EU endpoints. There is no consumer release and no free tier.
The move lands eight months after Google Ventures led a $200 million round into Synthesia at a $4 billion valuation. Google now sells a competing product.
What is Gemini 3.8 Live Avatar?
Gemini 3.8 Live Avatar is a real-time video layer bolted onto Google’s speech-to-speech model. It generates a talking human face that lip-syncs to the model’s own audio output while the conversation is happening, rather than rendering a video after the fact.
Google announced it on September 24, 2026 via the official Gemini models blog. The company describes the result as “near real-time visual presence.”
The pitch is customer service and sales. An agent that can be seen, Google argues, holds attention better than a disembodied voice.
How the avatar pipeline works
The model processes audio and visual input at the same time — including camera feeds and screen shares — and emits synchronized audio and video in one stream. Google says lip-sync and expression adaptation hold across all 97 supported languages “without degrading video fidelity or introducing visual drift.”
Two engineering details matter more than the demo reel.
- Asynchronous tool calling. The avatar can fetch data in the background while still talking, so it does not freeze mid-sentence waiting on a database.
- Automatic language switching. The system detects a speech change and follows it without losing video alignment, per Google’s documentation.
- SynthID watermarking. Every frame and every second of audio carries an imperceptible watermark, embedded at generation time.
- Custom personas. Enterprises can build an avatar from a single reference photo and an audio sample — but custom creation sits behind a strict allowlist.
- Pre-built library. Every Gemini Enterprise customer gets stock avatars immediately, with no approval step.
The allowlist is the tell. Google is shipping the easy half of the product broadly and gating the half that creates deepfake exposure.
How much does Gemini 3.8 Live Avatar cost?
Introductory rates are live through December 31, 2026: $0.75 per million text input tokens, $3.00 per million audio input tokens, $12.00 per million audio output tokens, and $1.00 per million video output tokens. Pricing analysis from DataNorth estimates roughly $0.39 per minute of agent speaking time at those rates.
That works out to about $23.40 an hour of live agent talk time.
Hold that number against the voice-only tier. When Gemini 3.8 Live launched on September 15, Wealth Engine priced it at $0.84 an hour. Adding a face multiplies the bill by roughly 28x.
| Token type | Introductory rate (through Dec 31, 2026) | Standard rate (from Jan 1, 2027) |
|---|---|---|
| Text input | $0.75 / M tokens | $1.50 / M tokens |
| Audio input | $3.00 / M tokens | Blended into $1.50 in / $7.50 out |
| Audio output | $12.00 / M tokens | $7.50 / M tokens |
| Video output | $1.00 / M tokens | Not separately published |
| Effective cost | ~$0.39 / minute of speech | Not published |
What happens on January 1, 2027
Google published standard rates of $1.50 per million input tokens and $7.50 per million output tokens starting January 1, 2027. It did not publish a corresponding per-minute figure, and the video output line disappears from the standard table.
Buyers signing annual contracts this quarter are pricing a product whose 2027 unit economics are not fully disclosed. That is a procurement problem, not a technical one.
Who is affected by the Gemini 3.8 Live Avatar launch?
The standalone AI avatar vendors absorb the hit first. DataNorth names Synthesia, HeyGen and D-ID as the companies most directly pressured, because Google has folded video generation into the same call that already handles the voice.
That collapses a two-vendor stack into one line item.
Synthesia, HeyGen and the awkward Google connection
Synthesia raised $200 million at a $4 billion valuation on January 26, 2026, in a round led by Google Ventures. The company says it has over one million users and counts 90% of the Fortune 100 as customers.
CEO Victor Riparbelli framed the raise around scale: “This funding round is about scaling that vision. We see a rare convergence of two major shifts.”
Eight months later, Google’s own cloud division is selling into the same buyer. GV is a financially separate entity from Google Cloud, but the optics for Synthesia’s next round are not good.
HeyGen is the other exposed name. The company announced it crossed $200 million in ARR on June 25, 2026, doubling in eight months, with 30 million users and 85% of the Fortune 100.
CEO Joshua Xu has staked the company on quality rather than volume: “People don’t want more AI slop. They want to communicate with trust, clarity, and presence in every language and every format where their audience lives.”
That is a reasonable defense against a commodity bundle. It is a harder defense against a bundle that is already inside the enterprise contract.
Which customers are already using it?
Google named Salesforce Agentforce and Cox Automotive as launch customers. Both are large, both already run voice agents, and both are exactly the accounts a standalone avatar vendor would list in a pitch deck.
No usage volumes were disclosed for either. No conversion or containment metrics were published.
What did Google not publish?
Quite a lot, and the omissions are consistent. Google shipped no latency figure — the product is described as “near real-time” with no millisecond number attached.
There is also no benchmark of any kind. No lip-sync accuracy score across the 97 languages. No interruption-recovery measurement. No evaluation harness.
This is a pattern in recent frontier launches. OpenAI’s GPT-6 Sol pricing disclosure was similarly thin on the operational numbers buyers actually need.
For a product sold on the claim that faces convert better than voices, the absence of a single conversion or engagement metric is the most interesting thing in the announcement.
Why does SynthID matter financially?
SynthID is Google’s watermark, woven into both the audio and the video at generation time. Every avatar frame carries it.
The commercial logic is liability, not ethics. An enterprise deploying a synthetic human face into a customer conversation inherits impersonation risk, and a provable watermark is what a legal department needs before signing.
Trust is the binding constraint on this category. Disclosure failures in AI deployments have proven expensive — see the 84-day gap in OpenAI’s Medicare incident disclosure — and a watermark that regulators can verify is cheaper than the alternative.
The gap: SynthID only proves a video came from Google. It does not stop a competitor’s unwatermarked model from doing the same job.
Frequently asked questions
When did Gemini 3.8 Live Avatar launch?
September 24, 2026. Google announced general availability for Gemini Enterprise customers on US and EU endpoints the same day.
How many languages does Gemini 3.8 Live Avatar support?
97 languages, with automatic switching. Google states lip-sync and facial expression hold across all of them without visual drift.
Can anyone create a custom avatar?
No. Custom personas built from a reference photo and audio sample require allowlist approval. Pre-made avatars are available to all Gemini Enterprise customers without approval.
Is Gemini 3.8 Live Avatar available to consumers?
No. It ships inside Gemini Enterprise through the Agent Platform Studio and the API. There is no consumer tier and no free access.
How much does a one-hour avatar conversation cost?
At the introductory estimate of roughly $0.39 per minute of agent speech, about $23.40 per hour of talk time. Actual cost depends on how much of the hour the agent is speaking.
Does Gemini 3.8 Live Avatar publish latency numbers?
No. Google describes the experience as “near real-time” and has not released a millisecond figure or any benchmark.
Which companies does this threaten most?
Synthesia, HeyGen and D-ID — the standalone avatar-video vendors now competing with a feature bundled into an enterprise contract customers already hold.
The bottom line
Gemini 3.8 Live Avatar is a distribution play, not a technology breakthrough. Google has not shown that its avatar is better than Synthesia’s or HeyGen’s — it has published no benchmark that would let anyone check.
What Google has done is remove the reason to buy one separately. An enterprise already paying for Gemini Enterprise voice agents now gets video in the same API call, on the same invoice, under the same compliance review.
That is how bundles kill categories, and it is why a $4 billion valuation set in January looks less comfortable in September.
The skeptical read cuts the other way too. At roughly $23.40 an hour of agent speech, the avatar costs something like 28 times the voice-only tier, and Google has published nothing showing that a face pays for that gap. Standalone vendors survive if faces turn out to be a nice-to-have. They do not survive if faces convert.
Watch January 1, 2027. When the introductory rates expire and Google declines to publish a per-minute standard price, the real margin structure of this product becomes visible — and so does whether the bundle was priced to win share or to hold it.
Leave a Reply