A weekly host video while Tavus Griffin is preview-only: Sume, priced

If you need a presenter on screen every week and Tavus Griffin is not available to you, Sume's Avatar 1.0 talking video is the route to price: a 20-second host update costs $3.68 at the standard tier, $4.90 at plus and $11.00 at max (read 2026-10-03). Twelve weekly updates at the default plus tier come to $58.80.

What Tavus publishes today is narrow. Its site calls Griffin "the first human interaction model", says "48% thought it was a real person" in live video testing, and marks the model as a preview; the page gives no availability date, no price, no latency or resolution figures, and I found no mention of a Griffin-Lite variant on it (read 2026-10-03). A tool you cannot yet book is not a plan for next Monday's update, so this article prices what you can render now.

Live conversation versus a rendered clip

Griffin is described as a model that sees, hears, speaks and moves during a live video conversation. A weekly update is a different job: one script, one take, a file you can review before it goes out. Sume does not do the live part. Its avatar video is an asynchronous render of a script spoken by a ready avatar, delivered as an MP4 after the job completes.

That difference is usually an advantage for scheduled content. A rendered clip can be read, approved and re-rendered. A live session cannot be unsaid, and a scheduled update is exactly where a reviewer should get one look before the audience does. Budget that review into the week, not the render.

The price per update

Avatar Video bills per second by quality tier. The rates below are Sume billable amounts with no product image attached; adding a product_image adds a small per-second increment, shown in the last column.

Submit one

The canonical route is POST /v1/avatar-1.0/talking-video. Reference a ready avatar with avatar_handle, send exactly one of script or video_inputs, and pick quality. Scripts are accepted when Sume estimates the video at 4 to 60 seconds, and resolution is currently 720p. The aspect_ratio options are 1:1, 3:4, 9:16, 4:3 and 16:9, with 9:16 as the default.

import os, requests

API = "https://api.sume.com"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}",
     "Idempotency-Key": "weekly-update-2026-w41"}

script = ("Good morning. Three things shipped this week: faster exports, "
          "a new template, and fewer failed renders. Details are in the "
          "changelog. See you next Monday.")
body = {"avatar_handle": os.environ["AVATAR_HANDLE"], "script": script,
        "quality": "standard", "aspect_ratio": "16:9"}
r = requests.post(f"{API}/v1/avatar-1.0/talking-video", headers=H, json=body)
print(r.status_code)
print(r.json())
# Poll GET /v1/jobs/<request_id>/status, then /result, as in the jobs docs.

What to do this week

Render one update at standard and one at plus from the same script and compare them side by side before you commit to a tier for the quarter. Standard is the fastest path and plus is the balanced default, so for an internal update standard may be enough, and for a customer-facing one plus may be worth $1.22 more per 20 seconds.

Keep scripts short. Spoken delivery of about 20 seconds is a few sentences, and the 4 to 60 second window means a long update should be split into two jobs rather than stretched. A fixed intro and outro can be rendered once and reused instead of paid for every week.

Add captions in the request if the update will be watched muted. The docs say captions are burned into the final MP4 from the spoken script after generation, and preview stills are never captioned.

Finally, use an Idempotency-Key per week, as in the sketch, so a retry returns the same job. If you run several presenters at once, remember that concurrency is plan-based: Pro processes 4 jobs at a time and accepts up to 24 queued or processing, and extra valid jobs wait as queued instead of failing.

What Sume does and does not do

Sume renders a scripted talking video from a ready avatar and returns an MP4. It does not hold a live, two-way video conversation, it does not claim Griffin-like behaviour, and it does not have Griffin or Griffin-Lite in its catalog. If real-time conversation is the requirement, Tavus is the vendor to ask, and its page currently lists no way to book Griffin.