Full-duplex or script-driven avatar: six questions before you pick

Tavus Griffin-Lite, reported on 2026-10-01, is a full-duplex video conversation model, in research preview for invited testers (per a third-party tracker, read 2026-10-05; I read no more detail than that). Sume Avatar 1.0 talking video is script-driven: you send a script or an ordered list of scenes, and a job returns a finished clip. Before you pick, answer six questions. They tell you whether you need a live conversation, or a rendered clip.

The six questions

Question 1 is the one that decides the rest. If a person has to interrupt the avatar, a script cannot give you that. If the avatar says the same lines to everyone, a script gives you a file you can review, caption and reuse.

What a script gives you that a call does not

A script-driven clip is a durable artifact. The result includes a public media.sume.com video, and may include preview_image_url and scene_previews. You can poll the job, read its events, and run the same text again with another quality tier: standard, plus (the default) or max.

You can also add silence beats. A scene with voice.type: "silence" has a required duration and no speech, so the avatar can hold a pause where a listener would nod. That is a scripted pause, not a reaction to a real person.

What a preview-only model leaves open

For Griffin-Lite, the only fact I can state is the research-preview status. Price, latency, resolution, API access and safety terms are things to ask the vendor. Do not fill them in from memory.

If a launch date matters to your plan, build the script-driven version first. The avatar, the script and the captions you make now will still be valid when a live option opens.

Decision in one line

• Need two-way conversation: wait for access, and ask the vendor the open questions.
• Need a finished clip for a feed, an ad or a lesson: use Avatar 1.0 now.
• Not sure: render one 15-second clip, read its preview stills first, and see whether a clip answers the need.