21:9 AI video with sound for $3: MiniMax H3 Max 1080p vs Seedance 2.5

The cheapest 21:9 AI video with sound on Sume is MiniMax H3 Max at 1080p: a 15 second clip is $3.00, against $21.43 for Seedance 2.5 at 1080p and $8.71 at 720p. minimax-h3 has native audio always on, and minimax-h3-max returns native stereo audio, so both answer a 21:9 request without a separate sound step.

Kling 4.0 adds 21:9 and two-channel stereo audio, per the fal explainer (read 2026-10-05). The full model is due in October. The comparison below uses what the Sume catalog lists today.

What each model lists

The video docs describe minimax-h3 as 5 to 15 seconds at native 480p or 768p, with 2K and 4K as billed upscales. minimax-h3-max is the faster 768p variant with a 1080p tier that is a latent refinement of the native 768p render, also 5 to 15 seconds. Both report image, video and audio input references. The catalog marks 21:9 as allowed for both.

Seedance 2.5 accepts 4 to 30 seconds at 480p, 720p and 1080p, in 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16, plus auto and adaptive. Its audio is optional. The difference in price is large because Seedance is billed by video tokens and MiniMax H3 by the second.

21:9 clips at 15 seconds

Choosing between them

Use H3 Max when the clip is 15 seconds or shorter and you want a 1080p wide frame with sound for a few dollars. Use Seedance 2.5 when you need more than 15 seconds, or a ratio such as 4:3 that the MiniMax entries do not list in the data read for this post. Check supported_aspect_ratios and supported_durations on GET /v1/videos/models before you submit.

Scale the figures to a project. Ten 15 second 21:9 clips at 1080p on H3 Max are $30.00. The same ten on Seedance 2.5 at 1080p are $214.30, and at 720p they are $87.10. If the clips are only for a social cut, 768p on H3 Max at $1.50 each is $15.00 for ten, which is under the price of a single 15 second Seedance clip at 1080p.

The MiniMax entries take references too: up to 9 images, 3 videos and 3 audio clips, and at most 12 references in total, per the Sume model notes. Seedance 2.5 lists image, video and audio references as well. Use the references to keep a product or a character consistent over several wide shots. Native audio is always on for H3, so describe the sound you want in the prompt, for example ambient street noise or quiet music, and expect to get a track whether you ask for one or not.

Remember that 1080p on H3 Max is a latent refinement rather than a native render, as the docs say, so judge a draft at 768p ($1.50 for 15 seconds) before you pay $3.00 for the 1080p version.

A 21:9 H3 Max request

This builds a 21:9 request for H3 Max and prints the expected price. It sends only when SUME_API_KEY is set.

import json, os, uuid

PER_SECOND = {"768p": 0.10, "1080p": 0.20}
seconds, resolution = 15, "1080p"
payload = {
    "model": "minimax-h3-max",
    "prompt": "Wide city canal at dusk, boats pass, ambient street sound",
    "duration": seconds,
    "resolution": resolution,
    "aspect_ratio": "21:9",
}
print(json.dumps(payload, indent=2))
print("expected price", round(PER_SECOND[resolution] * seconds, 2))

key = os.environ.get("SUME_API_KEY")
if key:
    import requests
    r = requests.post("https://api.sume.com/v1/videos",
        headers={"Authorization": f"Bearer {key}",
                 "Idempotency-Key": str(uuid.uuid4())},
        json=payload, timeout=60)
    print(r.status_code)

After you submit

The job is asynchronous: poll the URL in the response until it is completed, then download from the content URL, as in the video generation docs. The final amount is usage.cost, and the jobs docs explain the status and result routes. If a clip is rejected or fails, the error is on the job, so read it before you retry, and keep the same key only for an exact retry of the same body. A new prompt needs a new key.

A last practical point is the reserve. Sume holds the estimated amount at submit, so a 15 second 1080p H3 Max clip holds $3.00 until it finishes, and ten parallel clips hold $30.00. Keep the balance above that before you start the batch.