Make a vertical Short clip with curl and jq: submit, poll, download

You can make a vertical Short clip with nothing but curl and jq: POST to https://api.sume.com/v1/videos with aspect_ratio: "9:16", read polling_url from the reply, loop until status is completed, then download unsigned_urls[0]. The script below does that in under 20 lines and asks for 1080p, which is the ceiling YouTube states for uploaded Shorts.

Two YouTube pages set the targets. Understand three-minute YouTube Shorts (read 2026-10-10) says a Short can be up to three minutes and must have a square or vertical aspect ratio. Get started creating YouTube Shorts (read 2026-10-10) says you can upload Short videos with a maximum resolution of 1080p. One clip from this script is 3 to 10 seconds, so it is a building block for a longer Short, not the whole thing.

The script

It uses Gemini Omni Flash 1.1 because the Sume docs list it at 3 to 10 seconds, 360p/720p/1080p/4K, in 16:9 or 9:16, with native synced audio. Set SUME_API_KEY in your environment first; the key is never written into the script.

set -euo pipefail
API=https://api.sume.com
AUTH="Authorization: Bearer $SUME_API_KEY"
job=$(curl -sS -X POST "$API/v1/videos" -H "$AUTH" \
  -H "Content-Type: application/json" -H "Idempotency-Key: short-001" \
  -d '{"model":"gemini-omni-flash-1.1","aspect_ratio":"9:16",
       "resolution":"1080p","duration":8,
       "prompt":"A barista pours latte art, close vertical shot, cafe sounds"}')
url=$(jq -r .polling_url <<<"$job")
while :; do
  r=$(curl -sS "$url" -H "$AUTH")
  s=$(jq -r .status <<<"$r")
  [ "$s" = completed ] && break
  case "$s" in failed|cancelled) jq -r '.error // .status' <<<"$r"; exit 1;; esac
  sleep 15
done
curl -sS "$(jq -r '.unsigned_urls[0]' <<<"$r")" -H "$AUTH" -o short.mp4
jq '{status, cost: .usage.cost}' <<<"$r"

Why each field is there

The Video Generation page says the surface follows the OpenRouter video API with a few Sume differences. Those differences decide what you can leave out of the request.

What can go wrong

Check the loop's exit paths. A job can end as failed or cancelled, and the script prints the error field and stops rather than polling forever. The docs say a video usually takes from 30 seconds to several minutes, so a 15-second sleep is fine for one clip; for a batch, move to the callback_url webhook instead of many pollers.

If you change the model, change the numbers with it. The docs give seedance-2.5 4 to 30 seconds and wan-3.0 2 to 30 seconds, so an 8-second request fits all three, but a 2-second request fits only Wan. Read GET /v1/videos/models before you hard-code a duration.

Billing is visible in the last line. The poll response carries usage.cost, and the docs say Sume reserves the provider list price times 1.25 at submit; the cost field is the billable amount.

Check the file before you upload it

Do not trust the request alone; read the result. Sume's Video inspect page describes a probe that reads one media.sume.com clip your workspace already owns and returns probe facts plus sample stills, without re-encoding the source. Import the file first with POST /v1/media-imports, then send its URL as video_url with an Idempotency-Key.

Use the probe to confirm what you asked for: a vertical frame, a duration close to 8 seconds, and an audio stream. Use the stills to look at the first frame, because the opening seconds are what a Shorts viewer sees first. If the probe shows something other than what you requested, fix the request and resubmit under a new idempotency key; reusing the old key replays the original job.

From one clip to a Short

YouTube allows up to three minutes, and this script yields up to 10 seconds, so a full-length Short is a stitching job. Sume's Timeline 1.0 page describes joining ordered video slots with one audio spine into a single MP4. Keep the render at 1080 by 1920 or below for YouTube. Plan the join before you generate: eight 8-second clips make 64 seconds, and 22 make 176 seconds, just under the three-minute ceiling. The 1080p ceiling on the YouTube page is why the script does not ask for 4K, even though the model lists it.