Atmee avatar billed by the minute vs Sume reserve and refund
Atmee's LiveKit plugin bills by the minute, capped at 3,600 seconds by default. Sume reserves a clip price upfront and refunds failures. Compared.
Atmee's LiveKit plugin bills by the minute from the moment the avatar sees your agent in the room, and max_duration_seconds defaults to 3600, the most a session can bill if your agent dies without a trace (read 2026-10-05). Sume's lip-sync and avatar jobs work the other way: the price for a clip is known at submit, held as a reservation, captured on completion and refunded on failure.
What the Atmee page says
The plugin README states that billing starts when the avatar sees your agent and runs by the minute. It lists what stops billing: the agent leaving or the room closing, calling aclose(), a rendering worker failure (sessions silent for minutes are closed), or hitting max_duration_seconds.
The default of 3600 is described as a hard ceiling. If you run short sessions, set it lower in code.
Sume's shape
On Sume you ask for a finished clip. For H3 Max lip sync the price is ceil(duration_seconds) times the per-second rate, and the API reserves it at admit. Completion captures it and failure refunds it. There is no open session to leave running, so no ceiling to configure.
The cost is that you give up interactivity. You wait for a job, then play a file. Choose by what the user needs: a conversation, or a prepared answer.
| Question | Atmee plugin | Sume lip-sync job |
|---|---|---|
| Unit | Minutes in the room | Seconds of audio, rounded up |
| Known before start | Only the ceiling, 3,600 s by default | Reserve at admit |
| Runaway risk | Bounded by max_duration_seconds | None; the job ends |
| On failure | Worker failure closes silent sessions | Reservation refunded |
| Interaction | Live | Asynchronous clip |
Budget both
For a live session, price the worst case as the ceiling in minutes times the rate you were quoted, and set the ceiling to match your longest legitimate call. For a Sume clip, use admission preview before submitting large banks. The 1,000-reply math is in the cost post.
Sources
Related posts
More in Comparisons
- Best API for text to speech: six checks to run before you pick
Compare TTS APIs on price per million characters, request cap, latency, languages, controls and word timing, with MAI-Voice-2.1 and Sume TTS 1.0 numbers.
- Beyond Presence 50 credits a minute vs Sume lip sync per second
Beyond Presence bills live speech-to-video at 50 credits a minute. Sume bills a rendered lip-sync clip per second. Here is how to compare the two units.
- bitHuman local .imx or cloud avatar vs a hosted Sume avatar job
bitHuman renders avatars locally from .imx files or in its cloud, which changes your CPU bill. A hosted Sume avatar job moves the render off your agent.
- Can an AI avatar see me on a video call? Griffin perception vs Sume
Tavus says Griffin perceives visual context: 3.73 of 5 on VideoFDB perception vs 4.20 for humans. Sume avatar clips cannot see you, but a clip can be inspected.
Written by Sume