Wan 3.0 vs Seedance 2.5: Price, Control and What Nobody Has Measured

Alibaba's Wan 3.0 and ByteDance's Seedance 2.5 both generate a 30-second clip in one pass, and the cheaper of the two costs $6.00 at 1080p.

The pricier one costs about $17 for the same clip. That gap is the first thing a budget notices, but it is not what decides this choice. Wan 3.0 is a plain metered API that also reads documents. Seedance 2.5 is a heavier toolkit: up to 50 reference files, editing, extension, and an API that BytePlus now says is live.

We did not run a blind test of the two, and this guide does not rank how they look. It compares what each vendor documents. We opened Alibaba's Model Studio pricing page and BytePlus's ModelArk docs on 6 October 2026 for every price and limit below, and each gap is marked where it appears.

What are Wan 3.0 and Seedance 2.5, and where do they overlap?

Both are closed, API-first video models from the two biggest Chinese platform companies. Alibaba Cloud sells Wan 3.0 through its Model Studio. ByteDance sells Seedance 2.5 through BytePlus ModelArk, the cloud arm that also serves its text and image models.

Each promises a single continuous clip of up to 30 seconds. BytePlus says so in its Seedance 2.5 prompt guide, and a Reuters report relayed by AI Understanding says Alibaba rolled Wan 3.0 out officially on 25 August with the same 30-second ceiling.

Neither company has published weights for these versions. HokAI's review of Wan 3.0 found nothing on Hugging Face, GitHub or ModelScope, and traced the open-weight rumour to a single unverified LinkedIn post. Treat both as rented, not owned.

The overlap ends at the clip length. Everything after that is a different product philosophy.

Wan 3.0 costs between a third and half as much per second

Price is the cleanest difference, so start there. The table uses Alibaba's international list prices and BytePlus's own 5-second examples at 16:9 with no video input.

The two bottom rows are our arithmetic. BytePlus bills Seedance by token, not by the second, and prints examples only for 5-second clips. We scaled them linearly, which fits how it says tokens are counted (duration times pixels times frame rate), but the real bill is whatever the API returns.

Alibaba's pricing page prints the Wan figures as list prices and flags a limited-time 30% discount, so a bill today may come in lower. It also lists a second ID, wan3.0-video-prime, at $0.068 for 480p and $0.28 for 1080p.

Input video changes the maths on both sides. Alibaba bills input and output seconds together. BytePlus charges a 5-second 720p clip $1.156 with no video input and up to $4.838 with a 30-second input video.

Failed generations are free on both platforms. BytePlus says content-moderation failures are not billed either.

Seedance has one cost lever Wan does not document: draft mode. BytePlus splits it into two billed tasks. Step one makes a draft video, billed as 480p. Step two renders the final clip at your target resolution. If you iterate on ideas before you commit, drafting at the 480p rate is a cheaper way to reject a bad shot.

BytePlus also sets a minimum token charge when the input includes video, tied to resolution, aspect ratio and output length. A tiny input clip can therefore cost more than its length suggests, and its pricing calculator is the safer way to estimate.

Where Wan 3.0 wins: price, simplicity and documents

Wan 3.0 wins on cost per finished second, and the margin is not small. At 1080p a 30-second clip is $6.00 against roughly $17.07, which pays for nearly three Wan attempts for every Seedance one.

It also asks less of you. You send a prompt, optionally a video, and pay by the second. There is no token calculator, and we found no minimum-charge rule for video input or draft-then-final workflow to learn.

Its stranger strength is input type. Per Reuters, Alibaba says Wan 3.0 can turn documents, spreadsheets, slides and web pages into video. HokAI's record lists PDF input alongside text, image, audio and video. If your source material is a product spec or a slide deck, that is a shorter path than writing a prompt by hand.

One cost trap runs the other way. Alibaba's own page lists the older wan2.7-t2v at $0.15 per second for 1080p, below Wan 3.0's rate. If you do not need a 30-second shot, the older model is the better buy.

Alibaba's page also carries a rival model from the same shelf. HappyHorse 1.1 lists 1080p output a little under Wan 3.0's rate, and it is worth a test if clip length is not your constraint.

Wan-Animate-2 covers a different job, character animation, and is not priced here.

Where Seedance 2.5 wins: control, editing and a live API

Seedance 2.5 wins on what you can steer. Per BytePlus's prompt guide, one request takes up to 50 reference assets: up to 30 images at up to 4K, up to 10 videos with a combined 30 seconds, and up to 10 audio clips with the same cap.

The ratio line is a real change from Seedance 2.0, which supported six fixed output shapes. Timestamps are new too: 2.0 responded only to shot numbers, while 2.5 reads integer-second marks such as "0-3s".

The task list is wider than Wan's documented one. BytePlus lists subject, motion, style, audio and storyboard references, first and last frames, video editing, audio editing, forward and backward extension, and a generated in-between segment that joins two clips.

BytePlus splits its tasks into locked and unlocked ones, and the split matters for planning. Editing, first-and-last-frame generation and extension lock the output aspect ratio to the input, and editing also locks the duration. Reference tasks, storyboards and keyframes stay free, so you choose the ratio and length yourself.

It also advertises native generation in more than 10 languages, which matters if your ads run in several markets. We could not find a comparable claim in Alibaba's documentation for Wan 3.0, so that row stays blank rather than guessed.

Then there is the news that changes the whole comparison. Both BytePlus pages we opened carry a banner saying the official Seedance 2.5 API is live. HokAI's own record, last reviewed on 14 September, still called API access unconfirmed, so that record is now behind.

One conflict is worth flagging. The reseller Atlas Cloud advertises Seedance 2.5 output up to 4K, but BytePlus's price table has no 4K row for it. The 4K rows belong to Seedance 2.0, where a 5-second 4K clip lists at $3.89.

Which one looks better? Nobody has measured both

This is the honest hole in every comparison of these two. We found no independent benchmark that scores both. HokAI's records for Wan 3.0 and Seedance 2.5 both say none exists, and a leaderboard check on Artificial Analysis turned up no Seedance 2.5 entry.

A same-prompt test from Atlas Cloud is the closest thing. It ran ten briefs through Wan 3.0, Seedance 2.5 and MiniMax H3, and reports no clean sweep. Seedance kept in-frame text crisp, MiniMax held human identity, and Wan coped best with the longest prompts.

Read that test with its context. Atlas Cloud resells all three models, the clips are video we could not score ourselves, and the post is a vendor blog dated 21 August.

So the working rule is simple. Treat any "X beats Y" headline on these two as a claim, not a measurement. The HokAI model leaderboard shows benchmark scores where they exist, and for both of these it will show gaps.

If quality decides your purchase, generate your own prompt on both before you commit. Both bill only for successful generations, so a short test costs a few dollars, and Alibaba's free quota covers 30 seconds on Wan.

What goes wrong with each, and what ByteDance admits

ByteDance deserves credit here. Its prompt guide ends in a FAQ of failure modes, and the list is long enough to plan around.

On text and sound:

• Subtitles can appear despite a "no subtitles" line.
• Reference videos that carry subtitles tend to pass them on.
• English words can be misspelled, with E swapped for A or I.
• Chinese characters can come out wrong.
• Background music can survive a "no BGM" instruction.

On images and characters:

• Strong emotion words can make pupils glow.
• Characters can swap with the wrong reference image or voice.
• The same prompt and image look different from Seedance 2.0, so mixed workflows drift in style.
• Some JPG files fail with an HTTP 400 error until converted.

Edits come back about 0.3 seconds shorter than the input unless you set the frame count to 8n+1. Complex jobs that stack several reference and editing demands are unstable, and BytePlus advises splitting them into a reference pass and an edit pass.

Alibaba has published nothing comparable for Wan 3.0. HokAI's record finds no system card, no safety benchmark and no training-data disclosure. That is a gap in documentation, not evidence the model is cleaner.

Who should pick Wan 3.0, and who should pick Seedance 2.5?

Pick Wan 3.0 if the clip is a single shot, the budget is tight, or the source is a document. Pick Seedance 2.5 if you need to match a character, a camera move or a brand look across several shots, or to edit and extend footage you already have.

• A marketer turning a slide deck into an explainer: Wan 3.0, since documents are a stated input.
• A studio keeping one character consistent: Seedance 2.5, for the reference stack.
• A team editing footage it already shot: Seedance 2.5, since Wan's editing shows up only in its billing rules.
• A solo creator testing ideas cheaply: Wan 3.0 at 480p, or the older Wan 2.7 for short clips.
• A buyer who needs real 4K: neither, today. BytePlus prices 4K only on Seedance 2.0.

For a broader field, the AI video generator directory lists the tools built on these engines, Kling from Kuaishou and Higgsfield among them. Our wider video generator buyer's guide covers the category by job.

If the answer depends on your stack and your budget, Smart Match shortlists models from a plain description of the job.

How do you test both for under $5?

A three-prompt test is cheap enough to run before you commit a budget. Both vendors bill only successful generations, and you can read the clips yourself.

Run it like this:

1. Pick three prompts you would actually ship: one with on-screen text, one with a person, one with a planned camera move.
2. Generate each at 720p and 5 seconds on both models.
3. Count attempts, not just results. Note how many generations each prompt needed before you accepted a clip.
4. Shuffle the files and let two colleagues pick favourites without knowing the model.
5. Divide Seedance's cost per accepted clip by Wan's and compare it with the break-even ratios below.

For Seedance, write the prompt the way BytePlus recommends: bind each asset by number, use timestamps, and state negative constraints plainly.

Asset bindings: Image 1 is the sneaker, Audio 1 is the voiceover.
0s-3s: Close-up of the sneaker from Image 1 on a wet street, slow push in.
3s-8s: Wide shot, the runner steps into frame, voice from Audio 1.
No subtitles. No BGM; environmental sound only.

Audio matters in this test. Seedance generates sound and picture together in one pass, and HokAI's record lists audio output for Wan 3.0 as well. Judge the clips with the sound on, or you will miss the part that fails most often in the ByteDance FAQ.

If you cut the winners into a finished piece, a desktop editor such as CapCut handles the assembly. Our video editing and creation guide compares the finishing tools, and the MiniMax and Lightricks pages show who else makes video models worth adding to the shortlist.

What about switching, data and rights?

Switching cost is low on paper and real in practice. Both are HTTP APIs with their own model IDs: wan3.0-video on Alibaba and dreamina-seedance-2-5-260628 on BytePlus. A prompt that works on one will not behave the same on the other, because Seedance rewards numbered asset references and timestamps that Wan's docs never mention.

Seedance also has its own cheaper fallback. BytePlus lists Seedance 2.0 at $0.15 per second for 720p, well under 2.5's rate, with Fast and Mini variants cheaper again. If your clips are short and your references few, the older model may be all you need.

Plan the free tier honestly. Alibaba grants 30 seconds of combined input and output on Wan in Singapore for 90 days from activation. BytePlus lists no equivalent for Seedance 2.5 on the pages we read.

Data handling is the open question on both. ByteDance documented a block on real-face images for Seedance 2.0, and HokAI could not confirm it carries over to 2.5. Alibaba has not confirmed a provenance mark for Wan 3.0 output either. Ask each vendor in writing before you use either on footage of real people.

Rights are the part nobody here documents well. We found no commercial-use terms for generated clips on the pages we opened, so read the platform agreement before you sell the output.

Seedance 2.5 is also online-only. BytePlus lists offline, batch-style inference as not supported yet, which matters if you queue thousands of clips overnight. A host such as Replicate may offer its own queueing, though we did not check.

Is price per second the wrong unit?

A sharp reader will push back here. Price per second ignores retries. You pay for every attempt, and if Seedance's reference controls get you a usable clip in one try while Wan takes four, the cheap model is the expensive one.

That is fair, and it is also a number nobody has published. What we can do is set the break-even. At 720p Seedance costs 2.31 times as much per second as Wan, so it only wins on cost if it needs under 43% of Wan's attempts. At 1080p the ratio is 2.85, which cuts the threshold to 35%.

Those are steep hurdles for a model whose own FAQ lists subtitle leaks, text errors and reference mix-ups. They are not impossible hurdles. A job with a locked character and a fixed camera move may well clear them.

So the verdict stands, with its edge named. Wan 3.0 is the default buy for single shots. Seedance 2.5 is the specialist, worth the premium only when the control it adds removes retries you can count.

What would reverse the call

Four events could do it. Artificial Analysis listing both models in one arena would replace our "nobody has measured" with a number, and that is the single most useful event. Alibaba ending its 30% promotion would raise Wan's effective price. BytePlus adding a 4K tier to the 2.5 rate card would reopen the resolution row. A confirmed open-weight Wan 3.0 would change the rights question completely.

Watch the Alibaba and BytePlus pricing pages first: both were updated in the past ten days, and BytePlus's limited-time discounts on Seedance 2.0 Fast and Mini are scheduled to end on 7 October.

You can check how both models compare with Kling 3.0, Veo 3.1 and LTX-2.5 in the model recommender, and see the photo-to-video angle in our free photo-to-video guide. The next price move on either page will tell you who is chasing whom.

Frequently asked questions

Is Wan 3.0 cheaper than Seedance 2.5?

Yes, per second. Alibaba lists Wan 3.0 at $0.10 for 720p and $0.20 for 1080p, while BytePlus's examples put Seedance 2.5 at $0.231 and $0.569. A 30-second 1080p clip works out near $6.00 on Wan against about $17 on Seedance.

Can Seedance 2.5 edit and extend an existing video?

Yes. BytePlus documents video editing, audio editing, forward and backward extension and a generated in-between segment that joins two clips. Editing locks the output aspect ratio and duration to the input, and the result can come back about 0.3 seconds shorter.

Is Wan 3.0 open source?

No evidence says so. HokAI's review found no Wan 3.0 weights on Hugging Face, GitHub or ModelScope, and traced the open-weight rumour to one unverified post. Older Wan 2.x releases are the open ones, so treat Wan 3.0 as API-only.

Does Seedance 2.5 support 4K output?

Not on BytePlus's price table. The rate card for Seedance 2.5 stops at 1080p, and the 4K rows belong to Seedance 2.0. A reseller advertises 4K for 2.5, so confirm with your provider before you plan around it.

Which model has the better video quality?

Nobody has published an independent test that scores both. A reseller's same-prompt test found no clean winner, with Seedance stronger on in-frame text and Wan coping better with long prompts. Generate your own prompt on both, since failed generations are not billed.

Covered in this guide

• Wan 3.0: Wan 3.0 is Alibaba Tongyi Lab's video model, in public beta since August 6, 2026, generating up to 30 seconds of native 1080p video from text, image, audio, video, or documents.
• Seedance 2.5: Seedance 2.5 is ByteDance's 2026 video generation model, producing one continuous multi-reference clip per pass with synchronized audio, launched inside Jimeng AI and Doubao Pro.
• Alibaba Cloud: Alibaba Cloud, founded 2009, is China's top cloud provider (35% share) and builds the Qwen AI models, which passed 1 billion Hugging Face downloads in 2026.
• ByteDance: ByteDance is a Beijing-based consumer-internet group whose short-video apps and Seed AI models power the Doubao assistant.
• CapCut: AI-powered video editor for creating trending content across all platforms
• HappyHorse 1.1: Built by Alibaba's ATH Business Unit, HappyHorse 1.0 generates synchronized video and audio together from a text, image or reference-image prompt.
• Higgsfield: AI video and image studio, used in 238 countries, that runs Kling, Veo, Seedance and Sora plus its own camera and character tools on one credit plan.
• Kling 3.0: Kling 3.0 is Kuaishou Technology's proprietary AI video model, built for native 4K output with synchronized multilingual audio in one generation pass.
• Kling: Kuaishou's text-to-video platform with native 4K, lip-synced audio in 5 languages, and free daily credits. Used by 60M+ creators; plans from $6.99/month.
• Kuaishou: Kuaishou Technology (HKEX:1024), founded 2011, runs China's second-largest short-video platform with 413M DAUs and created Kling AI, top-ranked video generator in 2026.
• Lightricks: Jerusalem-based creative software maker whose LTX division builds an open-weights video and audio model family, split from Facetune in 2026.
• LTX-2.5: Lightricks' open-weights video and audio model that renders synchronized 4K HDR clips from text, image or video input, self-hostable and fine-tunable.
• MiniMax: Listed in Hong Kong since January 2026 and controlled by founder Yan Junjie, MiniMax builds the MiniMax M3 language model and the open-weight H3 video model from its Shanghai base.
• MiniMax H3: MiniMax H3 (Hailuo 3.0) is a 33B open-weight omni-modal model, released July 31 2026, generating 2K/24fps video with native stereo audio in one forward pass.
• Replicate: Run open-source AI models via API without managing any GPU infrastructure.
• Veo 3.1: Google DeepMind's Veo 3.1 generates 4, 6 or 8-second video clips with native synchronized dialogue and audio, rendering up to 4K.
• Wan-Animate-2: Wan-Animate-2 is Alibaba's open-weight 14B model for character animation, released in 2026, feeding driving video straight into a diffusion transformer.

Sources