Learn
What does a waiting agent cost? We measured it
We left a live Somebody agent on a page for a minute and recorded everything it sent and received. The answer was nothing. Here is the method, the result, and how streamed avatars bill the same wait.
Most visitors to a page never talk to the agent on it. They look, read and leave. So the cost of an agent that is present matters more than the cost of one that is talking, because presence is what you pay for most of the time.
We say a Somebody agent costs nothing while it waits. That is a claim you should be able to check, so we measured it.
What we measured
On 6 November 2026 we opened somebodyai.co in an automated Chromium browser at 1440 × 900, waited for the agent’s face to appear, let the page finish loading, and then recorded the next 60 seconds without touching anything. We counted:
- every HTTP request the page or the agent made, and the bytes transferred;
- every WebSocket opened or used;
- every WebRTC connection created, which is how live video and audio streams are carried;
- every video element playing a live media stream.
We ran it four times.
The result
| During 60 seconds of waiting | Run 1 | Run 2 | Run 3 | Run 4 |
|---|---|---|---|---|
| HTTP requests | 0 | 0 | 0 | 0 |
| Data transferred | 0 KB | 0 KB | 0 KB | 0 KB |
| WebSocket connections | 0 | 0 | 0 | 0 |
| WebRTC connections | 0 | 0 | 0 | 0 |
| Live media streams | 0 | 0 | 0 | 0 |
Once the page has loaded, a waiting agent talks to no server at all. The face is drawn by the visitor’s own device. No session exists on our side, so there is nothing to meter and nothing to bill. Minutes start when the visitor starts a conversation and stop when it ends.
What this does not show
- The first load is not free of bytes. The agent’s face and player are downloaded once with the page, like any image or script. The measurement covers what happens after that.
- It is one page and one browser. We measured our own homepage in desktop Chromium. We have not published figures for phones here.
- A conversation does use minutes, including its pauses. Once someone is talking to the agent, time counts until the conversation ends.
How streamed avatars handle the same wait
Most live avatars are video. A server renders the face on a GPU and streams it to the visitor. That stream needs a session, and a session is what gets billed. Here is what each vendor’s own documentation said on 6 November 2026.
| Platform | What its documentation says about idle time |
|---|---|
| Tavus | Billing is based on active session runtime, not only the time spent speaking, and an open, idle conversation can keep accruing usage. A conversation with no participant closes after five minutes by default. |
| Anam | Silence, a muted microphone or an idle avatar does not pause an active session. |
| Lemon Slice | The idle timeout defaults to 60 seconds and resets when the avatar is talking. The website widget plays a looping welcome video until the visitor clicks. |
| bitHuman (cloud) | Credits pay for active session time, talking or idle, by the second. Its web embed can also render in the browser, and starts on tap. |
| Beyond Presence | Managed agents time out after five minutes of inactivity to prevent unnecessary credit usage. |
| HeyGen LiveAvatar | Active sessions are checked every minute and credits deducted for the elapsed time. |
This is why streamed-avatar widgets usually show a still image or a looping clip until the visitor clicks, and close idle sessions after a minute or a few minutes. It is a sensible design for a streamed product. It also means the face on the page is not the agent until someone starts a session.
Why it matters for a website
Take a page with 10,000 visitors a month where 3% start a four-minute conversation. That is 1,200 minutes of conversation and 9,700 visitors who only look. With a Somebody agent, those 9,700 see a live face and cost nothing. You can run the numbers for your own traffic with the pricing calculator.
Check it yourself
Open any page with a Somebody agent, open your browser’s developer tools, select the Network tab and wait. After the page has loaded you should see no requests from the agent until you start talking.
