2026-10-07

Tavus Griffin-Lite claims vs a script-driven avatar API: a checklist

Griffin-Lite is full-duplex and a research preview for select testers. Check each claim against what you can integrate today.

5 min readSume

All posts

Can you build on Tavus Griffin-Lite today? Tavus lists it as a research preview for select testers only, so the practical answer is no for a public launch. What you can ship now is a scripted, asynchronous avatar video: Sume Avatar 1.0 takes a script, returns a job, and gives you a 720p MP4.

Tavus introduced Griffin this month as a full-duplex model: it listens and speaks at the same time and reacts to interruptions instead of waiting for a turn to end. That is a different product from a rendered clip, and the checklist below separates the two.

What Tavus says about Griffin-Lite

From the Tavus Griffin page: Griffin is described as a Human Interaction Model that handles audio and video together; Griffin-Lite averages 0.43 seconds of latency; it generates 720p video in real time; in a face-to-face study 48% of participants believed Griffin-Lite was human; and it is available only as a research preview to select testers. Tavus says further alignment and safety procedures are required before a safe release.

Those are vendor statements for a live, interactive system. They tell you nothing about a script-to-MP4 job, and the reverse is also true.

Claim by claim

Use this table to decide which side of the line each requirement is on.

Live full-duplex vs rendered avatar clip (Tavus page and Sume docs read 2026-10-07)
Requirement	Griffin-Lite (Tavus page)	Sume Avatar 1.0 (docs)
Listens while it speaks	Yes, full-duplex	No: a script goes in, a video comes out
Public availability	Research preview, select testers	Developer API with a key
Delivery	Real-time video	Async job: poll /v1/jobs/{id}/status then /result
Resolution	720p	720p at this time
Input	Live conversation	script or ordered video_inputs, 4-60 s
Review before spend	Not described	Preview first frames, then generate-video

What to prepare while you wait

If your product is a live conversation, plan consent, a disclosure line and a fallback now, because Tavus itself flags disclosure and safety as the gate to a wider release. If your product is an ad, a micro-drama episode or a spokesperson clip, you do not need a duplex model: you need a stable avatar handle, a script, and a review step.

The review step is concrete in Sume's API: create an avatar video preview, check scene_previews[], then call generate-video. Quality can be overridden at that last call.

Decision rule

Choose live when the viewer's reply changes what the avatar does next. Choose rendered when you can write the lines in advance and want to post the same file everywhere.

What to do this quarter

A research preview for select testers is not an integration plan. Write down the behaviours you need from a live avatar, such as interruption and latency, and keep a scripted fallback that you can ship today. Revisit the Tavus page when access opens, and re-run the checklist against the new terms.

Do not promise real-time to customers yet.
Ship scripted episodes on an async API.
Re-read vendor pages before you cite them.

Keeping the comparison honest

Compare like for like. A full-duplex model is a live conversation; Avatar 1.0 is a scripted render. They solve different problems, and the right choice depends on whether the viewer talks back. For a published episode or ad, a script is an advantage because you can approve it. For a live agent, you need the live model once it ships.