Wan 3.0 Video API

Pricing (per second)

Billed per second. Without a reference video you pay for output seconds. With a reference video (video edit, extend, character swap) you pay for INPUT video duration + OUTPUT duration — matching the upstream, which also caps these jobs at input + output ≤ 30 seconds. Reference images, audio, documents and web pages cost nothing extra.

• Cost = rate × billable seconds. **Without a reference video**, billable seconds = OUTPUT seconds, and reference images, audio, documents and web pages cost nothing extra. **With a reference video** (video edit, extend, character swap), billable seconds = INPUT video duration + OUTPUT duration. Toggling the audio track does not change the price.
• With duration: -1 (smart duration) we hold the 30-second worst case, then charge the real output length once upstream reports it and refund the difference.
• Failed tasks are not billed.

Create a task

POST /api/v1/video/generations — async: create, then poll the same endpoint for the result.

# Step 1: Create task (text-to-video)
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3.0-video",
    "mode": "prime",
    "prompt": "A paper airplane gliding over a calm lake at sunrise, camera drifting alongside",
    "resolution": "720P",
    "ratio": "16:9",
    "duration": 5
  }'

# Omni-reference: address materials positionally in the prompt
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3.0-video",
    "prompt": "The person in video 1 picks up the object from figure 2 and places it on the table",
    "reference_image_urls": ["https://example.com/object.png"],
    "reference_video_urls": ["https://example.com/person.mp4"],
    "resolution": "720P",
    "duration": 8
  }'

# Document to video: feed a deck straight in
curl -X POST https://api.apimodels.app/v1/video/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3.0-video",
    "prompt": "A polished product launch film based on this deck",
    "file_url": "https://example.com/launch-deck.pptx",
    "resolution": "1080P",
    "duration": 10
  }'

# Step 2: Poll status
curl "https://api.apimodels.app/v1/video/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer YOUR_API_KEY"

Four ways to drive it

⚠️ Frame mode (first_frame_url / last_frame_url) and reference mode (reference_* / file_url / link_url) are mutually exclusive — mixing them makes upstream fail the task, so we reject it locally first and name both colliding sides. file_url and link_url are likewise one or the other.

In omni-reference mode the prompt addresses materials positionally — "figure 1", "figure 2", "video 1", "audio 1", with images, videos and audio numbered independently. For example: "the person in video 1 holds the object from figure 3 and plays guitar on the chair from figure 4." That is how several characters, props and environments stay consistent inside one shot.

Request parameters

Getting the result

The create call returns data.taskId right away and the video renders in the background. GET the same endpoint with that task_id and read data.state: pending means still running, completed puts the MP4 URL in data.resultUrls[0], and failed puts the reason in data.failMsg. To skip polling, pass callback_url and we POST you when the task finishes.

• Measured turnaround: 480P 5s came back in about 356 seconds on the standard tier (Alibaba quotes 1-5 minutes); mode:"prime" returned the identical prompt and parameters in 58 seconds. Poll every 10-15 seconds.
• Output is H.264 MP4 at 30fps; a measured 480P 16:9 clip came back as 832×480.
• Result videos are kept for 30 days and then deleted — download or re-host anything you want to keep.
• You are charged on success; failed tasks are not billed.

Things worth knowing

• The 30 seconds is one continuous native shot, not stitched segments — that is the biggest single difference from models capped at 5-10 seconds.
• The two tiers have matching capabilities per Alibaba: same omni-reference inputs, same 30s / 30fps ceiling, same audio. The difference is speed and price, not quality.
• Reference material can be sent as raw base64 / data URLs — we stage it to an accessible URL before generation, identical to sending a URL. Upstream oss:// temporary URIs are passed through verbatim.
• Image limits: JPEG / JPG / PNG (no alpha channel) / BMP / WEBP, 240-8000px per side, aspect ratio within 8:1, ≤20MB each.
• Why not call Alibaba directly: Wan 3.0 is served only from the China (Beijing) and Singapore regions — we probed US/Virginia and it does not carry the model at all — and the China station requires mainland Chinese real-name verification to register. Here you need neither an Alibaba Cloud account nor that verification.
• Auth is the same as every other model: send Authorization: Bearer YOUR_API_KEY. One key covers every video, image, audio and LLM model on the platform.

Related