Gemini 3.8 Live Avatar Is Generally Available—Here Are the API Price and Limits

Google moved Live Avatar into general availability on September 25

Google announced Gemini 3.8 Live with Live Avatar on September 24, 2026. On September 25, 2026, Google Cloud said the feature was generally available in Gemini Enterprise and “ready for enterprise production.” The callable model is gemini-3.8-live. Google documents US and EU endpoints and says provisioned throughput is available. Gemini 3.8 Live Extended Thinking is a different product and remains in private preview.

This is a model and API launch, not merely a demo. Developers can build web, mobile and kiosk experiences that accept live audio, images or video and return spoken responses with a synchronized video avatar. Availability does not establish workload fitness: AccessAllGPT did not authenticate an account or verify entitlement, regional routing, quota or production behavior.

The API has six modality-specific prices

Google’s Agent Platform pricing table lists Gemini 3.8 Live API in a non-global region at $0.75 per million text-input tokens, $1 per million image/video-input tokens and $3 per million audio-input tokens. Output is $4.50 per million text tokens, including response and reasoning, $12 per million audio tokens and $1 per million avatar-video tokens. The table displays the same rate above and below 200K input tokens, although the model card separately caps Live context at 128K.

The table marks cached input as unavailable and says input and output are charged only for requests returning HTTP 200. Those are token rates, not a call, minute or completed-conversation price. A useful budget must meter each modality and include retries, tools, idle connection policy, provisioned capacity and the application around the model. AccessAllGPT did not receive a billing receipt and does not infer how many video or audio tokens a representative minute consumes.

Live Avatar changes the output contract and the limit

The Gemini 3.8 Audio model card says standard Gemini 3.8 Live accepts audio, images, video and text with a token context window up to 128K, and returns audio and text with up to 64K output. With Live Avatar, output becomes audio, video and text, and the listed output limit falls to 24K. The API overview describes avatar video as MP4 and the session transport as a stateful WebSocket connection.

The same model card says Live Avatar can support a few minutes of continuous interaction rather than extended hours. That sentence is a material product boundary for reception, coaching, claims, education and kiosk plans. Design a visible handoff, restart or non-avatar continuation path; do not promise an open-ended video conversation from a launch reel.

The avatar can keep talking while a tool runs

Google says asynchronous function calling lets the model trigger tools or API calls in the background while dialogue continues. The Cloud launch gives a hotel check-in style interaction as an example, and says interruption recovery can preserve conversation context and backend transactions. These are vendor descriptions, not results independently reproduced by AccessAllGPT.

Concurrency creates an application-state problem. The avatar may acknowledge work before a tool succeeds, fail after an optimistic statement, or receive a user correction while the previous action is still running. Keep the tool state authoritative outside the model. Surface pending, succeeded, failed and cancelled states; use idempotency keys; and require a fresh policy check before money, messages, records or permissions change.

“97 languages” and “24 supported languages” need scope clarification

The launch article says Live Avatar can transition across 97 languages while adapting lip-sync and expressions, and the Cloud announcement says Gemini 3.8 Live understands and speaks 97 languages with automatic detection. The Live API overview retrieved September 27 instead says “Converse in 24 supported languages.” Google may be describing different layers or levels of support, but the reviewed pages do not reconcile the numbers.

Do not turn the larger launch claim into a support guarantee. For each required language, verify input understanding, generated speech, voice availability, code switching, transcript quality, lip synchronization, policy behavior and fallback. Record whether the result uses a documented supported language, a demonstrated capability or an unsupported observation. The discrepancy itself should remain open until Google publishes a scope map.

Custom avatars are not self-service general availability

Google offers a curated library of preset avatars. Creating a custom avatar from a reference image—and, in the Cloud demo, an audio sample—requires enterprise allowlisting and verification. The general availability of the model therefore does not mean every customer can immediately upload a person or brand character and deploy it.

Procurement should separate base-model access, Live Avatar entitlement and custom-avatar approval. Before submitting a likeness, establish documented consent, usage scope, revocation, retention, deletion and incident ownership. Test look-alike rejection and account offboarding. These are AccessAllGPT controls, not claims that Google’s verification fails.

SynthID is transparency metadata, not the whole misuse control

Google says all generated Live Avatar audio and video streams carry imperceptible SynthID watermarks. That can support provenance and detection, but the reviewed launch pages do not publish a false-positive rate, false-negative rate, survival result for every transformation or service-level commitment for downstream detection. AccessAllGPT did not generate or inspect a stream.

Make the experience visibly disclose that the participant is AI, even when watermark detection is unavailable. Preserve consent and generation records, restrict export where the use case does not need it, and give users a route to a human. A hidden watermark should complement—not replace—clear presentation, authorization and abuse response.

The model card names limits the demos do not show

Google says Gemini 3.8 Audio is based on Gemini 3 Pro and delegates architecture, training-data, implementation and several safety details to that model’s card. It names hallucinations, occasional slowness or timeouts and a January 2025 knowledge cutoff. The “few minutes” avatar-duration limit is also in the card rather than the launch headline.

The card says Google performed automated, human, child-safety and red-team evaluations, and concludes the 3.8 Audio variants are unlikely to reach a tracked or critical Frontier Safety capability level by reference to Gemini 3.7 Flash. This is a vendor safety assessment, not an independent test of a particular avatar application, identity workflow or tool policy.

Run a conversation-and-action acceptance test

Start with one preset avatar, one endpoint and a narrow read-only workflow. Freeze cases for interruption, cross-talk, silence, camera loss, language switches, unsafe input, tool delay, tool denial, duplicate tool completion, user correction and the published session-duration boundary. Record first audio, first video, end-to-end turn time, transcript differences, tool-state accuracy, reconnect behavior, tokens by modality and cost per accepted conversation.

Then test the consequence boundary with inert tools. The avatar must not claim completion before durable tool success, repeat a side effect after reconnect, or conceal an incomplete action behind fluent conversation. Evaluate accessibility without assuming a human-like face helps every user: include captions, keyboard operation, reduced-motion needs, screen-reader flow, audio-only fallback and a non-avatar route.

The decision: pilot a bounded interaction, not a synthetic employee

Pilot Live Avatar when visual presence materially improves one short, measurable interaction and the team can obtain the needed entitlement, meet consent requirements, meter all modalities and provide a deterministic fallback. Use a preset avatar first. Add a custom avatar only after allowlisting, identity rights and removal procedures are confirmed. Keep consequential tools behind application-owned state and authorization.

Wait when the workflow needs hours-long continuity, an unverified language, guaranteed global routing, a custom likeness without approval, or a cost model that assumes audio and video tokens behave like text. Reject the avatar layer when plain voice, text or a human channel meets the outcome with less identity, accessibility and operational risk. General availability answers whether Google sells the service now; it does not answer whether the service belongs in a particular customer interaction.

Disclosures

AccessAllGPT did not receive early access, API credits, a Google Cloud account, an enterprise allowlist, product briefing, test data or compensation for this article. Google and Google DeepMind did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google, Google DeepMind, OpenAI or organizations cited. Publication-wide relationships are listed on the disclosures page.