90-second AI video API: three 30 s Wan 3.0 jobs and what they cost
A 90-second AI video on Sume is three 30-second jobs. Wan 3.0 totals $5.64 at 480p, $11.25 at 720p and $22.50 at 1080p. A shot-list plan with a Python loop.

A 90-second AI video on Sume is three 30-second jobs, the largest chunk any model accepts. On Wan 3.0 the three jobs cost $5.64 at 480p, $11.25 at 720p and $22.50 at 1080p.
Each job is a separate POST /v1/videos request and a separate file, so the piece becomes a short shot list of three scenes joined in an editor.
What does a 90-second piece cost?
Wan 3.0 bills $0.0625, $0.125 and $0.25 per second at 480p, 720p and 1080p after Sume's 1.25 multiplier. Jobs are rounded up to the cent individually.
| Model | Tier | Per 30 s job | Three jobs |
|---|---|---|---|
| wan-3.0 | 480p | $1.88 | $5.64 |
| wan-3.0 | 720p | $3.75 | $11.25 |
| wan-3.0 | 1080p | $7.50 | $22.50 |
| seedance-2.5 | 720p | $17.34 | $52.02 |
How should you plan three scenes?
Write the 90 seconds as three scenes with one clear change between each, and give each job one scene. A scene boundary hides the seam, which a mid-action cut cannot do. Keep the subject the same with shared input_references, and keep the aspect ratio and resolution fixed.
- Scene 1 sets place and character.
- Scene 2 starts from scene 1's last frame via
frame_images, then moves the action on. - Scene 3 closes the piece with a still beat that is easy to cut to a title.
Does audio match across the three files?
Not automatically. Each job generates its own sound, so ambient audio may jump at the join. If the piece needs a single track, mute the clips and lay one soundtrack over the joined file. Sume's video router lets you pass generate_audio per request on models that support it.
What does the loop look like?
A small script submits the three scenes with idempotency keys.
import os, requests
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
scenes = [
"A fisherman prepares his boat at first light",
"The boat leaves the harbour into mist",
"The boat returns at sunset, gulls overhead",
]
for i, scene in enumerate(scenes, 1):
r = requests.post(
"https://api.sume.com/v1/videos",
headers={**H, "Idempotency-Key": f"doc-90-{i}"},
json={"model": "wan-3.0", "prompt": scene, "duration": 30,
"resolution": "720p", "aspect_ratio": "16:9"},
)
r.raise_for_status()
print(i, r.json()["id"])Where do you save money?
Draft each scene at 480p for 5 seconds ($0.32 each, $0.96 for three) before spending $11.25 at 720p. A reroll of one 30-second scene at 720p costs $3.75, so replace single scenes rather than the whole piece.
Sources
Related posts
More in Use cases
- AB 853 from Jan 2027: keep originals, not just uploads
California AB 853 adds platform provenance duties from Jan 1, 2027. Store the original Sume outputs you generate so you can answer provenance questions later.
- Agency approval loop for AI video: 480p proof, then the final on Sume
Show clients a cheap 480p Seedance 2.5 proof, get sign-off, then pay for the 1080p final. Cost table for a 30 s clip: $8.06 proof, $42.65 final, plus rounds.
- AI action figure box from a selfie: a gpt-image-2.5 reference edit
Make the AI action-figure-in-a-box image from one selfie: send it as an input reference to gpt-image-2.5, then letter the name on the box in code.
- AI animated short film pipeline: shots, Timeline, and a score
Higgsfield published a guide to AI animated shorts. Here is the job as a Sume API pipeline: shots from /v1/videos, a Timeline cut and a Music Router score.
Written by Sume