Real-time AI avatar vs video avatar API: the difference

A real-time AI avatar is a live face that listens and answers during a session, rendered while the conversation happens; a video avatar API renders a finished video file from a script you send, and you download it after the job completes. They are different products: HeyGen and Synthesia each sell a live avatar API next to their video API, while Sume's avatar API renders files only.

Vendor facts below come from HeyGen's and Synthesia's own docs, read on 2026-09-28. Sume facts come from Generate avatar video, Webhooks, and the Sume API reference.

How does a real-time avatar work?

Synthesia's docs put it plainly: an Interactive Avatar "isn't a video you generate and download. It's a live participant that Synthesia renders in real time" and places into a LiveKit room you control. Your own agent handles listening and turn-taking, and Synthesia renders the lip-synced face.

HeyGen's LiveAvatar works in two modes. In FULL mode HeyGen manages the whole conversation (speech-to-text, LLM, text-to-speech, turn-taking, memory). In LITE mode you keep your own agent and LiveAvatar supplies only the avatar layer.

Either way, both vendors bill live use by the minute, and the output is a live stream, not a file you keep.

Which vendors sell live avatars, and which render videos?

Listed alphabetically. Prices and limits are as each vendor states them, in the units it uses.

What does a video avatar API return?

A rendered avatar video is a job, not a stream. On Sume you send an avatar_handle and a script to POST /v1/avatar-1.0/talking-video, and in the default async mode the response carries the job's status_url and result_url right away. The sync mode waits at most 30 seconds and bounds the HTTP wait, not the job.

When the job completes, the result can include a public media.sume.com video. Webhooks are terminal events only (job.completed, job.failed, job.canceled); there are no progress or partial deliveries.

Each video covers an estimated 4-60 seconds of script, billed per second at $0.184/s standard, $0.245/s plus, $0.55/s max (no product image). The request itself is covered in Talking avatar video API.

Can Sume run a live, talking-back avatar?

No. Sume's avatar video and speech routes are jobs you poll or get a webhook for; the API reference describes TTS 1.0 as an async job with poll or webhook, non-streaming. There is no session, room, or stream to join; What is an AI avatar? covers the same limit.

What Sume fits is a recorded answer: the words are known before anyone watches. FAQ videos with an AI avatar shows that pattern, one clip per question.

Which one do I need?

• Pick a real-time avatar if a person talks to it and the answer depends on what they say: a support face on a website, a practice conversation, a kiosk. Budget for per-minute session time.
• Pick a video avatar API if the script is written first: training steps, product explainers, FAQ answers, ads. You get a file you can review, caption, and host anywhere.
• Check the live product's limits before you build. Synthesia's operational trust page lists no session transcript, recording, or webhook system, bust framing only, and no stock avatars in live sessions.
• To have an AI agent decide what an avatar says, see AI avatar vs AI agent.