Google did not ship a prettier face for last week’s Live model. It shipped a presence channel.
On September 24, 2026, Google’s product blog from Shuo-yiin Chang and CJ Zheng introduced Gemini 3.8 Live with Live Avatar: near real-time visual presence paired with the live dialogue stack, available the same day in Gemini Enterprise. Configure docs updated the same date show the operator switch in plain API English. Set response_modalities to ["VIDEO"], pass avatar_config, keep talking on the Live WebSocket path. The Verge’s Emma Roth filed the same day: Enterprise-only, lip-sync and expressions, ninety-seven languages, SynthID, preset plus custom avatars. The Register followed early September 25 UTC with the same ship facts and a longer anthropomorphism caution.
This is a primary-source review. No hands-on theater. Distinct from Vault’s Live dual-SKU note and from yesterday’s TTS dual-studio note. The wrong read is treating Live Avatar as makeup on gemini-3.8-live. The right read is a gated channel SKU for support, sales, and training agents that now have a face, a voice pairing, and a different session budget.
What it is
Live Avatar is visual presence on Gemini Live for enterprise agents. The blog’s claim is simple: conversation is multimodal, so the agent should listen, look, speak, and show expression. The feature pairs near real-time video generation with speech so the experience listens, sees, and speaks with a dynamic visual persona. Precise lip-sync, natural expressions, and fluid turn-taking are the marketed behaviors. Asynchronous tool calling keeps dialogue moving while tools run in the background. Multilingual speech-to-speech synchronization is claimed across ninety-seven languages without degrading video fidelity or introducing visual drift.
The callable model string on the Enterprise Agent Platform docs is gemini-3.8-live. The model card lists GA with a September 24, 2026 release date. Modalities: text in and out, image input, audio in and out, video in and out. Token limits: 128K input, 64K output. Live API supported. Thinking, structured output, context caching, and several other Agent Platform features are marked not supported on this card. Regions for availability and PayGo include multi-region us and eu plus us-central1. Knowledge cutoff: January 2025.
Live Avatar is not a separate model id in the configure docs. It is a configuration of the Live path: VIDEO modality plus avatar_config. That distinction matters for procurement. You are not buying a third sibling beside Flash TTS and Flash-Lite TTS. You are enabling a presence mode on the Live Enterprise surface.
What changed
Enterprise gate on day one. Blog and Verge agree: available in Gemini Enterprise starting September 24. Consumer Live users do not get this as a casual toggle in the coverage we verified. If your roadmap assumed a free Gemini app face for brand demos, rebuild the roadmap.
VIDEO modality as the enable switch. Configure-live-avatars is unambiguous. In generation_config, set "response_modalities": ["VIDEO"]. In avatar_config, set a prebuilt avatar_name (example in docs: Ben) and pair a prebuilt voice through speech_config (example: Puck). Console path: Agent Platform > Studio > Stream realtime, switch model to gemini-3.8-live, select Live Avatar, choose avatar and voice, optional camera input, start session.
Custom avatars are allowlisted, not self-serve. From a high-quality reference image, developers can generate an animated avatar that preserves likeness and brand styling. Blog and docs: custom creation is enterprise allowlisting / select customers. Docs require PNG preferred, RGB, minimum 704×1280, 720p or higher, under 5 MB, bust framing with head and shoulders over sixty percent of the frame, neutral expression, no minors or celebrities. You are responsible for consents and rights. That is a legal and brand-ops ticket before it is a design ticket.
Continuous presence while tools run. The hotel check-in demo in the blog is the operator metaphor: tool calls and data fetches happen in the background while the avatar keeps the conversational beat. Marketing and CX teams that script “please hold while I look that up” as dead air need a new turn-taking design.
Trust stack named on launch. SynthID watermarking is woven into audio and video output. Blog points to the model card for the broader safety approach. The Register uses the ship to revisit Google’s own research warnings on anthropomorphism and undue trust. Operators should treat that as a briefing input for risk and brand, not as a reason to ignore the SKU.
Bright line against Live SKUs and TTS studios
Do not collapse three Vault tickets into one “Google voice” budget.
Live is two SKUs covered the Live dialogue family split (including Extended Thinking lanes) as interactive conversation SKUs. TTS is two studios covered gemini-3.8-flash-tts versus gemini-3.8-flash-lite-tts as exact-recitation generation SKUs. Live Avatar sits on the Live interactive path with VIDEO out and an avatar persona. It is not a TTS studio upgrade. It is not “make Live prettier” as a free cosmetic. Session economics, concurrency, and Enterprise gating are different from a text-to-speech rate card.
Also distinct from voice-is-not-the-sku (Codex CLI voice), one-claude-is-the-workspace, default-is-opus, harness-is-the-product, coworkers-with-logins, and seat-is-the-workspace. If your thread title says “avatar” and the body still argues TTS language counts, you are in the wrong document.
What works
A clear enablement contract. VIDEO + avatar_config + voice_config is a checklist a platform engineer can implement without inventing product language. Console parity for Studio Stream realtime lowers the first demo cost for Enterprise tenants who already live in Agent Platform.
Presence that survives tool latency. Async tool execution with continuous presence is the feature that makes support and sales agents feel less like IVR after a function call. For hotel, retail, and onboarding flows that need lookup mid-turn, that is the practical unlock the blog is selling.
Language as a presence problem, not only an ASR problem. Ninety-seven languages with lip-sync adaptation is a GTM claim aimed at global CX and training. Verge restates the no-drift language switch. Localization leads should test their top languages for visual credibility, not only transcript quality.
Brandable surface with controls. Prebuilt library for speed. Custom allowlist for brand characters that need likeness control. SynthID gives trust and brand teams a detection story to put in the appendix.
Model card honesty on limits. No thinking mode, no structured output, no context caching on the 3.8 Live card. That prevents the fantasy that Live Avatar is a full Agent Platform kitchen sink with a face glued on.
What breaks
Enterprise-only means most brand teams cannot pilot on a personal Gemini seat. If procurement has not cleared Gemini Enterprise, you do not have Live Avatar. Stop storyboarding custom faces for a consumer demo.
Audio-video session budgets are tight without compression. Start-and-manage-session docs (updated September 24): without context window compression, audio-only sessions are limited to about fifteen minutes and audio-video sessions to about two minutes because of token limits. Connection lifetime is around ten minutes; a goAway arrives about sixty seconds before the end. Session resumption can continue context across reconnects within documented windows. PayGo caps concurrent sessions at 1,000 per project; Provisioned Throughput customers are outside that cap. A training avatar that “just talks for half an hour” needs compression and resumption designed in, or it dies mid-demo.
Custom avatars are a consent and allowlist pipeline. No rights clearance, no reference image that meets framing rules, no account-team allowlist: no custom face. Marketing that promises “our CEO avatar next sprint” without legal is fiction.
Anthropomorphism is a product risk, not a vibe. The Register’s framing is useful even if you reject the tone: humanlike presence can inflate trust. Support scripts, disclosure, and escalation to humans need to be rewritten for a face that looks attentive while a tool fails.
Wrong SKU shopping. Buying Live Avatar to solve audiobook narration or high-volume dubbing is a category error. That work still belongs on the TTS studios. Buying it as a free cosmetic on an existing Live audio agent without rechecking video session limits will page the on-call during the first long call.
Camera input changes the threat model. Docs allow optional camera share into the Live session. CX and security need a policy before agents “see” buyer environments, IDs, or screens.
Who it is for
Primary: Enterprise CX, sales enablement, and learning teams already on Gemini Enterprise who need a present agent for support, guided selling, or interactive training, and who can staff avatar selection, disclosure, and session architecture.
Secondary: Platform engineers wiring Live WebSocket sessions who must add VIDEO modality, avatar_config, compression, and resumption without breaking audio-only agents.
Not the buyer: Teams that need consumer Gemini faces tomorrow, pure TTS narration at Flash/Flash-Lite rates, or a thinking-mode research agent with a bust silhouette.
Pricing and limits (documented)
Public blog and configure docs do not publish a separate Live Avatar per-minute sticker in the pages verified for this review. Treat commercial terms as Enterprise contract work. What is documented and operator-relevant: model gemini-3.8-live GA on September 24; PayGo and Provisioned Throughput supported on the model card; 1,000 concurrent PayGo sessions per project; audio-video session length ~2 minutes without compression; ~10 minute connection with goAway; custom avatars allowlist-only; SynthID on outputs. Do not invent a rate card. Ask Cloud account for the presence SKU line before finance models a face for every agent seat.
What to do this week
- Confirm Enterprise entitlement before creative. If you are not on Gemini Enterprise, Live Avatar is not a pilot. Route the ask through Cloud, not through a consumer Gemini screenshot.
2. Name the SKU correctly in the roadmap. Label it presence / Live Avatar on gemini-3.8-live with VIDEO modality. Keep TTS Flash / Flash-Lite and non-avatar Live audio on separate lines. Refuse one “Google voice” budget.
3. Run a two-minute honesty test. Prototype an audio-video session without compression and watch the documented limit bite. Then enable context window compression and session resumption for any flow that must outlive two minutes or a ten-minute socket.
4. Pick prebuilt first. Choose one library avatar and one voice. Write disclosure copy that says the agent is AI, watermarked, and when a human takes over. Save custom likeness for an allowlisted brand character with legal sign-off.
5. Rewrite turns for continuous presence. Replace dead-air holds with talk-while-tool patterns. Define what the avatar says when a tool fails. CX owns that script, not only engineering.
6. Map language and trust. Test your top three languages for lip-sync credibility. Brief brand and risk on SynthID plus anthropomorphism. Decide camera-input policy before anyone enables it in Studio.
7. Staff the channel owner. Presence is a channel. Assign who owns avatar selection, who owns Enterprise quota and concurrency, and who owns the human escalation path. If those names are blank, you bought a face without an operating model.
Live Avatar is not makeup. It is a gated presence channel on Gemini Live. Treat it like a new surface with session math, Enterprise locks, and brand risk, or you will staff it like a cosmetic patch on last week’s dual SKU and ship the wrong agent.
