Longest single AI video clip by API in 2026: 30 seconds, two Sume ids

The longest single AI video clip you can request from Sume is 30 seconds, and two ids reach it: seedance-2.5 (4-30 s) and wan-3.0 (2-30 s). Every other video id tops out at 15 seconds, except gemini-omni-flash-1.1, which stops at 10. That matches the vendor side: Magic Hour's tracker, read on 2026-10-06, lists Seedance 2.5 at up to 30 s and both Kling VIDEO 3.0 and MiniMax H3 at 15 s.

The numbers below come from the Sume video docs and the Video Router docs, and from the tracker. Prices are the provider list times 1.25.

Length limits side by side

The maximum clip, the cost of that clip at the cheapest listed resolution, and at 720p where it exists:

Why 30 seconds is not always the right ask

A 30-second job is one prompt and one take. If the clip fails, you pay a reserve and get a refund, but you wait for the full length again on the retry. Two 15-second jobs cost the same per second on Wan and Seedance, finish in parallel, and let you pick the better half. Sume's catalog does not accept a seed, so a retry is a new take, not a repeat of the old one.

Wan 3.0 is the cheaper long clip by a wide margin: 30 seconds at 720p is $3.75 against $17.34 for Seedance 2.5. The trade is what the model does with references and motion, which only a test on your own prompt answers.

What happens when you ask for 31 seconds

The catalog is enforced. A duration outside a model's supported_durations returns 400 unsupported_capability, and no provider call is made. The same applies to a 16 s request on kling-3. Read supported_durations from GET /v1/videos/models and clamp in your own code.

import os, requests

r = requests.get(
    "https://api.sume.com/v1/videos/models",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
r.raise_for_status()
for m in r.json()["data"]:
    d = m["supported_durations"]
    print(f"{m['id']:<26} {min(d):>3}-{max(d):<3} s")

Going past 30 seconds

Chain jobs. Generate each shot as its own job, then join them on a timeline. For the cost side of the longest single job, see what a 30-second 1080p Seedance job reserves, and for the 2.0 comparison see Seedance 2.5 or 2.0.

Choosing between the two 30-second ids

Both ids take text, a first and last frame, and image, video and audio references, so the decision is mostly price, ratios and the look you want. Seedance 2.5 takes 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16 plus auto and adaptive, and goes up to 1080p. Wan 3.0 takes 16:9, 4:3, 1:1, 3:4 and 9:16 plus auto and adaptive; it has no 21:9. If you need an ultrawide trailer, that alone decides it.

On price, Wan 3.0 bills the same rate per second at any duration, so a 30-second 1080p clip is $7.50. Seedance 2.5 bills on video tokens, which grow with frame size and length: the same 30 seconds at 1080p is $42.65. The gap is large enough that a draft on Wan, then a final on Seedance only for the shots that need it, is a reasonable way to spend a budget. Neither id accepts a seed, so a draft does not reproduce on the final; it is a preview of the idea, not of the frames.

A quick rule for picking the clip length

Pick the length from the story, not from the limit. A product reveal is 4 to 6 seconds, a talking explainer shot is 8 to 12, and a scene that needs a full beat of setup and payoff is where 20 to 30 seconds helps. Longer single takes also get more chances to drift: faces, hands and text are most likely to wobble late in a long clip. If the first 10 seconds of a test are right and the last 10 are not, cut the request to 15 seconds and extend the story with a second job that starts from the last frame of the first.