Try Wan 3.0 for 25 cents: a 2-second 720p clip on Sume

The smallest Wan 3.0 job you can run on Sume is 2 seconds, and at 720p it bills $0.25; at 480p it bills $0.13. Wan 3.0 (wan-3.0) takes 2 to 30 seconds, so it covers both very short tests and clips as long as Seedance 2.5's 30 seconds. Limits and rates come from the Video Router docs, read on 2026-10-06, and the request shape is in the video generation docs.

Pin the resolution. The provider's default for Wan 3.0 is 1080p, and the 1080p rate is double the 720p rate, so a request that omits resolution can cost twice what you expected.

The request

This submits a 2-second 720p text-to-video job. It needs a key and a funded balance, and it returns 202 with a polling URL.

import os, requests

r = requests.post(
    "https://api.sume.com/v1/videos",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "wan-3.0",
        "prompt": "A hand turns a vinyl record, dust in a shaft of light",
        "duration": 2,
        "resolution": "720p",
        "aspect_ratio": "9:16",
    },
    timeout=60,
)
print(r.status_code, r.json())

Price at every resolution and length

Wan 3.0 bills a flat rate per second, so the total is the rate times the seconds. The list rates are $0.05, $0.10 and $0.20 a second, and Sume bills them times 1.25.

What Wan 3.0 accepts

The catalog lists text-to-video, image-to-video with a first and a last frame, and reference-to-video. References can be up to 10 images, up to 5 videos totaling at most 15 seconds and at least 16 frames per second, and up to 5 audio clips totaling at most 15 seconds. Aspect ratios are auto, adaptive, 16:9, 4:3, 1:1, 3:4 and 9:16. There is no 21:9. Audio is supported. A seed is not.

A 2-second clip is useful for more than a test. It is the length of a loop for an email header, a button hover or a transition, and at 480p it is $0.13. For anything longer, draft at 480p, check the motion, and rerun the winners at 720p.

Where a first test goes wrong

The common mistake is an omitted resolution, which falls back to 1080p. The next is a duration of 1 second, which is below the model's 2-second floor and returns 400 unsupported_capability. A third is a 21:9 ratio, which Wan does not list. Each of these is rejected before any provider call, so none of them costs money, but each wastes a minute.

If you are unsure about a value, read the catalog first: validate against `/v1/videos/models`. And for the default-resolution trap in detail, see Wan 3.0 defaults to 1080p.

A cheap way to test a prompt set

Because the floor is 2 seconds and the rate is flat, Wan 3.0 is a cheap way to test a list of prompts. Twenty prompts at 2 seconds and 480p cost $2.60 in total; the same twenty at 720p cost $5.00. You will not judge a full shot from 2 seconds, but you can judge whether the model understands the subject, the lighting and the camera move you asked for, which is where most prompts fail. Keep the winners, lengthen them to 8 or 10 seconds, and rerun. That two-stage routine spends cents on the rejects and dollars only on the keepers.

Image-to-video for the same price

The first-frame mode costs the same per second as text-to-video. Send a still in frame_images with frame_type: first_frame, keep duration: 2, and you get a short animation of your own image for the price in the table. Add a last_frame image and Wan will move between the two, which is a quick way to build a before-and-after transition.