Google's Gemini avatars go GA, and the video bills at 6,192 tokens a second

Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise on 24 September, adding a generated, lip-synced video face to the live dialogue models it announced the week before. Animated avatars are not new, and the feature itself is incremental. What is genuinely useful is that Google published the meter: the avatar's video stream bills at 6,192 tokens for every second it speaks.

That number does most of the work. At the list rate of $1.00 per million tokens for avatar video output, a speaking avatar runs about $0.37 per minute, or roughly $22 per hour of actual speech. The audio it is lip-syncing to bills at 25 tokens per second against a $12.00 per million rate — about 1.8 cents a minute. Putting a face on a voice agent therefore costs on the order of 21 times what the voice alone costs, before any input or reasoning tokens enter the calculation.

What actually shipped

Live Avatar pairs near-real-time video generation with Gemini's native speech-to-speech dialogue, so the model listens, sees and speaks through a visual persona instead of returning audio alone. Google says it holds lip-sync and facial expression across 97 languages with automatic language detection, and can switch language mid-conversation without, in its words, "degrading video fidelity or introducing visual drift."

• Asynchronous tool calling — the model triggers tool and API calls in the background and keeps talking while they run, rather than going silent during a lookup.
• Live visual understanding — camera feeds and screen shares are processed alongside audio in the same session.
• Avatar library and custom avatars — preset avatars are available to all customers; custom avatars generated from a single high-quality reference image are gated behind enterprise allowlisting and verification.
• SynthID watermarking — embedded in every generated audio and video stream.
• US and EU endpoints, with provisioned throughput available through a Google Cloud sales representative.

Gemini 3.8 Live Extended Thinking, announced alongside the base Live model, remains in private preview. Google says Live Avatar was first previewed at Cloud Next 2026.

The billing mechanic that matters more than the rate card

The part of the pricing page that will shape real bills is not the per-token rate. The Live API charges per turn for everything sitting in the session context window, and Google states plainly that tokens from past turns "are re-processed and billed in every new turn," up to the configured context window limit. A live voice session accumulates context quickly: at the documented rate of 25 tokens per second of audio, ten minutes of conversation is roughly 15,000 tokens — and every subsequent turn pays for all of them again.

No discount is listed for that. The Gemini 3.8 Live API rows on Google's pricing table show cached input pricing as N/A, while Gemini 3.1 Pro Preview on the same table lists cached input at $0.20 and $0.40 per million tokens. Whatever context caching does for batch and chat workloads, it is not offered here. Long conversations are the expensive case, and the avatar's 6,192 tokens per second rides on top of them — though Google does note that video is billed only while the avatar is actively speaking, not while it listens.

What Google has not shown

The announcement carries no latency figures and no benchmarks. Every quality claim attached to it comes from customers rather than from Google's own published measurements or from independent evaluation.

Gemini 3.8 Live improved interruption handling, multilingual conversations, and tool-call reliability.

Equal AI also says it handles over a million live calls daily across nine Indian languages. Cox Automotive describes an Autotrader shopping assistant built on the feature, and Salesforce says it is exploring pairing Gemini 3.8 Live with Agentforce. These are vendor-supplied testimonials and are worth reading as such. At the time of publication, no independent testing of Live Avatar was available.

If you are building voice agents

The practical read is that asynchronous tool calling is the part worth caring about, and it does not require the avatar. Dead air while a tool runs is the most reliable way to make a voice agent feel broken; letting the model acknowledge a request and keep talking while a lookup completes addresses that directly. The face is a separate decision with a separate meter.

If you do turn it on, model it as roughly 37 cents per speaking minute layered on top of audio and reasoning costs, and check whether visual presence earns that against a plain voice session running under two cents a minute. One scope note before you plan around it: this is a Gemini Enterprise release, not a general Gemini API feature — Google's public Live API documentation does not mention Live Avatar, and custom avatars require allowlisting on top of that.