Wan 3.0 and Wan 3.0 Prime: what they make, what they cost, how to use them
By GetStepSnap · Checked against Alibaba Cloud's own pages and fal.ai on October 22, 2026
Wan 3.0 is Alibaba's newest video model. It makes 2 to 30 seconds of video in one go, at 480p, 720p or 1080p, and the sound comes out with the picture: voices, sound effects and music. You can start from a sentence, from a first frame, or from up to 20 references, and it can edit a clip you already have. It opened to everyone on August 24, 2026.
Wan 3.0 Prime is the same model, run faster. Alibaba calls it the high-speed version and says it can do everything Wan 3.0 does; it costs about 40% more a second. Alibaba hasn't said how much faster it is, so unless you're waiting on it, plain Wan 3.0 is the better buy.
Both are paid by the second through an API: a 5-second 720p clip costs $0.50 on Wan 3.0 and $0.70 on Prime. And unlike Wan 2.2, Wan 3.0 isn't open source, so there are no weights to run on your own graphics card.
Wan 3.0 vs Wan 3.0 Prime
| Wan 3.0 | Wan 3.0 Prime | |
|---|---|---|
| What it is | Alibaba's all-in-one video model | The same model, run faster |
| Out | August 6, 2026 on Alibaba Cloud; open to everyone August 24 | August 20, 2026 |
| Per second at 480p | $0.05 | $0.068 |
| Per second at 720p | $0.10 | $0.14 |
| Per second at 1080p | $0.20 | $0.28 |
| Length | 2 to 30 seconds, or let it pick | The same |
| Sound | Voices, sound effects and music, on by default | The same |
| Speed | No figure published | Alibaba says much faster; no figure published |
Pick Prime when time costs you more than the extra 40%: a client waiting on a draft, or a batch that has to be done tonight. For everything else, the plain model makes the same video for less. Both share the same length, sizes, modes and request limits.
If speed is the whole point, look at FLUX 3 Max too: fal says it makes a 10-second clip in about 11 seconds, at $0.17 a second at 720p once its launch discount ends on November 7. FLUX 3 Max: price and speed.
What Wan 3.0 can do
- From a sentence. Describe the scene and get a video. No picture needed.
- From a first frame, or a first and last frame. Give it the opening picture, or the opening and closing ones, and it fills in the motion between them.
- From references. Up to 20 at once: images, videos, audio clips, documents and web pages, with at most 10 images, 5 videos and 5 audio clips. In the prompt you call them Image 1, Video 1, Audio 1 and so on.
- Edit a video with words. Hand it a clip and say what to change, instead of making the whole thing again.
- Make a video longer. Extend a clip forward, backward or both ways, as long as the old part and the new part add up to 30 seconds or less.
- Several shots in one clip. Write it as a storyboard: numbered shots a few seconds long, each with its timestamps.
Every video is an MP4 at 30 frames a second, in 16:9, 4:3, 1:1, 3:4 or 9:16, or a shape that follows your references. Leave the length to the model and it picks one that suits the prompt.
What changed from Wan 2.7
Length, mostly. Wan 2.7 made 2 to 15 seconds; Wan 3.0 goes to 30. Fifteen seconds is a shot. Thirty is a small scene, with time for a line of dialogue and somebody reacting to it.
The references got wider too. Besides images, video and audio, Wan 3.0 reads documents (Word, Excel, PowerPoint, PDF, plain text, Markdown, Keynote, Pages and Numbers) and web pages, so you can hand it a product sheet or a slide deck and ask for the video. Editing a clip with words, new in 2.7, carries over.
Alibaba is open about what still needs work: the texture of the sound, and words that appear on screen. Keep on-screen text short, or add it afterwards.
What a clip costs
Billed per second of finished video, so a clip twice as long costs twice as much. These are list prices on fal.ai, the same as Alibaba Cloud's international ones; Alibaba has run time-limited discounts on its own channels.
| Size | 5 seconds | 10 seconds | 30 seconds |
|---|---|---|---|
| 480p | $0.25Prime $0.34 | $0.50Prime $0.68 | $1.50Prime $2.04 |
| 720p | $0.50Prime $0.70 | $1.00Prime $1.40 | $3.00Prime $4.20 |
| 1080p | $1.00Prime $1.40 | $2.00Prime $2.80 | $6.00Prime $8.40 |
How to write a Wan 3.0 prompt
- Go in Alibaba's order: what the video is about, then the shots, the dialogue, the sound, the style, and last a short list of what you don't want. Skip the parts you don't care about.
- For more than one shot, number them and give each its timestamps, like Shot 1 [0-5s], Shot 2 [5-10s], each starting where the last one ends. Alibaba's guides keep a shot to a few seconds (2 to 5 in the prompt guide, 4 to 6 in the model guide). For one continuous take, start with "Generate single shot."
- Write dialogue as who says what: The grandmother says: "Patience." Add "Lip sync" so the mouth matches. If you want nobody talking, write "No dialogue", or the model may add some.
- Name the sounds you want in their own sentence (rain on the glass, traffic). For silence under the scene, write "No background music".
- Move the camera in plain words: push in, pull out, orbit, handheld follow. Or hold it still: "Fixed shot, camera static."
- Keep words on screen to a few, or add them later. It still gets text wrong sometimes, so "No subtitles" is worth adding.
- Working from references? Call them Image 1, Video 1, Audio 1 in the prompt so it knows which is which.
Five prompts to start from
Our own, written to Alibaba's prompt guide. We haven't run them on this site yet.
Three shots with timestamps
Shows: A short scene cut into numbered shots, 15 seconds
A rainy night in a tiny ramen shop. Shot 1 [0-5s]: wide shot, steam rolls off the counter as the cook, in a white headband, ladles broth into a bowl; rain streaks the window behind him. Shot 2 [5-10s]: close-up, chopsticks lift glossy noodles, warm light on the steam. Shot 3 [10-15s]: medium shot, a tired office worker takes the first bite, closes her eyes and smiles. Sound: rain on the glass, a simmering pot, soft traffic outside. No dialogue. No subtitles.
A line of dialogue
Shows: A voice that says the line, with the lips in sync
Generate single shot, handheld follow. A grandmother and her grandson bake bread in a sunny farmhouse kitchen. She dusts the flour off her hands. The grandmother says: "Patience. Good bread can't be rushed." Lip sync. He sneaks a pinch of dough into his mouth and grins. Warm morning light. Sound: a wooden spoon against a bowl, the oven ticking. No background music. No subtitles.
Product shot for a phone screen
Shows: A 9:16 ad with a slow orbit
Vertical 9:16 product video. Generate single shot. A matte black water bottle stands on wet rocks beside a mountain stream; drops of water slide down its side as the camera slowly orbits it, then a hand reaches in and picks it up. Crisp morning light, 85mm, shallow depth of field. Sound: rushing water and birds. No dialogue. No text on screen.
Drone shot at sunrise
Shows: Wide scenery, a rising camera and music
Generate single shot. A drone sweeps low over a misty pine forest at sunrise, then rises to reveal a still mountain lake as golden light breaks through the clouds. Calm and cinematic. Sound: wind and distant birds, with soft piano music. No dialogue.
Anime summer
Shows: An animated style and fast action
Anime style. Generate single shot, handheld follow. A girl on a red bicycle races down a steep seaside hill in summer; her straw hat flies off and she twists around to catch it, laughing, as the ocean glitters below. Bright colors, smooth animation. Sound: cicadas, the rattle of the bicycle, the wind. No dialogue.
Is Wan 3.0 open source?
No, and that's the big change from the last generation. Wan 2.1 and 2.2 put their weights on Hugging Face, which is why so many people run them in ComfyUI on their own graphics card. Wan 3.0's weights aren't published: on October 22, 2026, Alibaba's Wan-AI page on Hugging Face still stopped at the Wan 2.2 family. You use it through Alibaba's services, Alibaba Cloud Model Studio and Qwen Cloud, or through an API platform.
A word on sites named after Wan
Search for Wan 3 and you'll find plenty of sites with Wan in their name. Most of them aren't Alibaba's: like us, they run the model through an API and set their own price. Check what you'd pay per second against the table above before you buy credits anywhere.
What we're doing
GetStepSnap runs new AI models without a subscription: you buy credits once and use them on any tool. Today we make images. We plan to add Wan 3.0 here and will put the price per clip right on this page. Until then, a strong first frame is half of a good video, and you can make one with our photo generator.
Make a first frameQuestions
What is Wan 3.0?
Alibaba's newest video model. It makes 2 to 30 seconds of video in one go, at 480p, 720p or 1080p, with sound, from a sentence, a first frame, a first and last frame, or up to 20 references. It can also edit a video you give it and make a clip longer.
What is Wan 3.0 Prime?
The faster version of Wan 3.0. Alibaba calls it the high-speed version and says it can do the same things; it isn't a newer or better model. It costs about 40% more a second at 720p and 1080p ($0.14 against $0.10 at 720p), and Alibaba hasn't said how much faster it is.
How much does Wan 3.0 cost?
It's billed by the second of video: $0.05 at 480p, $0.10 at 720p and $0.20 at 1080p on Alibaba Cloud's international price list and on fal.ai. A 5-second 720p clip is $0.50 and a 30-second 1080p clip is $6.00. Prime is $0.068, $0.14 and $0.28 a second.
How long can Wan 3.0 videos be?
2 to 30 seconds in one generation, twice Wan 2.7's 15. Or leave the length to the model and it picks one that fits the prompt. A longer video comes from extending a clip, with the old and new parts adding up to 30 seconds at most.
Does Wan 3.0 make sound?
Yes, by default: voices speaking your lines, sound effects and background music, made with the picture. You can turn it off. Alibaba says the sound quality is still improving.
Is Wan 3.0 open source? Can I run it in ComfyUI?
No. Wan 2.1 and 2.2 put their weights on Hugging Face, so people run them in ComfyUI on their own graphics card. Wan 3.0's weights aren't published: on October 22, 2026, Alibaba's Wan-AI page on Hugging Face still stopped at the Wan 2.2 family. You use it through Alibaba's services or an API.
How do I use Wan 3.0?
Through an API. Alibaba sells it on Alibaba Cloud Model Studio and Qwen Cloud as wan3.0-video and wan3.0-video-prime, and you can call it there without applying. API platforms carry it too: on fal.ai it's alibaba/wan-3.0 and alibaba/wan-3.0-prime, each with text-to-video, image-to-video and reference-to-video.
When did Wan 3.0 come out?
It appeared on Alibaba Cloud on August 6, 2026, and Alibaba announced it for everyone on August 24. Wan 3.0 Prime was added on August 20.
Sources, checked October 22, 2026: Video generation model Wan3.0 officially launched (Alibaba Cloud, August 24, 2026), Wan3.0 video generation guide (Model Studio), wan3.0 video generation prompt guide (Model Studio), Model lifecycle and updates (Model Studio), Wan3.0 vs. Wan3.0-Prime (Alibaba Cloud blog, October 7, 2026), Wan2.7 vs. Wan3.0 (Alibaba Cloud blog, August 27, 2026), Wan3.0 at general availability (Alibaba Cloud blog, August 27, 2026), Wan 3.0 on fal.ai, Wan 3.0 Prime on fal.ai, Wan-AI on Hugging Face.
