Ten-shot 21:9 storyboard test on Seedance 2 Mini: $3.52 at 480p

To test a ten-shot storyboard cheaply on Sume, run each shot as a 4 second 480p clip on seedance-2-mini: it is $0.35 a shot and $3.52 for ten, against $5.63 on Seedance 2 Fast, $7.03 on Seedance 2 and $10.75 on Seedance 2.5. Mini keeps the 21:9 ratio and takes a first and last frame, so a ten-frame plan is ready to try today.

Kling 4.0 takes up to 10 keyframe images in one clip, per the fal explainer (read 2026-10-05). It is not on Sume yet. A storyboard test with ten shots shows which keyframes carry the idea, whatever model you use for the final.

The family

The Seedance 2.x family on Sume runs 4 to 15 seconds at 480p, 720p and 1080p, with text-to-video, image-to-video, an end frame, and image, video and audio references. The prices follow the video-token formula with a different rate for each model: Mini is the lowest, then Fast, then Seedance 2, then Seedance 2.5 at $0.0214 per 1,000 tokens before the 1.25 margin.

Because the four models share the same envelope, you can run the same ten requests on each and compare. Only the model field changes. Seedance 2.5 adds a 30 second ceiling and is the one to use for the final if you need more than 15 seconds.

The 4 second floor is also the cheapest length to test. A 4 second clip at 480p on Mini is 864 x 496 pixels, which is about 40,176 video tokens by the formula in the pricing tables, and the price scales in a straight line with seconds, so a 6 second clip is one and a half times the 4 second price.

Ten 4 second 480p shots, by model

Stills and the next step

Add the stills. Ten GPT Image 2.5 frames at 2560 x 1088 and medium quality are about $0.11, so a ten shot storyboard with stills is about $3.63 on Mini. Use each still as the first_frame of one shot, and the next still as its last_frame if you want the clip to end on it.

A pass at 480p tells you about composition, pacing and continuity. It does not tell you about fine detail, so judge those on the final model. Keep the shots that work, drop the rest, and rerun only the kept shots at 720p or 1080p.

A reasonable order is to write the ten prompts, make the ten stills, run the ten Mini shots, and look at them in sequence on a timeline. At that point you will know which shots need a different still, which need a longer take, and which are fine. You will have spent under four dollars to learn it, which is less than a single 15 second Seedance 2.5 clip at 720p.

Ten shots in one loop

This prints the ten request bodies for Mini and the total. It does not send them.

The loop uses the same formula as the pricing tables, so the Mini line matches the table above. Change W and H for another size, or SECONDS for another length, and the other three lines follow. The 21:9 sizes are different from the 16:9 ones, so check the dimension table in the Seedance 2.5 post before you rely on the 21:9 numbers.

import json

RATE_PER_1000_TOKENS = {
    "seedance-2-mini": 0.0070,
    "seedance-2-fast": 0.0112,
    "seedance-2": 0.0140,
    "seedance-2.5": 0.0214,
}
W, H, SECONDS = 864, 496, 4

def price(model):
    tokens = W * H * SECONDS * 24 / 1024
    return tokens / 1000 * RATE_PER_1000_TOKENS[model] * 1.25

for model in RATE_PER_1000_TOKENS:
    print(model, round(price(model), 2), round(10 * price(model), 2))

shots = [{"model": "seedance-2-mini", "prompt": f"Shot {i + 1}",
          "duration": 4, "resolution": "480p", "aspect_ratio": "21:9"}
         for i in range(10)]
print(json.dumps(shots[0]))

Run the ten shots

Send each shot to POST /v1/videos with its own Idempotency-Key, poll the jobs, and read usage.cost for the final amount. The video docs show the request and the polling flow, and the jobs docs show the status routes.

Run the shots with a short pause between submits if your plan has a low capacity, and keep each Idempotency-Key stable, for example board-shot-03, so that a retry returns the same job and not a second charge.