Most voice AI still talks to you from inside a plain chat window — a waveform, maybe a transcript, no face. On September 24, 2026, Google changed that for its enterprise customers: Gemini 3.8 Live Avatar, a feature that puts a lip-synced, expressive video face on top of Gemini's live voice conversations, reached general availability inside Gemini Enterprise, with endpoints live in the United States and the European Union.
This isn't Google's first live-voice product — Gemini 3.8 Live and a heavier 'Extended Thinking' variant for complex reasoning both launched September 15, 2026, and LumiChats already compared the base Live model against OpenAI's GPT-Live and Meta's Muse Voice Transcribe. Live Avatar is a layer on top of that same Live model: it doesn't change what Gemini can reason about, it changes how the conversation looks while Gemini is doing it. Here's what's actually new, what it costs, and — because a talking face raises the stakes on being wrong — where the feature's own documented limits are.
Quick summary: Gemini 3.8 Live Avatar reached general availability in Gemini Enterprise on September 24, 2026, adding a near-real-time, lip-synced video avatar on top of Gemini's existing Live voice model (which itself launched September 15). It supports 97 languages with automatic detection, accepts audio, camera feeds, and screen shares as input, and can keep talking while background tool calls run — Google calls this asynchronous function calling. Pricing is token-based: avatar video output costs $1.00 per million tokens, but a talking avatar burns roughly 6,192 tokens per second of speech, working out to about $0.37 per minute of active talking — and you're billed only while the avatar is speaking, not while it's listening. The real catch: no one has published an accuracy benchmark for Live Avatar itself. Reviewers instead point to an unrelated finding — a rival model, Claude Opus 4.5, scoring roughly 92% on structured support tasks in Sierra's τ-bench but collapsing to around 25% on document-grounded retrieval — as a cautionary analogy: a confident face doesn't make an underlying answer more correct, and nothing about Live Avatar's own accuracy has actually been measured yet.
What Live Avatar actually does
Live Avatar takes Gemini's existing speech-to-speech pipeline and adds a synchronized video layer: precise lip-syncing, natural facial expressions, and turn-taking Google says is fluid enough for real conversations, not just scripted lines. It processes live visual input at roughly one frame per second — camera feeds or a shared screen — alongside audio, and watermarks its output using Google's SynthID system so the video can be identified as AI-generated. Google says it's designed to recover from mid-sentence interruptions without losing conversational context, and the avatar can keep talking while it waits on a tool call or API result — what Google calls asynchronous function calling (Source: techjacksolutions.com).
Pricing: what a talking face actually costs
| Input/Output Type | Price per 1M tokens | Notes |
|---|---|---|
| Text input | $0.75 | Standard Gemini Live rate |
| Video/image input | $1.00 | Camera feed or screen share |
| Audio input | $3.00 | Speech-to-speech |
| Text/reasoning output | $4.50 | Underlying model reasoning |
| Audio output | $12.00 | Synthesized voice |
| Avatar video output | $1.00 | ≈6,192 tokens/sec of speech ≈ $0.37/min |
The number worth remembering isn't the $1.00-per-million-token line item — it's the roughly $0.37-per-minute figure that falls out of it, because that's what shows up on an invoice. Google only bills avatar output while the avatar is actively speaking, not while it's silently listening to the user, which matters for any use case with long listening pauses, like a guided walkthrough or a tutoring session (Source: eesel.ai).
Where Google is pointing this
- Hotel and kiosk check-in experiences, where a face reduces the friction of talking to a screen.
- Vehicle-shopping and retail assistants that can highlight parts of the screen while talking through them.
- Multilingual call-center deployments — Google's own materials reference use cases built around handling over a million daily calls.
- Language tutoring and guided product walkthroughs, where seeing a face model pronunciation or point at a UI element adds value over audio alone.
Two tiers of avatar exist: pre-built avatars are self-serve for any Gemini Enterprise customer today, but custom avatars — built from a single reference image and audio sample — are allowlist-only, gated behind a verification process, and require going through Google's sales team rather than a simple API toggle. Building a production deployment also means working with Google's Agent Development Kit, which Google itself frames as a real engineering project, not a no-code switch (Source: [eesel.ai](https://www.eesel.ai/blog/gemini-3-8-live-avatar)).
The catch nobody should skip
Adding a photoreal, lip-synced face to an answer doesn't make that answer more correct — it just makes an incorrect answer feel more confident. No one has published an accuracy benchmark for Live Avatar itself — but reviewers covering the release keep citing an unrelated, uncomfortable data point as a warning: on Sierra's τ-bench, a rival model (Claude Opus 4.5) scores roughly 92% on structured support tasks but collapses to around 25% on document-grounded retrieval, where it has to pull the right fact from a knowledge base rather than reason generally. A plain text chatbot that's wrong looks like a chatbot that's wrong. An avatar that's wrong, with a straight face and a steady voice, is a materially different user-trust problem — and it's one the avatar layer itself does nothing to solve (Source: eesel.ai).
How this fits next to GPT-Live and Muse
LumiChats covered the three-way voice race between GPT-Live, the base Gemini 3.8 Live, and Meta's Muse Voice Transcribe when they were competing mainly on price and latency. Live Avatar doesn't reopen that price war — Gemini's underlying Live token rates are unchanged — it adds a dimension none of those three had: a visual presence. Whether that's a genuine improvement or just a more expensive way to deliver the same voice answer depends entirely on the use case; for a 1:1 tutoring session or a kiosk, a face plausibly helps. For most back-end customer-support deflection, it's additional cost and additional risk for a benefit that's harder to measure.
01What is Gemini 3.8 Live Avatar?
A feature, generally available in Gemini Enterprise since September 24, 2026, that adds a real-time, lip-synced video avatar on top of Google's existing Gemini 3.8 Live voice model. It supports 97 languages and can process camera feeds or screen shares alongside audio.
02How much does it cost to use?
Avatar video output is billed at $1.00 per million tokens, but a talking avatar consumes about 6,192 tokens per second of speech — roughly $0.37 per minute of active talking. You're only billed while the avatar is speaking, not while it listens.
03Is this the same as Gemini 3.8 Live?
No — Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, the underlying voice models, launched September 15, 2026. Live Avatar, which launched September 24, is a video layer on top of that same Live model, not a new reasoning model.
04Does the avatar make Gemini's answers more accurate?
No — and there's no published benchmark showing it does. The avatar changes presentation, not underlying answer quality; reviewers point to an unrelated τ-bench result on a different model (Claude Opus 4.5, 92% on structured tasks vs. ~25% on document-grounded retrieval) as a reminder that a confident face never substitutes for verified accuracy.
05Can I build a custom avatar with my own face or brand character?
Google offers that, but it's allowlist-only: it requires a single reference image and audio sample, a verification process, and direct engagement with Google's sales team — it isn't a self-serve option the way pre-built avatars are.
Judging whether a feature like this is actually worth the latency and cost for your use case is hard to do from a spec sheet — it's easier to do by putting the same prompt in front of several models and seeing how each handles it. LumiChats gives you access to more than 40 AI models, including leading voice- and vision-capable systems, in one subscription, so you can compare how a new feature like this performs against your own real questions before committing to it.
