Griffin replies in 0.43 s; a Sume video job takes minutes
Tavus says Griffin answers in 0.43 s on average at 720p, in a research preview. Sume video is async submit-and-poll, so use it for clips around a live call.

Tavus's Griffin page reports an average response time of 0.43 seconds (0.27 best, 0.59 worst) for a full-duplex video-to-video model that listens and speaks at the same time. A Sume video job is a different kind of call: you submit, it is queued or processing, and you poll or take a webhook until it completes, which Sume's docs say takes from about 30 seconds to several minutes.
So Sume clips belong around a live conversation, not inside it. Read on 2026-10-07: Tavus's Griffin page and Sume's video docs.
The two shapes
| Property | Tavus Griffin (as stated) | Sume video job |
|---|---|---|
| Interaction | Full-duplex video-to-video, listens while it speaks | Prompt in, finished file out |
| Response time | 0.43 s average, 0.27 s best, 0.59 s worst | About 30 s to several minutes by model and settings |
| Resolution | 720p | 360p to 4K depending on the row |
| Availability | Griffin-Lite research preview for select trusted testers; not for customers yet | Public API, pay per second |
Where a rendered clip fits
- Before the call: a 5 to 10 second branded intro, rendered ahead of time.
- After the call: a recap clip or a follow-up message.
- Between sessions: scene-setting footage for a lobby or waiting state.
- As a fallback: if the live model is unavailable, show a pre-rendered clip.
Planning for the wait
Sume accepts a valid job as queued when your workspace is at its concurrency limit and moves it to processing later, per the generation admission docs. If your product shows a clip right after a button press, render the clip earlier, not at the moment of the click. A webhook (callback_url on /v1/videos) avoids holding a poll loop open.
Be careful with the claims
Tavus states that Griffin-Lite is not available to customers at this time, so there is nothing to integrate today. The page's benchmark claims are Tavus's own and are not repeated here.
Handling the wait on the Sume side
Send an Idempotency-Key with each submit so a retry does not create a second job, and use a callback_url webhook rather than a tight poll loop if your service can receive one. On /v1/videos, the poll status is pending or in_progress until it reads completed.
What the Tavus numbers do and do not say
The figures are an average, a best case and a worst case from the vendor's own page, so plan for the worst case of 0.59 seconds, not the average. They describe reply latency in a live exchange, not the time to produce a finished, downloadable clip. Sume does not list Griffin in its video catalog, so there is nothing to compare row against row. The point of the contrast is the shape of the call, not a ranking.
Sources
Related posts
More in Comparisons
- Grok Imagine, Qwen Image and Imagen 4 Fast all cost 2.5 cents on Sume
Grok Imagine, Qwen Image and Imagen 4 Fast each bill $0.025 per image on Sume. What separates them: ratios, n, edits. Choose the cheap row that fits your job.
- H3 Max 1080p or Wan 3.0 1080p on Sume: 15 seconds costs $3.00 or $3.75
At 1080p MiniMax H3 Max costs $0.20 a second and Wan 3.0 costs $0.25 on Sume. The 15-second clip is $3.00 or $3.75; the features differ.
- HappyHorse 1.0 vs 1.1: which one renders 480P on Model Studio?
On Alibaba Model Studio only HappyHorse 1.1 lists 480P. Both versions take 3 to 15 seconds, and 1.0 adds a video-edit id. Neither is in Sume's catalog.
- Headshot background to neutral grey: cutout plus Pillow or an AI edit
Swap a headshot background for grey: Sume RMBG at $0.0225 plus a Pillow composite keeps the face pixels exact; an Ideogram 4.5 edit costs $0.075 or more.
Written by Sume