Learn
A waiting agent costs a download, not a video call
Most visitors never talk to the agent on your site. Here is what it costs to have one there anyway.
Most people who visit a website look around and leave. They never talk to the agent on the page. So the question that matters is simple: what does the agent cost while it is only waiting?
The answer depends on how the face gets onto the page. There are three ways.
| Approach | How it works | GPU in a data centre? | Who works this way |
|---|---|---|---|
| A living avatar | The face is sent to the visitor’s device once, together with the page. From then on it lives there. The device draws it and moves it, and the visitor can talk to it. | No | Only Somebody AI |
| Like a video call with a supercomputer | A GPU in a data centre draws the face and sends it as live video, to one visitor at a time. A GPU is the specialised hardware that AI runs on. It is expensive and in short supply. | Yes, for every second of the conversation | Most avatar platforms |
| Like a remote-controlled puppet | The visitor’s device draws the face, but a GPU in a data centre still works out every movement and sends it as a constant stream. | Yes, for every second of the conversation | Some avatar platforms |
Only a Somebody agent is a living avatar: once it has downloaded, you can talk to it, and no GPU in a data centre is involved. Every other avatar needs one, whether it sends the whole video or only the movements.
When someone does talk: only the voices travel
During a conversation, the only thing going back and forth is sound: the visitor’s voice one way, the agent’s voice the other. The face, including the moving lips, is still drawn on the visitor’s own device.
On other platforms the face is the most expensive part of the call. They send it as live video, or at the very least as a constant stream of instructions for how the face should move. Either way it takes specialised servers, GPUs again, working for every second of every conversation.
The delay is trivial by comparison. The visitor’s device holds the voice back for a brief moment, only so that the lips can move smoothly in step with it. A video call has to draw every picture in a data centre and send it across the internet before the visitor sees anything.
Carrying huge numbers of voice conversations at the same time is something the industry already knows how to do well. That is the only part we have to carry.
Video calls also run out of room
- Only so many at once. Each video call needs its own share of a GPU, the same specialised AI hardware everyone else is competing for. There is a fixed number of those, so every plan has a limit on how many calls can run at the same time. One visitor too many, and that visitor waits or sees no avatar.
- More room costs more. To raise the limit, GPUs have to be kept ready for you, used or not. You pay for your busiest minute, not your normal one.
- Visitors come in crowds. 10,000 visitors a month is about 14 an hour on average. But a launch, an email or a mention can send 500 people in ten minutes, and those are the visitors you most want to greet.
| 500 visitors in ten minutes | Somebody AI | Others |
|---|---|---|
| Does everyone see a live face? | Yes | Only up to your plan’s limit |
| What you arrange in advance | Nothing | Extra room, paid for ahead of time |
We are only bound by the audio conversation. A Somebody AI agent’s limit is how many voice conversations can run at once, not how many GPUs are free, and voice is the part the industry already runs at scale.
While it waits, a Somebody AI agent costs nothing to run
An idle Somebody AI agent costs nothing to run. It sits on the visitor’s own device, so no server and no GPU is working for it, whether the visitor stays for a minute or an hour.
We left our homepage open for one minute with nobody talking, four times, and counted.
| In one minute of waiting | Test 1 | Test 2 | Test 3 | Test 4 |
|---|---|---|---|---|
| Messages to our computers | 0 | 0 | 0 | 0 |
| Data sent or received | 0 | 0 | 0 | 0 |
| Open connections | 0 | 0 | 0 | 0 |
| Live video playing | 0 | 0 | 0 | 0 |
Nothing is sent, nothing is received and nothing is kept open. The clock starts only when someone starts a conversation.
The others cannot wait for free. A video call costs money for every second it is open, so each company has a rule for when nobody is talking. This is what their own help pages said on 6 November 2026.
| Company | Its rule for quiet time |
|---|---|
| Tavus | An open conversation keeps counting while it is quiet. It closes after five minutes with nobody in it. |
| Anam | Silence does not pause the clock. |
| Lemon Slice | Closes after 60 seconds of quiet. Shows a repeating welcome video until the visitor clicks. |
| bitHuman (cloud) | Charges per second, quiet or talking. Starts when the visitor taps. |
| Beyond Presence | Closes after five minutes of quiet. |
| HeyGen LiveAvatar | Checks every minute and charges for the time that has passed. |
This is why a video-call avatar usually shows a still picture or a repeating clip until someone clicks, and hangs up soon after things go quiet. A Somebody agent has no call to hang up, so the live face is there from the moment the page opens.
The only cost before talking: two downloads
| When | What arrives | Size |
|---|---|---|
| The page opens | The agent: its face, background and player | about 5.9 MB |
| A conversation starts | A library and AI models, loaded once, not once per avatar | about 15 to 16 MB |
Both are ordinary files, handed out the way any picture on a website is, and both then run on the visitor’s own device. Visitors who only look never receive the second one.
What it means for your website
| Situation | Somebody AI | Others |
|---|---|---|
| Visitors who only look | Cost nothing to run | A call that is charged for, or hung up |
| A crowd arrives | Everyone sees a live face | Only up to your plan’s limit |
| Someone talks | Only the voices travel | A GPU works for every second |
| What arrives on the device | Two downloads, once | A live stream, all the time |
With 10,000 visitors a month and 3 in every 100 starting a four-minute conversation, that is 1,200 minutes of talking and 9,700 visitors who cost nothing. Try the pricing calculator with your own numbers.