Seedance 2.0 Fast via Dawnwood Models: $0.15/s at 720p, about 40% below fal.
In shortSeven video models (Seedance 2.0 Fast, 2.0 Mini, 2.0, 2.5 and the new 2.5 Flash, plus Wan 3.0 and Wan 3.0 Prime) over one HTTP API that takes the BytePlus ModelArk request body as-is. Seedance 2.5 also renders native 1080p and 4K. One key, one wallet: you pay per second of generated video, only for generations that succeed. Sold by Dawnwood Inc., a US company, with Stripe invoices.
USD per second of output video, by resolution. Billed on the duration you request. Seedance 2.5 renders native 1080p and 4K (4K is 2160 pixels on the short side); Wan renders native 1080p; the other Seedance models top out at 720p. Video-reference inputs are available in the console's Model Studio, not on the API yet. To make a clip without code, use Model Studio or the /video command in Discord, at the same prices.
| Model | 480p / s | 720p / s | 1080p / s | 4K / s | 5 s clip, 720p | fal, 720p / s | vs fal |
|---|---|---|---|---|---|---|---|
| Seedance 2.0 Fast | $0.07 | $0.15 | — | — | $0.75 | $0.242 | 38% lower |
| Seedance 2.0 Mini | $0.047 | $0.10 | — | — | $0.50 | $0.155 | 35% lower |
| Seedance 2.0 | $0.078 | $0.167 | — | — | $0.835 | $0.303 | 45% lower |
| Seedance 2.5 | $0.097 | $0.209 | $0.52 | $1.892 | $1.045 | $0.473 | 56% lower |
| Seedance 2.5 Flash | $0.078 | $0.167 | — | — | $0.835 | — | — |
| Wan 3.0 | $0.055 | $0.11 | $0.22 | — | $0.55 | — | — |
| Wan 3.0 Prime | $0.075 | $0.154 | $0.308 | — | $0.77 | — | — |
New: Seedance 2.5 in 1080p and 4K. Seedance 2.5 renders natively at 1080p for $0.52/s and at 4K for $1.892/s: set "resolution" to "1080p" or "4k". A 5-second 4K clip costs $9.46, so a common pattern is to iterate at 720p and render the keeper in 4K.
New: Seedance 2.5 Flash. The faster, lower-priced Seedance 2.5 tier: the same 30-second clips and request shape as 2.5, about 1.8× faster and priced like Seedance 2.0. Up to 10 reference images and 5 reference audios, 480p/720p. Seedance 2.5 Flash API.
Below list price. Seedance 2.5 is $0.209/s at 720p on Dawnwood Models: 10% below BytePlus ModelArk's published rate (about $0.232/s) and 56% below fal ($0.473/s).
Real people as references. In the console's Model Studio, Seedance 2.5 accepts photorealistic photos of real people as reference images: yourself, or cast who have agreed to it. Tick KYC on the reference and it goes into the upstream portrait library for that one generation, then is deleted when the task finishes. On a standard Seedance endpoint, realistic faces in reference images are usually rejected by input moderation, and fal does not offer a portrait-library route. Not on the API yet.
fal prices are its public 720p rates as of 2026-11-02. Wan is priced at Alibaba Cloud's list rate plus 10%, and you don't need an Alibaba Cloud account. Credits: 1 credit = $0.10, so a 5-second 720p Seedance 2.0 Fast clip is 7.5 credits. Full comparison across Seedance providers: Seedance API pricing compared.
Mint a key in the console under Developers (it is a Dawnwood personal access token, dw_pat_v1_…, with the model_api scope). Create a task, then poll it.
# 1. create a task
curl -s https://models.dawnwood.ai/v1/video/generations \
-H "Authorization: Bearer $DAWNWOOD_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: shot-0001" \
-d '{
"model": "seedance-2.0-fast",
"content": [
{"type": "text", "text": "A woman in a red coat turns at the top of a rain-soaked stairwell, slow push-in"},
{"type": "image_url", "image_url": {"url": "https://example.com/lead.png"}, "role": "reference_image"}
],
"resolution": "720p", "ratio": "9:16", "duration": 5, "generate_audio": true
}'
# -> {"id": "cgt-dw-..."}
# 2. poll until status is succeeded / failed / cancelled
curl -s https://models.dawnwood.ai/v1/video/generations/cgt-dw-... \
-H "Authorization: Bearer $DAWNWOOD_API_KEY"
import os, time, requests
BASE = "https://models.dawnwood.ai/v1"
H = {"Authorization": f"Bearer {os.environ['DAWNWOOD_API_KEY']}"} # dw_pat_v1_...
task = requests.post(f"{BASE}/video/generations", headers=H, json={
"model": "seedance-2.0-fast",
"content": [{"type": "text", "text": "Two strangers share an umbrella at a night bus stop, handheld"}],
"resolution": "720p", "ratio": "9:16", "duration": 5, "generate_audio": True,
}).json()
while True:
t = requests.get(f"{BASE}/video/generations/{task['id']}", headers=H).json()
if t["status"] in ("succeeded", "failed", "cancelled"):
break
time.sleep(10)
print(t["content"]["video_url"] if t["status"] == "succeeded" else t["error"])
print(t["billing"]) # credits, usd, refunded
Wan 3.0 and Wan 3.0 Prime take the same content[] body as Seedance: a text prompt, image_url items as first_frame, last_frame or reference_image, and audio_url reference audio. Two differences: resolution can be 1080p, and duration runs from 2 to 30 seconds. In the prompt, refer to references the Wan way — Image 1, Image 2, Audio 1, numbered in the order they appear. Prime gives the same quality several times faster; pick it when someone is waiting on the result.
curl -s https://models.dawnwood.ai/v1/video/generations \
-H "Authorization: Bearer $DAWNWOOD_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "wan-3.0",
"content": [
{"type": "text", "text": "Image 1 walks out of the bakery into morning snow, handheld follow"},
{"type": "image_url", "image_url": {"url": "https://example.com/lead.png"}, "role": "reference_image"}
],
"resolution": "1080p", "ratio": "9:16", "duration": 5, "generate_audio": true
}'
ModelArk: change the base URL and the key. POST /v1/contents/generations/tasks and GET /v1/contents/generations/tasks/{id} are aliases of the endpoints above, the body is the same model / content[] / resolution / ratio / duration / generate_audio, and the official dreamina-seedance-… model ids are accepted.
# before: BytePlus ModelArk
BASE = "https://ark.ap-southeast.bytepluses.com/api/v3"
KEY = os.environ["ARK_API_KEY"]
# after: Dawnwood Models (same path, same body, same task object)
BASE = "https://models.dawnwood.ai/v1"
KEY = os.environ["DAWNWOOD_API_KEY"] # dw_pat_v1_..., model_api scope
requests.post(f"{BASE}/contents/generations/tasks", headers={"Authorization": f"Bearer {KEY}"}, json=body)
fal: the prompt becomes a text item in content[], each image URL becomes an image_url item with a role, and aspect_ratio becomes ratio. Side-by-side code: Migrating Seedance from fal or ModelArk.
429 with a Retry-After header. Wait that many seconds and resubmit.Idempotency-Key header: a retried request with the same key returns the same task and is never charged twice.GET /v1/video/generations/{id} every 10 seconds or so; finished video URLs stay valid for 24 hours."refunded": true.402 paid_credits_required.Dawnwood Inc. is a US company and the API is sold under our Terms. They forbid generating a real person's likeness without authorization and sexual content. With authorization, use the KYC option described above. The upstream model runs its own moderation: when it rejects a request we return that rejection as-is and refund the credits. We do not offer a way around it.
$0.15 per output second at 720p and $0.07 at 480p. A 5-second 720p clip costs $0.75. Only successful generations are charged.
Seedance 2.5 renders natively at 1080p for $0.52/s and at 4K for $1.892/s: set "resolution" to "1080p" or "4k". A 5-second 4K clip costs $9.46, so a common pattern is to iterate at 720p and render the keeper in 4K. Seedance 2.0 Fast, Mini, 2.0 and 2.5 Flash render at 480p or 720p.
Yes. Wan 3.0 is $0.11/s at 720p and $0.22/s at 1080p; Wan 3.0 Prime, the faster tier, is $0.154/s at 720p. Same endpoint, same content[] body, same key and wallet as Seedance, and no Alibaba Cloud account needed. In the prompt, refer to references as Image 1, Video 1, Audio 1.
No. Credits are reserved when a task is created and returned in full if the task fails or is rejected upstream. The task's billing field shows refunded: true.
Partly. Purchased credits are spent first. When they run out, up to 30 credits of your signup or gift credits are spent as a free trial, about four 5-second 720p Seedance 2.0 Fast clips. Credits included in a subscription can't be spent on the API. Once the trial is used, a request without enough purchased credits returns 402 paid_credits_required.
Yes. The request body and task object follow ModelArk's content-generation task format, and /v1/contents/generations/tasks is an alias of /v1/video/generations, so a ModelArk integration only changes its base URL and key. Official dreamina-seedance model ids are accepted too.
Each API key has a limit on tasks running at once. Past it, task creation returns 429 with a Retry-After header; wait that many seconds and resubmit with the same Idempotency-Key.
Dawnwood Inc., a US company. Credit purchases go through Stripe and come with Stripe receipts and invoices.
In the console's Model Studio, Seedance 2.5 accepts photorealistic photos of real people as reference images: yourself, or cast who have agreed to it. Tick KYC on the reference and it goes into the upstream portrait library for that one generation, then is deleted when the task finishes. On a standard Seedance endpoint, realistic faces in reference images are usually rejected by input moderation, and fal does not offer a portrait-library route. Not on the API yet.
Our Terms forbid generating real people's likeness without authorization and sexual content. Requests rejected by the upstream model's moderation are returned with that rejection as-is; we do not offer a way around it.
Dawnwood is a studio for short-drama series: season bible, episode scripts, consistent characters and sets, rendering with QC, delivery. The same token that calls Dawnwood Models can drive that pipeline over its own API or MCP, so your agent works on episodes instead of stitching prompts together.