Griffin-Lite's 0.27 to 0.59 s latency in 25 fps frames vs Sume clips

Tavus says Griffin-Lite's video generation latency averaged 0.43 seconds on H100 GPUs, with a best case of 0.27 seconds and a worst case of 0.59 seconds (Tavus page, read 2026-10-09). At 25 frames per second, a common video rate, that is 6.75, 10.75 and 14.75 frames: about 7, 11 and 15 frames. Sume does not offer a live full-duplex model; it renders finished clips of at least 4 seconds.

Putting the latency in frames

The frame conversion makes the spread easier to feel. A 25 fps frame lasts 0.04 seconds. The best case is 6.75 frames behind, the average 10.75 and the worst 14.75. A viewer on a call would notice the worst case as a pause before the face reacts.

Availability, as Tavus states it

Griffin-Lite is a research preview that launched on 1 October 2026. The page says it will not be available for use for customers at this time, though it is available for select trusted testers, and that Tavus is working on safe disclosure features. The page gave no pricing for Griffin-Lite. Anything beyond those statements, including a launch date, is not on the page and is not claimed here.

Reading vendor numbers carefully

The Tavus figures are for video generation latency on H100 GPUs, as stated on the page. They are not an end-to-end conversation delay, and the page does not say what a viewer would see on a consumer connection. Treat the 25 fps frame conversion as a way to compare scale, not as a prediction of the experience.

The page also reports that the model is a research preview and gives no price. Any budget for a future live avatar is therefore a guess. A recorded clip has a price you can compute today: for example 15 seconds at plus is 15 x 0.245 = $3.675.

What Sume gives you in the meantime

Sume's Avatar 1.0 talking video is a recorded job: you send a script and poll a job. The docs accept an estimated 4 to 60 seconds. At standard that is 4 x $0.184 = $0.736 for the shortest clip, and 60 x $0.184 = $11.04 for the longest. The job returns a finished MP4 you can play with no live link, which suits prepared messages, not conversation.

If you are evaluating Griffin-Lite for a future live use, prepare what carries over: a consistent avatar, a script bank, and a place to label the clip as AI. Those are the parts you can build today with a recorded avatar and update when access opens.

• Live reaction in under 0.6 s: not on Sume.
• A prepared 4 to 60 second spoken clip with a reusable avatar: Sume Avatar 1.0.
• Poll with the job status URL; see Jobs and results.