Wan 3.0: Tested Against Wan 2.7, with Pricing and API Guide

Wan 3.0 is Alibaba's newest video model, released in August 2026. One model now covers text to video, image to video, and video to video, makes clips up to 30 seconds at up to 1080P, and can read a PDF, slide deck, or webpage and turn it into a video. It is API only: there are no open weights to download.

On Modellix it costs $0.045 to $0.18 per second, so a 6-second 720P clip is $0.54. We ran it against Wan 2.7 on the same prompt, then tried the two things 2.7 cannot do: a single 30-second shot and a video made from a PDF.

Short version: Wan 3.0 finished our test clip 2.5x faster than 2.7 at the same 720P price, and it followed the camera direction more closely. The 30-second shot kept the same object from start to end. The PDF test pulled a real fact from the document into the narration. Watch the defaults, though: 1080P, square video, and audio are all on unless you say otherwise.

Wan 3.0 at a Glance

	Wan 3.0
Maker	Alibaba
Released	August 2026, API only
Modes	Text to video, image to video (first frame, first and last frame, or up to 10 reference images), video to video (up to 5 clips, 15 s total)
Clip length	2 to 30 seconds per call
Resolution	480P, 720P, 1080P
Audio	Generated audio, on by default
Extra inputs	A document (PDF, PPTX, DOCX, XLSX, MD and more, up to 50 pages) or a public webpage
Open weights	No. The last open Wan release is the 2.2 family
Price on Modellix	$0.045/s at 480P, $0.09/s at 720P, $0.18/s at 1080P
Model IDs on Modellix	wan3.0-t2v,  wan3.0-i2v,  wan3.0-v2v

Specs from Alibaba Cloud Model Studio and the Modellix model schemas.

Wan 3.0 vs Wan 2.7: Same Prompt, 6-Second Test

We gave both models one prompt that tests several things at once: a named camera move, small hand motion, liquid, readable text on a sign, and reflections. Settings: 6 seconds, 720P, 16:9, one take each, audio off where the model allows it.

A small orange robot barista in a tiny Tokyo alley cafe at night. It steams milk, then slowly pours latte art shaped like a heart into a white cup. A neon sign reading "OPEN 24H" glows in the window and reflects in rain puddles outside. Slow dolly-in from the doorway to a close-up of the cup. Cinematic, warm tungsten light.

Same prompt, same settings, only the model changed. Generated on Modellix: 6 seconds, 720P, 16:9, one take each.

Wan 3.0 T2V

720P · 187 s · $0.54 · audio off

Copy PromptTry it

Wan 2.7 T2V

720P · 468 s · $0.54 · audio on by default

Copy PromptTry it

Open Wan 3.0 in the Playground

	Wan 3.0	Wan 2.7
Time to finished clip	187 s	468 s
Billed	$0.54	$0.54
Camera move as written (doorway to cup)	Yes	No starts inside
Visible milk pour	Brief	Clear stream, heart forms on screen
Neon sign text "OPEN 24H"	Correct	Correct
Audio	Off, as requested	On (2.7 has no off switch)

Wan 3.0 did what the prompt asked in order: it started outside the doorway, pushed in, and ended on a close-up of the cup. Wan 2.7 skipped the doorway and opened at the counter, but it gave us the better pour, with a visible stream of milk drawing the heart. Both spelled the neon sign correctly. The robot in 2.7 also has a friendlier face, which nobody asked for and nobody minded.

The bigger gap was time: 187 seconds versus 468. At the same $0.09 per second, that makes 3.0 the default for anything you iterate on.

What changed in the API

If you are moving code from 2.7, these are the fields that break or behave differently:

Two of these cost money if you miss them. Both models default to 1080P, and on 3.0 that is $0.18/s, double the 720P rate. And 3.0 defaults to a square 1:1 frame, so set ratio every time.

When to stay on Wan 2.7

Keep 2.7 if you depend on negative_prompt, if you bring your own soundtrack through audio_url, or if you render a lot of 1080P:  Wan 2.7 is $0.135/s at 1080P against $0.18/s on 3.0. The 2.7 family also has reference-to-video, video editing, and image models that 3.0 does not replace yet. Our Wan 2.7 API guide covers those.

30-Second Shots and PDF to Video: What Only Wan 3.0 Does

One 30-second shot, no cuts

Long clips usually fail by drift: the subject slowly changes shape, color, or identity. We asked for one continuous shot of a paper boat moving through a town, with a list of things it had to pass.

One continuous shot, no cuts. A small white paper boat floats down a rain-filled street gutter in an old European town at dusk. The camera follows it at water level as it passes a bakery window, slips under a stone footbridge, spins once in a small whirlpool, and drifts out into a wide canal where lantern light reflects on the water. Keep the same paper boat, same folds, same lighting direction for the whole shot.

The boat kept the same shape, folds, and color for all 30 seconds, through the bakery, under the bridge, and out onto the canal. The "spin in a whirlpool" became a slow turn in place, and the water in the gutter sometimes flows a little uphill. It took about 13 minutes and cost $2.70. For comparison, stitching five 6-second clips costs the same per second but gives you five chances for the boat to change.

A PDF fact sheet becomes a narrated clip

file_url takes a document instead of, or alongside, a prompt. We passed NASA's two-page MAVEN mission fact sheet with one line of direction: "Turn this mission fact sheet into a short cinematic explainer with narration." 10 seconds, 720P, audio on.

The narration says: "After a 442 million mile journey, Maven fires its engine to capture orbit." That figure is in the PDF, not in our prompt. The spacecraft follows the fact sheet's illustration, and the on-screen caption reuses the document's own headline. One leak: the page header "NASA Facts" shows up as a faint vertical label at the right edge, so crop or mask headers in documents you send. This clip cost $0.90.

Webpage to video: pick the page carefully

link_url takes a public page instead of a file. Our first try, the Wikipedia article on latte art, was refused before generation with "This URL is not supported" and was not billed. Our own AI product video guide went through: 8 seconds, 720P, $0.72.

Wan 3.0 found the serum bottle that the article uses as its test product and cut a short product montage around it, label text included. It did not keep the cap straight: the bottle switches between a dropper and a plain lid from shot to shot, the kind of slip a human editor would catch. If a page is refused, save it as a PDF and send it through file_url instead.

Wan 3.0 Pricing: $0.045 to $0.18 per Second

Wan 3.0 bills per second of output. The price depends on resolution and on where you buy it:

Alibaba prices are list prices from the Model Studio page, before any promotions in the Alibaba console. "Other regions" means Beijing, Tokyo, Frankfurt, Virginia, and Hong Kong.

So Modellix is 10% below Alibaba's international (Singapore) endpoint and about 9% above Alibaba's other regions. What you get for that is one key and one bill across Wan, Kling, Veo, Seedance, and 200+ other models, with no Alibaba Cloud account to set up.

Cost per clip on Modellix

Our own runs matched the table: $0.54 for 6 seconds at 720P, $0.90 for 10 seconds, $2.70 for 30 seconds, $0.72 for the 8-second webpage clip. Audio did not change the price. The refused webpage request cost nothing.

Three defaults that raise the bill

• Resolution defaults to 1080P. Leave it out and every second costs $0.18. Draft at 480P ($0.045/s) and render finals at 720P or 1080P.
• Ratio defaults to 1:1. Square video is a quiet surprise if you expected 16:9.
• Audio defaults to on. It is not billed extra, but turn it off if you add your own track.

How to Use the Wan 3.0 API

Is Wan 3.0 open source?

No. Alibaba's Wan-AI page on Hugging Face lists open weights for the 2.1 and 2.2 families, including the Wan 2.2 Animate models. Wan 2.5, 2.6, 2.7, and 3.0 are only available through APIs. If you want to run Wan on your own machine, 2.2 is the version to use; our ComfyUI image to video test shows what that takes without an NVIDIA GPU.

Your first Wan 3.0 request

Wan 3.0 is three model IDs on Modellix: wan3.0-t2v for prompts, documents, and webpages, wan3.0-i2v for first frames and reference images, and wan3.0-v2v when you send reference video. Calls are asynchronous: submit, get a task ID, then poll.

curl -X POST https://api.modellix.ai/api/v1/alibaba/wan3.0-t2v \
  -H "Authorization: Bearer $MODELLIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A small orange robot barista pours latte art shaped like a heart. Slow dolly-in.",
    "duration": 6,
    "resolution": "720P",
    "ratio": "16:9",
    "audio": false
  }'

# then poll the task until status is "success"
curl https://api.modellix.ai/api/v1/tasks/<task_id> -H "Authorization: Bearer $MODELLIX_API_KEY"

Full parameters are in the Wan 3.0 T2V docs.

Why Modellix is the best API for Wan 3.0

Create your account →

Six ways to run these models

Playground

No code. Sign in, open a model page such as Wan 3.0 T2V, upload your image or paste a prompt, and the clip, time, and cost show up in your browser. Best for picking a shortlist.

Coding agent, no install

Paste one sentence into Claude Code, Codex, or Cursor and let it read the docs:

My Modellix API key is mdlx-••••••••••••••••••••••••. Read https://docs.modellix.ai/get-started/index.md and https://docs.modellix.ai/alibaba/wan3.0-t2v.md, then generate a 6-second 720P 16:9 clip with alibaba/wan3.0-t2v for each line in prompts.txt, audio off. Run at most 4 tasks at a time, save the MP4s to ./videos, and report what each one cost.

Swap in your own key from the console, and only share it with agents you trust.

Agent Skill

Install the skill once and your agent reads each model's schema and fills in the request fields itself. Name the model in your prompt; if you don't, it asks which one to run. Works in any Agent Skills host. Setup is in the Skill guide.

npx skills add https://github.com/Modellix/modellix-plugin --skill modellix-design

Plugin

If your host has a plugin marketplace (Claude Code, Codex, Cursor), the plugin adds six skills (generation, image prompts, social cards, infographics, and local edits with FFmpeg or MoviePy) plus the Media Generation MCP, so commands like /modellix:video and /modellix:social-card work out of the box. In Claude Code, run /plugin marketplace add Modellix/modellix-plugin, then /plugin install modellix@modellix. Other hosts are in the Plugin guide.

CLI

For scripts and cron jobs. One command submits the task, waits, and prints the result URL; modellix-cli model batch takes a JSONL file for a whole list. See the CLI guide.

npm install --global modellix-cli
modellix-cli model run --model-slug alibaba/wan3.0-t2v --body-file request.json --wait

REST API

Submit, keep the task_id, poll, then download before the result URL expires. Field names vary by provider, so check each model's schema in the API guide.

Start from a prompt that already works

Prompts from Modellix tests that already work. Paste one into the Playground and swap the model. Generated on Modellix.

One 30-second shot

“One continuous shot, no cuts. A small white paper boat floats down a rain-filled street gutter in an old European town at dusk. The camera follows it at water level as it passes a bakery window, slips under a stone footbridge, spins once in a small whirlpool, and drifts out into a wide canal where lantern light reflects on the water. Keep the same paper boat, same folds, same lighting direction for the whole shot.”

Copy PromptTry it

Food and product detail

“Honey pours slowly over a stack of pancakes topped with berries, steam rising, macro shot, bright breakfast light.”

Copy PromptTry it

Multi-shot story in one call

“Shot 1: close-up of a chef's hands plating a dessert. Shot 2: wide shot of the busy restaurant kitchen. Shot 3: the chef looks up and smiles at the camera. Warm cinematic lighting.”

Copy PromptTry it

Camera move and expression

“Slow dolly-in on a woman sitting alone at a dim jazz bar counter, warm tungsten light. As a spotlight sweeps across her face, her expression shifts from calm to quiet surprise. Shallow depth of field, 35mm film look.”

Copy PromptTry it

Browse text-to-video models

FAQ About Wan 3.0

Is Wan 3.0 free?

No. Wan 3.0 is a paid API model with no free tier on Modellix, and Alibaba bills it per second too. A 5-second 480P draft costs $0.22 on Modellix. If you want to try every major video model without signing up for each one, Modellix is the fastest and cheapest way to do it.

Is Wan 3.0 open source?

No. The newest open-weight Wan models are the 2.2 family on Hugging Face. Wan 2.7 and 3.0 are API only.

How long can a Wan 3.0 video be?

2 to 30 seconds per call, in whole seconds. Wan 2.7 stops at 15.

How much does Wan 3.0 cost?

On Modellix, $0.045/s at 480P, $0.09/s at 720P, and $0.18/s at 1080P. A 30-second 1080P clip is $5.40. Alibaba's list prices are $0.10/s at 720P in Singapore and $0.083/s in its other regions.

Is Wan 3.0 better than Wan 2.7?

For most new work, yes: it was 2.5x faster in our test, follows camera direction more closely, goes to 30 seconds, and reads documents. Stay on 2.7 for negative_prompt, your own audio track, or cheaper 1080P.

What is Wan 3.0 Prime?

Some third-party hosts list a "Prime" tier at a higher price. Alibaba's Model Studio page documents one model, wan3.0-video, and that is the model Modellix serves.

Can I use Wan 3.0 in ComfyUI?

Yes, through an API node, since there are no weights to load locally. In our ComfyUI image to video test, Wan 3.0 blocked its own output twice for a photorealistic photo of a person (no charge either time), so use illustrated or non-human first frames when you can.