Is there a Tavus Griffin-Lite API? Not yet, and what to use

No public Griffin-Lite API is available. Tavus's Griffin page says the model "will not be available for use for customers at this time, though it is available for select trusted testers", and access is by a request form (read 2026-10-04). If you need to ship talking-face video this week, you need something that has an endpoint today.

What Tavus published

The page, dated 1 Oct 2026, describes Griffin-Lite as a research preview of Tavus's first Human Interaction Model, a full-duplex video-to-video model. It reports 720p generation in 320 ms chunks with 0.43 s average audio-to-video latency on H100 GPUs, and that 48% of people believed it was a real person after a one-minute call (read 2026-10-04).

It also reports VideoFDB scores of 3.83 out of 5 for generation and 3.73 for perception. Tavus says further alignment and safety procedures are required, and anticipates a release very soon after the safety concerns are addressed.

What Sume ships for a speaking face

Sume's avatar video endpoint turns a script into a finished clip. It is asynchronous rather than full-duplex: you submit, poll and download. The docs list scripts of 4-60 s, aspect ratios 1:1, 3:4, 9:16, 4:3 and 16:9 with 9:16 as the default, 720p output and three quality levels: standard, plus and max. See the avatar video docs.

That is a different product shape. If your use case is a live, interruptible conversation, a clip API does not replace it. If it is a prerecorded explainer, a greeting or a product demo, it does.

How to decide

Compare the two on three points.

Our recommendation for this week

Treat Griffin as research you can read about, not infrastructure you can plan on; the Tavus page makes no promise of a date. Build your pipeline on what is documented, and the stored post on what to render today lists concrete jobs.