Wan AI

What Wan AI is

Wan is Alibaba's video generation family and the web app that sells access to it. The marketing site is wan.video; the product lives at create.wan.video, and wan.video/pricing redirects there. The current video model is Wan 3.0: the homepage leads with "Try Wan3.0 Now" and "Introducing Wan3.0", and Alibaba Cloud Model Studio lists the model code wan3.0-video, with a faster paid variant, wan3.0-video-prime, alongside it. Images run on a separate line, wan2.7-image-pro.

What 3.0 adds is duration and input breadth, and Alibaba's Model Studio documentation states the limits plainly: 2 to 30 seconds in a single pass, output at 480P, 720P or 1080P, aspect ratios of 16:9, 4:3, 1:1, 3:4 and 9:16, a first frame and an optional last frame, and up to 20 reference assets per request, broken down as a maximum of 10 images, 5 video clips totalling no more than 15 seconds, and 5 audio tracks totalling no more than 15 seconds. A document or a public webpage can also drive a generation, one file or link per request, at most 100MB and 50 pages. The app around the model is a real product rather than a demo page: Canvas, Skills, Playground, TimeLine and Assets tabs, a CLI, an API tab, and official iOS and Android apps with a mobile web fallback for regions where the store listing is unavailable.

One thing has changed that matters more than any feature. Wan built its reputation on open weights, and that reputation is doing work the current product no longer earns. The open line stops at Wan 2.2. The Wan-Video organisation on GitHub tops out at the Wan2.2 and Wan2.1 repositories, each at roughly 17,000 stars, plus side projects such as Wan-Animate-2 and Wan-Dancer, and the Wan-AI organisation on Hugging Face has no Wan 3.0 checkpoint, only Wan2.2 variants and Wan-Dancer-14B. Wan 3.0 is a closed, paid model served through Alibaba's platforms and resellers. It is still a very good model: in ToolChase's own protocol, run through DaVinci AI, it was the strongest all-rounder of five video models at 52 of 65 points. It just is not an open model any more, and buying it as one is buying a 2025 story.

Pricing

Wan publishes every consumer price, with a working monthly and yearly toggle and no email-us wall, which puts it ahead of a good part of this category. The catch is that the toggle opens on yearly, so the first numbers you see are not what a monthly subscriber pays. There are three prices per paid tier: the yearly-billed rate, the monthly-billed rate, and a struck-through list price that the Gift Cards tab confirms is real.

All figures in US dollars, read from the rendered create.wan.video pricing page, including the DOM behind the billing toggle, as of September 2026. The page shows per-month figures only and never displays an annual total, so US$60 for a year of Pro and US$240 for a year of Premium are arithmetic rather than figures the vendor shows.

Three prices per tier, not two

The toggle defaults to "Billed Yearly (50%off)", so Pro reads US$5 and Premium US$20. Click "Billed Monthly" and the price element, which is a two-item slider, shifts to its second value: US$6.50 and US$26. The struck-through US$10 and US$40 are genuine list prices, not decoration: the Gift Cards tab sells one month of Pro for US$10 and one month of Premium for US$40 with no discount applied, which is what the sticker looks like if the promotion ends. The discount arithmetic matches the label on one toggle only. Yearly billing is 50% off list; monthly billing is 35% off, and nothing on the page says so.

Per credit the two paid tiers are identical: US$0.0167 per credit billed yearly and US$0.0217 billed monthly, for Pro and Premium alike. You are paying for concurrency and volume, not for a better exchange rate, and both subscriptions beat every top-up pack and every gift card below.

Credit packs and gift cards

The Gift Cards tab adds a second price list. Membership gift cards buy one month (31 days) of Pro for US$10 or Premium for US$40, redeemable within a year of purchase. Credit gift cards run 60 credits for US$3, 300 for US$10, 660 for US$20, 1,800 for US$50, 3,900 for US$100 and 8,000 for US$200. That last card, at US$0.025 per credit, is the best rate outside a subscription, marginally ahead of the 3,900-credit pack at US$0.0256, and still about 50% more per credit than either subscription. The restrictions are stated on the page: credits cannot be exchanged for memberships, refunded, transferred or withdrawn, and they expire two years after redemption. Subscriptions auto-renew, and the first month or year is non-refundable once activated.

API pricing is a separate purchase

The pricing page states that a subscription covers model creation on create.wan.video only, and that separate purchases are required for the API, DealDance and other products. API access goes through Alibaba Cloud Model Studio, which publishes two Wan 3.0 video models with six-region price tables on its own model pages, both last updated by Alibaba in September 2026. Prices are per second of output video.

So a 30-second 1080P clip through the API costs about US$4.95 on wan3.0-video in most regions, US$6 in Singapore, and about US$7.63 on the prime variant, which is roughly 54% dearer per second in exchange for faster generation. Resellers quote their own numbers; OpenRouter's, for example, is a percentage off the prime list price rather than a vendor figure. Budget from Alibaba's table.

How to budget credits, and why we cannot

Every tier and pack carries a line like "Accelerate: Up to 4800 images or 240 videos". Across all of them that resolves to a flat 0.25 credits per image and 5 credits per video. That is a ratio, not a rate card. Set it against the API: Premium's own strapline implies 240 videos for US$20, about eight cents a video, while the same model sells a single 30-second 1080P clip for US$4.95 through Alibaba Cloud. Both numbers cannot describe the same work. The claim is presumably reachable at the floor, on short low-resolution clips in whatever "Accelerate" mode is, and nowhere near it at the settings people actually want.

What a real generation costs in credits is published nowhere on the public site, and the generation screen requires an account. The same goes for the free tier: its credits come from a daily check-in whose amount is not stated. Those are the two numbers a buyer needs most, and neither is available without signing up.

What we tested, and what we did not

This is a pricing and specification review with one borrowed set of test results, so here is the boundary in plain terms.

Verified: every consumer price on the rendered pricing page, including the monthly toggle read from the DOM and the Credits and Gift Cards tabs; the API price tables on Alibaba Cloud's wan3.0-video and wan3.0-video-prime pages; the input limits in Alibaba's Model Studio guide; and the state of the Wan-Video GitHub and Wan-AI Hugging Face organisations.

Tested by proxy: Wan 3.0 as a model, through five of the nine tests in our video protocol, run on DaVinci AI rather than on create.wan.video.

Not tested: create.wan.video itself. The generation surface sits behind a login we did not have, so its interface, queue times, moderation behaviour, reliability, the free tier's daily credits, the editing surface, the mobile apps and the CLI are all unmeasured. That means the highest-weighted input to the score is a measurement of the model, not of the app that sells it, and the score table below says so line by line.

Test results for Wan 3.0

Wan 3.0 received the fullest run in our DaVinci AI review, and the findings transfer because the model is the same one create.wan.video markets. Scores are out of the points each test carries in our protocol, not out of 100.

Run through DaVinci AI in September 2026, using credits that vendor supplied for its own review. Four of the nine protocol tests were not run on Wan 3.0, so there is no publishable total out of 100, and nothing here measures the create.wan.video interface. Per-clip findings are on the DaVinci AI page.

Where Wan AI is the strongest pick

Wan is strongest when the shot has to hold together for longer than most models can manage. Native 30-second single-pass generation is a genuine differentiator: most competitors reach 30 seconds by extending shorter clips and pay for it in drift, and our own testing is where Wan earned its result, holding people, wardrobe and timed action across 15 and 30 second takes. If your work needs one coherent 20 to 30 second shot rather than a stitched sequence, this is the model to try first.

The second argument is the input surface. Up to 20 reference assets across image, video and audio, plus a document or webpage as the brief, is the widest set of controls we have catalogued in this category, with the caveat that the document input takes one file or link at a time. The third is price: US$20 a month billed yearly for 1,200 credits and eight concurrent video jobs sits below Runway Pro at $35, Kling Pro at $37, Luma Plus at $30 and Pika Standard at $28 on this site's own verified figures, though what those credits buy per clip remains unpublished.

Best for

Creators and small production teams who want long single-pass shots and heavy reference control at one of the lowest published prices in the category, and who are willing to work out real per-generation costs by testing rather than from a rate card. It suits anyone whose work needs a coherent 20 to 30 second take, or who wants to drive a shot from a stack of image, video and audio references rather than from a prompt alone. It is a poor fit for teams that came specifically for open weights and self-hosting, since that road ends at Wan 2.2, and for anyone who needs firm per-asset budgeting or a support contract before committing.

Key features

• Wan 3.0 generates up to 30 seconds in a single pass, with any duration from 2 to 30 seconds selectable
• Output at 480P, 720P or 1080P (the API default), at 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive, as MP4 at 30fps
• Up to 20 reference assets per request: a maximum of 10 images, 5 video clips totalling no more than 15 seconds, and 5 audio tracks totalling no more than 15 seconds, per Alibaba's Model Studio guide
• Document and webpage input, new in the 3.0 line: one file or link per request, at most 100MB and 50 pages, in doc, xls, ppt, pdf, txt and md formats among others
• First-frame pinning with an optional last frame, and video extension where input plus output stays within 30 seconds
• Audio as an input modality alongside text, image, video, file and link. Generated sound is less settled: the API guide exposes an audio switch on generated clips and Runway's listing describes native audio, but Alibaba's own model pages list video as the only output modality and its launch post says audio texture is still improving, so we treat generated sound design as unverified rather than as a headline feature
• A precision video editing surface separate from generation, marketed on the homepage and unmeasured here
• Text-to-image, image-to-image and image editing alongside video, served by a separate model line (wan2.7-image-pro)
• Concurrency scales with tier: 1 video and 1 image at Free, 3 and 3 at Pro, 8 and 5 at Premium
• Image upscaling, watermark-free downloads, 1080p output and 10 to 30 second videos, all gated to the paid tiers
• A full app shell rather than a prompt box: Canvas, Skills, Playground, TimeLine and Assets tabs, plus a CLI and an API tab
• Official iOS and Android apps, with a mobile web fallback where the store listing is unavailable
• API access through Alibaba Cloud Model Studio as wan3.0-video and the faster wan3.0-video-prime, plus hosting on Runway, resale through OpenRouter, and inclusion in aggregators such as DaVinci AI
• Apache 2.0 open weights for the older Wan 2.1 and Wan 2.2 lines and side models such as Wan-Dancer-14B and Wan-Animate-2, still published on GitHub and Hugging Face

Best-fit use cases

• Producing a coherent 20 to 30 second shot in one pass, where stitched extensions from other models would visibly drift
• Driving a shot from a stack of references (product images, a style clip, a voice track) rather than from a text prompt alone
• Turning a document or webpage directly into video, which is unusual enough in this category to be worth testing on its own
• Short-form social video for teams whose budget will not stretch to Runway, Luma or Kling Pro subscriptions
• Character-consistent sequences where the same person has to survive several shot changes
• Prototyping on the cheap web tiers, then moving proven prompts to the Alibaba Cloud API for production volume, budgeted as a separate purchase
• Self-hosting an older Wan 2.1 or 2.2 checkpoint under Apache 2.0 where data residency or cost rules out an API

Pros

• Genuinely cheap for the output class: US$20 a month billed yearly buys 1,200 credits and eight concurrent video jobs, under the Pro tiers of Runway, Kling, Luma and Pika
• Native 30-second single-pass generation is a real differentiator; most competitors get there by extending shorter clips and pay for it in drift
• The widest reference system we have catalogued in this category: up to 20 assets across image, video and audio, plus a document or webpage as the brief
• Strongest all-rounder of five models in our own video protocol, run through DaVinci AI: 52 of 65, holding people, wardrobe and timed action over 15 and 30 second takes
• Every consumer price is published, with a working monthly and yearly toggle and no email-us wall
• Both paid tiers price credits identically, so you are choosing throughput rather than being upsold a better exchange rate
• Subscription credits are cheaper per credit than every top-up pack and every gift card, so the default path is the good-value path
• API pricing is first-party and public, with two model variants and six regions on Alibaba Cloud's own pages
• A real product surrounds the model: canvas, timeline, asset library, CLI, API and shipped mobile apps
• The older Wan 2.1 and 2.2 weights remain Apache 2.0 and widely supported, a genuine fallback if you need to self-host

Cons

• Wan 3.0 is not open weights: the open line stops at Wan 2.2, and no 3.0 repository or checkpoint exists on either the GitHub or Hugging Face organisation
• The pricing page defaults to the yearly rate, so the headline US$5 and US$20 are not what a monthly subscriber pays; that is US$6.50 and US$26
• The advertised "50% off" applies only to yearly billing; monthly billing is 35% off list, and the page never says so
• Capacity claims like "up to 240 videos" are floor figures: taken literally, Premium would deliver two hours of 30-second video for US$20, which the same model's API price contradicts by a factor of about sixty
• Per-generation credit costs are not published anywhere on the public site, so you cannot budget before you pay
• The free tier's daily check-in amount is undisclosed, so the free plan cannot be evaluated without signing up
• A subscription explicitly does not cover API use; the page says separate purchases are required
• Credits are non-refundable, non-transferable, cannot be converted to a membership, and expire two years after redemption
• The first month or year is non-refundable once activated, so an annual plan is a genuine commitment
• "Sound design" is documented but not hands-on verified: Alibaba's Model Studio guide exposes an audio parameter that is on by default, so generated audio is something to test before you rely on it, not assume
• Support is a single service email address and a user guide, with no published response time or live channel
• Camera direction was the model's weak spot in our testing: locked-off ignored, tilt up misread as move along

What to watch out for

Four things, in order of how much they will cost you.

Check which billing toggle you are looking at. The pricing page opens on "Billed Yearly (50%off)", so the first prices you see, US$5 and US$20 a month, are the annual rates. Click "Billed Monthly" and they become US$6.50 and US$26. The page never shows an annual total, so nobody tells you that the Premium yearly plan is a US$240 commitment, and the terms say the first year is non-refundable once activated. The "50% off" badge is only true on the yearly toggle; monthly billing is 35% off the US$10 and US$40 list prices. Those list prices are real: the Gift Cards tab sells one month of Pro for US$10 and one month of Premium for US$40 with no discount at all.

Do not budget from the "up to" figures. Every tier and every pack carries a line like "Accelerate: Up to 4800 images or 240 videos", which resolves to a flat 0.25 credits per image and 5 credits per video no matter what you buy. That is a ratio, not a price list. Wan's own API sells the same model at US$0.165025 per second at 1080P in most regions, so a single 30-second 1080P clip costs about US$4.95 through the API, while Premium's claim implies 240 of them for US$20. Both numbers cannot describe the same work. We could not check the real per-generation cost because the generation surface requires a login, and that is precisely the number a buyer needs.

If you came for the open model, stop here. Wan earned its reputation by shipping Apache 2.0 weights, and a lot of writing still describes it that way. It is no longer true of the current model. The Wan-Video GitHub organisation tops out at Wan2.2, and the Wan-AI Hugging Face organisation has no 3.0 checkpoint either. Wan 3.0 is closed and paid. Alibaba has kept publishing smaller models openly, so the shift is specific to the flagship rather than a general retreat, but the practical effect is the same: you cannot self-host the model this page is about.

Your subscription does not buy API access. The pricing page states outright that the purchase is for model creation on create.wan.video only, and that separate purchases are required for the API and other products. Anyone planning to prototype in the web app and then move to the API should price that as two products, and should pick the right variant: wan3.0-video-prime is faster and about 54% dearer per second than wan3.0-video.

Two smaller notes. Credits expire two years after redemption and can never be refunded, transferred or turned back into a membership, so buying a large pack against future work carries real risk. And the homepage's "sound design" promise rests on an audio switch documented in Alibaba's Model Studio guide that we could not exercise, so treat generated audio as something to test rather than assume.

Is Wan AI worth it?

Yes, for a specific buyer, and the case rests on two things we could actually measure. The model is the best all-rounder we have run on people, wardrobe and timed action over long takes, and the consumer prices are low and fully published: US$20 a month billed yearly for 1,200 credits and eight concurrent jobs is below the Pro tiers of Runway, Kling, Pika and Luma on this site's verified figures. If your work is long, character-led shots and your budget is small, there is no obvious reason to pay more elsewhere.

It scores 4.1 rather than higher for reasons that are all on the commercial surface and one that is ours. The pricing page opens on the flattering toggle, the "50% off" badge is true on one toggle only, no annual total is shown, and the capacity claims are written at the floor while per-generation credit costs are published nowhere. Support is an email address. And the weights that made Wan's name are two versions behind the model you are paying for. None of that is about output quality; all of it is what a buyer has to deal with before generating a frame.

The reason that is ours: this score is model-led. Product quality carries the most weight, and we measured it by running the model through DaVinci AI, not by using create.wan.video. The app's interface, queue times, moderation and reliability are unmeasured, and the score can move in either direction once someone with a login tests them.

How we arrived at 4.1

The score is weighted across nine categories. The largest single input is the model measurement borrowed from our DaVinci AI protocol; the app itself is scored on what could be observed from outside a login.

The 5 point score is a product rating rather than a linear map of that total, and it reflects a strong model at a low published price, held back by a buying experience built to flatter the headline and by an app we could not test.

FAQ

Consumer pricing read from the rendered create.wan.video pricing page, including the Credits and Gift Cards tabs and the DOM behind the billing toggle; API pricing and input limits read from Alibaba Cloud Model Studio's wan3.0-video, wan3.0-video-prime and video generation guide pages; open-weights status read from the Wan-Video GitHub and Wan-AI Hugging Face organisations; all as of September 2026. Model test results come from our DaVinci AI review. ToolChase is not an affiliate of Wan or Alibaba and received no payment for this review. Scores, rankings and list positions are never sold. See our methodology for how we score tools.