Griffin-Lite vs Sume Avatar 1.0: what you can integrate this week

Compare what Tavus Griffin-Lite shows on its page with what Sume Avatar 1.0 ships in its API: access, input, output, latency claims, tiers and disclosure.

5 min readSume
All posts

This week you can integrate Sume Avatar 1.0 and you cannot integrate Tavus Griffin-Lite. Griffin-Lite is a research preview for select trusted testers that Tavus says is "not available for use for customers at this time", while Sume Avatar 1.0 is an API you can call now: create an avatar from a photo, a prompt or a profile, then render a talking video of 4 to 60 seconds. They are not substitutes for each other, because one is a live conversation model and the other renders clips.

Every Tavus fact below is from Tavus: Griffin, and every Sume fact is from the docs pages in the sources, all read 2026-10-05.

Side by side

The table uses only what each vendor states. Where Tavus gives no figure, such as a price, the cell says so, and I do not fill it in.

Griffin-Lite and Sume Avatar 1.0, read 2026-10-05
QuestionTavus Griffin-LiteSume Avatar 1.0
Can a customer use it today?No, research preview for select trusted testersYes, through the API
What it isReal-time model from one reference image, full duplexReusable avatar plus rendered talking video
InputOne reference imagePhoto, prompt or profile, then a script or scenes
OutputLive video call720p video file, 4 to 60 seconds
Latency stated0.43 s average audio-to-video on H100sJob-based: submit, then poll or receive a webhook
Quality controlNot stated on the pagestandard, plus (default) or max
PriceNot stated on the pagePer second by tier; $0.95 per avatar created
DisclosureTavus is working on safe disclosure featuresYou add your own disclosure to the clip

Where Griffin-Lite is the better idea

Griffin-Lite is built to hold a conversation: Tavus says it can "interrupt, adjust, back-channel, or be interrupted without losing its place". That cannot be done by a rendered clip, and no amount of tier choice changes it. If your product needs the viewer to talk back, you are waiting for a live model.

Where a rendered clip is the better idea

A rendered clip wins when the script is known and the result must be repeatable. You can preview the first frame before a render, regenerate stills, approve, and then change quality tier, because preview stills are tier-independent. You can ask for 1:1, 3:4, 9:16, 4:3 or 16:9. You get job status, events, cancel before generation starts, and signed webhooks. None of this is a claim about quality against Tavus, which I have not tested and the page does not let you test.

  • Fixed message at scale, such as one script rendered for many customers.
  • A known cost per second before you start.
  • A job you can track and cancel, with webhooks.
  • A disclosure line that you control.

A pilot you can run now

For a pilot this week, create one avatar, render a 15 second clip at plus, and judge whether a scripted clip meets your need before you wait for a live model. The first calls are on Create new avatar and Generate avatar video. The cost by length is on the API pricing page.

A fair warning applies to any comparison made between a preview and a shipping product. A preview page shows the best case the vendor chose to publish, and a shipping API shows the contract you will be held to. Use the first to learn what to expect from live avatars in general, and the second to decide what to build now. Do not write a purchase decision on the strength of a research preview.

When to look again

Revisit this comparison when Tavus opens access. The sensible test then is the same script, same viewer task, one live and one rendered, measured on what you care about: completion, trust or conversion. The page numbers, such as the Turing test and latency, are Tavus's own and are not a substitute for your measurement.

Sources

Related posts

More in Comparisons

All Comparisons posts

Written by Sume