Gemini 3.8 Live Avatar Is GA at About 37 Cents a Speaking Minute. The Avatar Only Exists in Gemini Enterprise.
Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise on September 24, nine days after the voice model shipped. Avatar video is metered at 6,192 tokens per second at $1.00 per million, which works out to about $0.37 per speaking minute before audio, reasoning, and the per-turn re-billing of the whole session. The catch is where it lives: the Gemini Developer API and the consumer app get audio only, custom faces and voices are allowlist-only, and the 97 languages in every headline are 24 in the Enterprise docs.
By T.W. Ghost
Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise today, September 24, 2026. The voice model underneath shipped nine days ago, on September 15, as a stable Gemini API model and a private preview in Gemini Enterprise; today's Google Cloud post promotes it to GA and adds the face. The avatar is a config, not a model: connect to the same gemini-3.8-live ID over the Live API, set response_modalities to ["VIDEO"], name a prebuilt avatar and a voice, and a lip-synced talking head streams back as MP4 over the WebSocket. Google's September cadence, which we tracked in the 3.8 Flash post, now has a video output.
It is also the only frontier-lab live API we found that returns avatar video. OpenAI's GPT-Live 1 went GA on September 10 at $0.05 a minute with image and video unsupported, and xAI, now SpaceXAI, sells Grok Voice 2.0 as speech to speech only at $0.08 a minute; our Realtime and Grok Voice notes cover both. Meta described a "Muse Realtime Avatar" in a research post yesterday, with no API, price, or date. Google's only "first" is "First previewed at Google Cloud Next 2026", which the Next '26 wrap-up does not mention, so we are not using the word.
The number Google does not print is the one you will budget against: 6,192 tokens per second of avatar video at $1.00 per 1M tokens, billed only while the avatar speaks, which is about $0.37 per speaking minute, our arithmetic from Google's rate. We read both Google posts, the pricing page, the Agent Platform docs, the model card, and Artificial Analysis. Here is what held up.
Thirty-Seven Cents a Minute Is Only the Video Meter
| Gemini 3.8 Live with Live Avatar | |
|---|---|
| Announced | September 24, 2026, GA in Gemini Enterprise; voice model stable on the Gemini API since Sep 15 |
| Model ID | gemini-3.8-live with response_modalities: ["VIDEO"] and an avatar_config; no separate avatar ID |
| Avatar video price | $1.00 per 1M tokens at 6,192 tokens per second, billed while speaking only (about $0.37 per speaking minute, our arithmetic) |
| Audio price | $3.00 in / $12.00 out per 1M at 25 tokens per second ($0.005 / $0.018 per minute) |
| Text and image price | $0.75 text in / $4.50 text out (response and reasoning); $1.00 image or video in; no cached-input rate |
| Context | 128K in / 64K out; 24K out with the avatar on |
| Knowledge cutoff | January 2025 (model card) |
| Avatar output | Minimum 704 x 1280 portrait or 1280 x 704 horizontal at 24 FPS, MP4 over a stateful WebSocket |
| Session length | "A few minutes of continuous interaction" (model card); maximum with the avatar on: - |
| Languages | 24 listed in the Gemini Enterprise docs; Google's posts say 97 |
| Latency | - (Google publishes no millisecond figure; the blog says "near real-time") |
| Availability | Gemini Enterprise Agent Platform only, us and eu multi-region plus us-central1; custom avatars and voices by allowlist |
The meters stack. Google lists six rates for the Gemini 3.8 Live API, all non-global, with no cached-input price and no change above 200K tokens, over conversion rates of 25 tokens per second of audio, 258 tokens per image, and 6,192 tokens per second of avatar video. A speaking minute is $0.3715 of video plus $0.018 of output audio, about $0.39; add a camera frame every second ($0.0155) and the user's own audio ($0.0045) and a fully loaded minute is roughly $0.41 before reasoning tokens and the per-turn re-billing below. All of that is our arithmetic on Google's rates; Google prints no per-minute total. Idle is free: "Video output charges only apply when the avatar is actively speaking; users are not billed for idle time (when avatar is listening)." Budget on speaking minutes, not session minutes.
The audio is cheap and unchanged. $3.00 and $12.00 per 1M at 25 tokens per second are $0.005 a minute in and $0.018 out, the same per-minute audio rate as the Gemini 2.5 Flash Live API this replaces. The Decoder ran the comparison on launch day: "An hour of voice conversation costs about $1.38 with Google versus at least $3.00 with OpenAI." Text moved against the predecessor, $0.50 to $0.75 in and $2.00 to $4.50 out, while image and video input fell from $3.00 to $1.00; the deltas are ours.
The Avatar Lives in Exactly One Place
Agent Platform only. The Gemini Developer API model page for gemini-3.8-live lists output as "Text and audio", and the Gemini API pricing page has no avatar row. The Agent Platform model page lists video as "Input and output", and the avatar docs point at one endpoint, {LOCATION}-aiplatform.googleapis.com. The consumer Gemini app and Search Live got 3.8 Live voice on September 15, no avatar; Gemini Enterprise for Customer Experience is "coming soon". On the Gemini API free tier you can hear this model and you cannot see it.
Prebuilt faces for everyone, custom faces for the allowlisted. The Cloud blog says custom avatar creation "is gated behind a strict enterprise allowlisting and verification process." The docs are blunter: "Custom avatars are only available to select customers. To request access, reach out to your Google Cloud account team. You are responsible for securing all consents and rights necessary for the processing of face and/or voice samples." A custom avatar is one reference image (PNG, 704 x 1280 minimum, under 5 MB, no minors, celebrities, or offensive content); a custom voice is a 10 to 20 second sample, also "only available to select customers". Google does not publish the size or names of the prebuilt library; the docs show one sample avatar, "Ben", and 30 prebuilt voices. Avatar count: dash.
Watermarked, with the consent burden on you. All generated audio and video carries a SynthID watermark, per both Google posts. The Gemini 3.8 Audio model card, published September 15, has no impersonation, deepfake, or SynthID evaluation table and no benchmark table at all; it says the 3.8 audio variants have no meaningful new capabilities over Gemini 3.7 Flash and leans on that model's frontier-safety finding. The mitigations are product rules: the curated library, the allowlist, the reference-image policy, and a clause that makes face and voice consent the customer's problem under Section 32 of the Service Specific Terms. The one avatar deployment Google quotes is Cox Automotive's Autotrader, a shopping assistant with live screen highlighting and tool calls; Equal AI and Specs are voice customers, and Salesforce Agentforce is a research collaboration.
The Benchmarks, Sorted by Who Reported Them
House rules: independent numbers first, stated as fact; Google's numbers labeled as Google's. One caveat up front: nobody, Google included, publishes a benchmark for the avatar. Lip-sync accuracy, visual fidelity, and visual latency are unmeasured by any party we could find. Everything below is about the voice model.
Independent. Artificial Analysis runs a Speech to Speech Index, a weighted average of Speech Reasoning (Big Bench Audio), Agentic Performance (tau-Voice), Arena Preference, and Task Success Rate. It carries no version number or update date, and AA's Intelligence Index v4.3.2, which we cite for text models and which has not changed since September 19, does not cover Live models. On September 15, Google's launch post said Gemini 3.8 Live Extended Thinking was "capturing the #1 overall spot on Artificial Analysis' Speech to Speech Quality Index (82.6)". The Decoder read the board that day and confirmed it: "The Extended Thinking variant ranks first on the Artificial Analysis Speech-to-Speech Leaderboard with 82.6 percent, ahead of OpenAI's latest GPT-Live-1 models." Two notes. AA's own name for the metric is "Artificial Analysis Speech to Speech Index"; there is no "Quality" in it. And the ranked model is Extended Thinking, still in private preview in Gemini Enterprise; the GA avatar runs on base Gemini 3.8 Live. Update (Oct 16): the top spot did not last; see the update below.
Now the vendor table. These are Google's self-reported numbers from the September 15 post, with no comparison models and no effort settings stated.
| Benchmark (Google's numbers) | Result |
|---|---|
| tau-Voice (Extended Thinking) | 68.6% |
| Sierra tau-Voice-banking (Extended Thinking) | 35.1% |
| Big Bench Audio (Extended Thinking) | 97.7% |
| Speech Agent Arena (Gemini 3.8 Live) | second place |
| ServiceNow EVA-Bench | no score; "balancing accuracy with conversational quality" |
| Avatar lip-sync or visual fidelity | - (no benchmark from Google or anyone else) |
Three honest observations Google's marketing does not lead with:
- The #1 belonged to the model still in private preview. The 82.6 is Extended Thinking's. Google's only claim for base 3.8 Live, the model behind the GA avatar, is "a second place in the Speech Agent Arena", with no word on who runs that arena or who finished first.
- Half the table has no denominator. "Second place", "balancing accuracy", and a banking score of 35.1% with no rival column are not comparisons. Google names the benchmarks and omits the field.
- The fast number is missing. Neither post gives a latency figure; "near real-time visual presence" and "low-latency streaming video" are the only descriptions. Meta's research post yesterday put its own Muse Realtime Avatar at about 870 ms, which is Meta's number for Meta's demo. We print a dash.
The Billing Fine Print
Every turn re-bills the whole session. The pricing page, verbatim: "Tokens from past turns are re-processed and billed in every new turn, up to the configured context window size limit." A turn is one user input plus the model's response, and you pay for "all tokens present in the Session Context Window during that turn". With audio at 25 tokens per second and a 128K window, a long conversation gets more expensive per minute as it goes, and Google publishes no worked example. One mercy from the Provisioned Throughput docs: "output video avatar tokens aren't included within the session memory", so the avatar's own frames do not compound.
Nothing to cache, batch, or reserve cheaply. The Agent Platform model page marks context caching, batch inference, fixed quota, tuning, structured output, and code execution "Not supported". Provisioned Throughput is supported, with a catch: the Live API PT page requires "a minimum of 10 GSUs" before any avatar session can run on reserved capacity, and a smaller purchase sends "the entire video avatar-based session" to pay-as-you-go. Google's worked example lands 10 concurrent five-minute avatar sessions at 69 GSUs. No dollar figure from us; the GSU table does not state its time unit in the text we could read.
Session mechanics. Up to 1,000 concurrent Live API sessions per project on pay-as-you-go, no limit under Provisioned Throughput. The default session is 10 minutes, extendable without limit through goAway notifications and session resumption. Without context compression, audio-only sessions stop at 15 minutes and audio-plus-video sessions at 2 minutes; no page says whether that 2-minute cap governs avatar output or only camera input, and the model card's "a few minutes of continuous interaction, rather than extended hours" is the only avatar-specific duration Google gives.
No SLA names the Live API. Google's Gemini Online Inference API SLA (last modified June 30, 2026) covers generateContent and streamGenerateContent at 99.5% monthly uptime and never mentions BidiGenerateContent, WebSocket sessions, or avatars. Session resumption stores text, video, and audio for up to 24 hours; for zero data retention, Google's own guidance is to decline it.
Two migration notes. Asynchronous function calling (behavior: NON_BLOCKING) is now the default, and turnComplete: true no longer means the model is idle. And Google's two doc sets disagree on thinking: the Agent Platform page says "Thinking: Not supported", the Gemini API page says "Supported (interleaved reasoning)", and both pricing pages bill "text (response and reasoning)" output. Expect reasoning tokens on the bill either way.
Ninety-Seven Languages in the Headline, Twenty-Four in the Docs
Every headline this morning says 97 languages, because both Google posts do: "Gemini 3.8 Live understands and speaks 97 languages, with automatic language detection." The Gemini Enterprise docs, the only surface where Live Avatar exists, say otherwise. The Live API overview: "Multilingual support: Converse in 24 supported languages." The "Configure language and voice" page lists exactly 24, with BCP-47 codes, from Arabic (Egyptian) to Vietnamese. A third Google page, the Gemini API Live guide, says 99. No documentation page prints 97. We print 24 languages documented for the Gemini Enterprise Live API, note that the launch posts claim 97, and leave the gap for Google to close. Swedish and Mandarin are not in the 24-row table; test before you promise.
Where You Can Run It
Gemini Enterprise Agent Platform: the Live API over wss://{LOCATION}-aiplatform.googleapis.com, or the console at Agent Platform > Studio > Stream realtime, where you pick Live Avatar, an avatar, and a voice. Model availability, Provisioned Throughput, and pay-as-you-go are all "Multi-region: us, eu" plus us-central1; there is no global endpoint, and every pricing row reads "Non-global". Data residency, CMEK, VPC-SC, and Access Transparency are listed as supported. Retirement is "September 25, 2027 or later".
Gemini Developer API and AI Studio: gemini-3.8-live and gemini-3.8-live-extended-thinking, both stable since September 15, audio and text out only, free tier included. Extended Thinking exposes low, medium, and high thinking levels there; in Gemini Enterprise it remains in private preview. Consumer: 3.8 Live voice in the Gemini app for Workspace AI Pro and Ultra subscribers and in Search Live, no avatar. Predecessor: gemini-live-2.5-flash-native-audio stays GA on Agent Platform (video input only) until December 13, 2026. Not on Bedrock, Foundry, or OCI.
One thing we could not settle: whether Live Avatar needs a Gemini Enterprise subscription or seat, or only a Google Cloud project on a pay-as-you-go plan. The docs speak of a "pay-as-you-go (PayGo) plan" per project and never mention seats; the Gemini Enterprise pricing page returned 404. Dash.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
GA in Gemini Enterprise Sep 24, 2026; ID gemini-3.8-live; avatar via response_modalities: ["VIDEO"] | Verified (Cloud blog, Agent Platform model page and avatar docs) |
| Avatar video $1.00 per 1M at 6,192 tokens per second, billed while speaking; about $0.37 per speaking minute | Verified (Agent Platform pricing page; the per-minute figure is our arithmetic, Google prints none) |
| "97 languages" | Contradicted (Enterprise docs list 24; Gemini API Live guide says 99; only Google's blogs say 97) |
| "#1 overall spot on Artificial Analysis' Speech to Speech Quality Index (82.6)" | Verified for Sep 15, for Extended Thinking (The Decoder read the board that day; AA's metric has no "Quality" in its name; the GA avatar runs on base 3.8 Live) |
| "First previewed at Google Cloud Next 2026" | Unverified (not in the Next '26 wrap-up or the Agent Platform launch post; the SDKs gained avatar fields on April 9 and 13) |
| Whole session context re-billed every turn; no caching, batch, or fixed quota | Verified (pricing page billing details; Agent Platform model page) |
Thinking on gemini-3.8-live | Contradicted (Agent Platform page says "Not supported"; Gemini API page says "Supported (interleaved reasoning)") |
| Avatar latency in milliseconds; maximum avatar session length; prebuilt avatar count | Unverified (Google publishes none; the model card says "a few minutes") |
Who Should Use It
Use Gemini 3.8 Live with Live Avatar if you are already a Gemini Enterprise customer on Google Cloud, you want a branded face on a short, tool-calling support or shopping interaction (the Autotrader pattern), your languages are inside the documented 24, and a prebuilt avatar will do. Budget about $0.39 per speaking minute for video plus output audio, rely on the idle-is-free rule, and keep sessions short, because the model card says "a few minutes" and the whole context re-bills every turn.
Skip it if you build on the Gemini Developer API, the consumer app, or any non-Google cloud, because the avatar is not there. Skip it if you need a custom likeness or voice without a sales conversation and an allowlist, an SLA that names the Live API, sessions longer than a few minutes, or a language list longer than 24. Skip it if you need Extended Thinking behind the face; in Gemini Enterprise it is still private preview.
The frontier is elsewhere: for voice without a face, Google's own Extended Thinking variant held the top of AA's Speech to Speech Index at 82.6 on September 15 and OpenAI's GPT-Live 1 is $0.05 a minute, while the text models that do the reasoning behind these agents live on a different board, where Claude Opus 5.5 (max) leads the current v4.3.2 Intelligence Index at 58. Our model comparison has the current standings side by side.
Update, October 16, 2026
- October 14: xAI, now SpaceXAI, released Grok Voice 3.0 and cited Artificial Analysis's board of October 13 for a Speech to Speech Index of 86.9, taking the top spot, at the same $0.08 per minute as Grok Voice 2.0, now deprecated. Still speech to speech only, no video output.
- October 15: OpenAI added native Decisions integration to
gpt-live-1in the API; its API model list and changelog still carry nogpt-live-1.1and no video output. Our GPT-Live-1.1 note covers the ChatGPT side of that name. - October 16, Artificial Analysis: a GPT-Live-1 (Astra, medium) row at 83.2 now sits above Google; AA prints no dates, so we cannot say when it appeared. Gemini 3.8 Live Extended Thinking (High) still scores 82.6 and ranks third. Google's "#1" was true on launch day and is not true now. The same board lists base Gemini 3.8 Live at 76.0: time to first audio 1.18 seconds (Extended Thinking 1.35 s, Grok Voice 3.0 0.81 s), a higher Arena Elo (1088) than Grok Voice 3.0 (1070) or GPT-Live-1 Astra (1056), the lowest Agentic Performance of that group (30.1%), and $0.84 per hour of input audio against $3.50 for Extended Thinking, $4.80 for Grok Voice 3.0, and $5.83 for GPT-Live-1 (Astra, medium). Those are benchmark-run costs that include text and reasoning tokens, not list prices (OpenAI's $0.05 a minute is $3.00 an hour), and none includes avatar video. AA still has no avatar column, and its Intelligence Index remains v4.3.2.
- October 16, Google side, unchanged: the pricing page still shows the same six meters and 6,192 tokens per second; the Enterprise docs (last updated October 15) still list 24 languages, while the Gemini API Live overview (last updated October 6) says 70, a fourth Google figure. The Gemini API deprecations page lists
gemini-3.1-flash-live-previewandgemini-2.5-flash-native-audio-preview-12-2025for shutdown on November 17, 2026, withgemini-3.8-liveas the replacement.
Sources
- Google Cloud: Gemini 3.8 Live with Live Avatar is now generally available
- Google: Gemini 3.8 Live with Live Avatar
- Google: Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
- Gemini Enterprise Agent Platform: Generative AI pricing
- Agent Platform docs: Configure live avatars
- Agent Platform docs: Configure language and voice
- DeepMind model card: Gemini 3.8 Audio
- Artificial Analysis: Speech to Speech Index
- The Decoder: Google launches Gemini 3.8 Live to take on OpenAI's GPT-Live-1
*Not sure whether a metered talking head belongs in your support stack, or whether a plain voice line at about two cents a minute does the job? Take the free 2-minute quiz and get matched. Then read the next model note in this series: Claude Sonnet 5.5.*