Google's AI Agents Now Have a Talking Face. It Costs About 21 Times the Voice, and Here Is Where That Is Worth Paying

On 24 September Google gave its real-time AI agent a face. Gemini 3.8 Live can now answer as a video avatar: a person, or a cartoon character, on screen, with lips that move with the words, holding a conversation in any of 97 languages. A day later Google made it generally available to businesses. It looks impressive, and for some businesses it will be worth having. The catch is on Google's own pricing page: once you do the arithmetic, a minute of the avatar talking costs about 21 times as much as a minute of the same agent talking by voice alone. This article sets out what Google shipped, what the face really costs, where we think it is worth paying for and where it is money thrown away, and what we would test before switching one on.

What Google Shipped on 24 and 25 September

Google announced Live Avatar on its main blog on 24 September. The Google Cloud blog followed on 25 September to say it is generally available, meaning ready for production and no longer a preview. It is an add-on to Gemini 3.8 Live, the real-time voice model Google released the week before. In business terms, this is what you get:

• A talking face. The agent can reply as video instead of audio only. Google says the avatar does "precise lip-syncing, natural expressions, and fluid turn-taking", and it keeps lip-syncing when the caller switches language partway through.
• 97 languages, detected automatically. The customer does not have to pick a language first. They just start talking.
• Eyes as well as ears. The agent can watch a live camera feed or a shared screen while it talks. Google's own demo is an insurance claim: the customer shows the damage on camera and the claim form fills itself in.
• Work in the background. The agent can look up an order or check a policy through your systems without leaving the conversation, and keeps talking while it does.
• Where it runs. Google names web pages, mobile apps and "interactive kiosks". It is available through Google Cloud from US and EU endpoints.

Two limits matter for a small business. Anyone can use the library of ready-made avatars, but a custom avatar made from a photo of a real face is, in Google's documentation, "only available to select customers". To get one you go through your Google Cloud account team and a verification process. And every second of audio and video the avatar produces carries SynthID, Google's invisible watermark, so it can always be identified as AI-generated.

The Price, Worked Out From Google's Own Pricing Page

Google's announcement gives no per-minute price. The pricing page does, but in tokens, so here is the conversion. The figures are Google's. The arithmetic is ours.

• Avatar video. Google bills avatar video at $1.00 per million output tokens, and says one second of avatar video is 6,192 tokens. That makes one minute of the avatar talking about $0.37.
• Voice alone. Audio output is $12.00 per million tokens, at 25 tokens per second of audio. That makes one minute of the agent talking by voice $0.018, a little under two cents.
• Listening is cheap. The customer's voice coming in is $3.00 per million tokens, about half a cent a minute. If you turn on the camera, a picture a second at 258 tokens each adds about $0.015 a minute.
• Silence is free for the face. Google says video charges "only apply when the avatar is actively speaking". While it listens, you pay nothing for the face.

So the face is about 21 times the cost of the voice it sits on. Google's pricing page lists avatar video as its own line and does not say whether you also pay for the audio when the avatar speaks. Our advice is to budget for both until your first invoice shows otherwise. Here is what that looks like at three sizes:

None of these numbers will break a business. But the difference is not a rounding error either. On the middle row, the face costs about $167 a month on top of the voice, and it only makes sense if it brings in more than $167 a month in sales or saves that much in staff time.

Where a Face Earns Its Keep

A face does not make the AI any smarter. It is the same model, the same answers and the same mistakes. What it changes is how much attention the customer gives it and how much they trust it. That matters in a few specific places:

• A screen in a physical space. A reception desk, a hotel lobby, a showroom, a clinic waiting area. Google gives hotel check-in as an example. A screen with a face on it gets walked up to. A screen with a microphone icon mostly gets ignored.
• A guided decision on your website. Cox Automotive is using it in Autotrader to walk car buyers through search, comparison and finance, highlighting parts of the screen as it talks. When the customer is choosing between options that cost real money, being guided by a face feels like being served, not processed.
• Showing, not telling. With the camera on, the customer can point at a broken part, a damaged parcel or a meter reading. The face matters less here than the eyes, but the two come together.
• A counter that serves many languages. If your walk-in customers speak three or four languages, one screen that switches language as soon as someone starts talking is a genuine new option. Until now you needed a person who happened to speak each one.

A face is a display feature, not an intelligence upgrade. Pay for it where your customer is already looking at a screen, and nowhere else.

Where a Face Is Money Thrown Away

For most of the small businesses we work with in Pakistan and the Gulf, most customer conversations happen on channels where there is nothing to look at:

• The phone line. A caller cannot see a face. Every minute of a phone call is voice-only, and for that the voice-only rate is the whole bill. This is where most after-hours leads are won or lost, and it is the job an AI voice agent on your phone line is built for.
• WhatsApp and website chat. Most people who type a question want a typed answer. Starting a video in reply to "what time do you close?" is slower and more intrusive than the question needs. A text assistant such as our AI Chat Assistant answers that for a small fixed monthly fee.
• Long explanations. You pay for every second the avatar talks. An agent that reads out your whole returns policy on camera is now expensive as well as boring.

Our opinion: most small businesses should not start with a face. Get a voice or chat agent answering well first. It is the cheaper part and it does most of the work. Add a face only at the one screen where people already stand and look.

The Small Print: Custom Faces, Consent and a Watermark

The first thing many owners will ask is "can it look like me, or like my receptionist?" Technically yes, and Google's own demo builds one from a single photo and an audio sample. In practice there are three hurdles:

• Access. Custom avatars are limited to select customers, approved through a Google Cloud account team. A small business on pay-as-you-go should expect to use the ready-made library for now.
• Consent is your job. Google's documentation puts the responsibility for "all consents and rights" for the face and voice on you. If a staff member lends their face and later leaves, that face should not keep talking to customers. Put it in writing before you build it.
• Rules on the photo. No images of children, no celebrities, and a neutral, front-facing head-and-shoulders shot at 720p or better.

The SynthID watermark means the video can always be identified as AI-made. We would go further and simply say so on screen. A customer who works out halfway through that the friendly face was not a person trusts you less than one who was told at the start. A small "AI assistant" label costs nothing.

One more point for readers in Pakistan and the Gulf: Google says the feature runs from US and EU endpoints. If your customers' conversations are covered by rules about where data may be processed, check that before a camera-enabled agent starts seeing faces, ID cards or documents.

What We Would Test Before Paying for a Face

We run always-on AI for our own customers, and the lesson that applies here is simple: measure the outcome, not the demo. A face always wins the demo. Whether it wins the month is a separate question. Before rolling it out, we would:

• Run it side by side. Half your website visitors get the voice agent, half get the avatar, for two weeks. Compare how many finish the conversation and how many book or buy, not how many say "wow".
• Write short answers into the instructions. The face is billed per second of speech, so "answer in two sentences, then ask a question" in the agent's system instructions directly cuts your bill.
• Turn the camera on only when it is needed. Camera input is cheap per minute, but it is also where privacy risk comes in. Ask for it only when the customer needs to show something.
• Set a monthly cap in Google Cloud. Put a budget alert on the project before launch, so a popular week does not turn into a surprising invoice.

Our Take, and Where to Start

Live Avatar is a real new capability, and at about 37 cents a talking minute it is within reach of a business with one busy screen. It is not a reason to put a face on every conversation. The smartest setup for most businesses is a layered one: voice on the phone, text on WhatsApp and the website, and a face only at the kiosk or on the one page where a customer is making a big decision.

If you are pricing the voice part first, our breakdown of what an AI voice agent costs per minute covers the whole bill, including the small print that applies whether or not there is a face. Our earlier guide to AI voice agents and missed calls explains where the money actually goes. And if you want help deciding which of your customer conversations deserve a face, a voice or a simple text reply, talk to us. We will tell you honestly if the answer is "not yet".