Seedance 2.5 vs WAN 3.0 vs Kling 4.0: Specs, References and Price Compared

Three of the video models people ask about most are live on WaveSpeedAI at the same time: Seedance 2.5 from ByteDance, WAN 3.0 from Alibaba’s Tongyi Lab and Kling 4.0 from Kuaishou. All three make clips of up to 30 seconds with audio, and all three have edit and extend endpoints, so the differences are in the details.

This is a spec-level comparison, not a benchmark. We have not run head-to-head quality tests, so the guide below compares what each model accepts and how it is billed, then shows how to test them against each other on your own material. Values come from the live WaveSpeedAI model pages and may change; check the linked pages before you build on them.

Quick answer

Pick Seedance 2.5 when you need a resolution above 1080p (a 4K tier), a large reference set (up to 30 reference images), or six aspect ratios including 4:3 and 3:4.
Pick WAN 3.0 when cost per clip drives the decision, when you want a 480p tier for drafts, or when you need the shortest minimum clip (2 seconds) and up to 5 audio references.
Pick Kling 4.0 when you need 21:9, keyframes between the first and last frame, or edits of source clips up to 30 seconds with a 720p or 1080p target.
Test all three on your own prompts; the specs below tell you what each accepts, not which result you will prefer.

Specs side by side

	Seedance 2.5	WAN 3.0 (and WAN 3.0 Prime)	Kling 4.0
Clip length	4 to 30 seconds (default 5)	2 to 30 seconds (default 5)	3 to 30 seconds (whole seconds, default 5)
Resolution	480p, 720p (default), 1080p, 4K	480p, 720p (default), 1080p	720p (default), 1080p
Aspect ratios	16:9, 9:16, 4:3, 3:4, 1:1, 21:9	16:9, 9:16, 1:1, 4:3, 3:4	16:9, 9:16, 1:1, 21:9
Native audio	Yes (`generate_audio`, on by default)	Yes (`generate_audio`, on by default)	Yes, synchronized stereo audio
First and last frame	Yes (`image`, optional `last_image`)	Yes (`last_image` on image-to-video)	Yes (`image`, `last_image`, plus optional `keyframes`; at most 10 images in total)
References	Up to 30 images, 10 videos and 10 audio clips (video and audio each capped at 30 seconds in total)	Up to 10 images, 5 videos and 5 audio clips (video and audio each capped at 15 seconds in total)	Up to 10 images, 5 videos and 3 audio clips (15 files at most)
Edit existing video	Yes, source trimmed to 30 seconds	Yes, up to 15 seconds of source	Yes, 3 to 30 seconds of source (longer is trimmed to 30)
Extend existing video	Yes, adds 4 to 30 seconds	Yes, adds 2 to 30 seconds	Yes, adds 3 to 30 seconds
Faster tier	Seedance 2.5 Turbo (720p, 1080p)	WAN 3.0 Prime	None on WaveSpeedAI

Where they differ

Resolution and aspect ratio

Seedance 2.5 is the only one of the three with a 4K tier. WAN 3.0 stops at 1080p, but it is the only one that also offers 480p, which is useful for cheap drafts before you spend on a final render. Kling 4.0 on WaveSpeedAI is 720p (default) or 1080p.

For aspect ratios, Seedance 2.5 covers the widest set (it adds 4:3, 3:4 and 21:9), WAN 3.0 adds 4:3 and 3:4 but not 21:9, and Kling 4.0 adds 21:9 but not 4:3 or 3:4. On image-to-video, Seedance 2.5 follows the aspect ratio of the input image.

Clip length

All three reach 30 seconds in a single request. The lower bound differs: WAN 3.0 starts at 2 seconds, Kling 4.0 at 3 and Seedance 2.5 at 4. If you need very short clips such as 2 to 3 second loops or stingers, that matters; if you only make 5 to 10 second shots, it does not.

References

Seedance 2.5 text-to-video takes the largest reference set: up to 30 images, 10 videos and 10 audio clips, with the combined video and audio each capped at 30 seconds. WAN 3.0 reference-to-video accepts up to 10 images, 5 videos and 5 audio clips, with the combined video and audio each capped at 15 seconds, and at least one reference is required. Kling 4.0 reference-to-video accepts up to 10 images, 5 videos and 3 audio clips, 15 files at most.

If your brief has many separate assets (a character sheet, several products, a set of locations), the 30-image limit is the practical difference. If it is built around a few images plus a reference video and a voice or music clip, all three can do it.

First frame, last frame and in-between frames

All three take a first frame and a last frame on image-to-video. Kling 4.0 image-to-video also accepts keyframes, in-between frames given in order, with image, last_image and keyframes together limited to 10 images. Seedance 2.5 and WAN 3.0 do not have an equivalent in-between frame list on image-to-video.

Edit and extend

All three have dedicated endpoints for changing or continuing a clip you already have.

Seedance 2.5 video-edit follows the length of the input video (inputs over 30 seconds are trimmed to 30, inputs under 4 seconds are padded to 4), accepts up to 30 reference images and 10 audio clips, and outputs 480p to 4K. Video-extend uses up to the last 30 seconds of your clip as context and adds 4 to 30 seconds.
WAN 3.0 video-edit revises up to 15 seconds of existing footage with image and audio references, and its video-extend adds 2 to 30 seconds.
Kling 4.0 video-edit takes a 3 to 30 second source, keeps the scene, subject and motion, and keeps the original sound by default. Video-extend adds 3 to 30 seconds.

If you edit long footage, note the source-length caps: 15 seconds for WAN 3.0 against 30 seconds for the other two.

Faster tiers

Seedance 2.5 Turbo (720p and 1080p) and WAN 3.0 Prime are faster, higher-throughput variants of their families with the same tasks. Kling 4.0 has no faster tier on WaveSpeedAI. If throughput or turnaround matters for batch work, test the faster tier alongside the standard one.

Pricing

All three are billed per use on WaveSpeedAI, with no subscription. List prices per second of generated video, at the time of writing:

	480p	720p	1080p	4K
Seedance 2.5	$0.18	$0.36	$0.90	$1.80
WAN 3.0	$0.05	$0.10	$0.20	n/a
Kling 4.0	n/a	$0.157	$0.235	n/a

A few things to keep in mind:

A 5 second 720p clip lists at about $1.80 on Seedance 2.5, $0.50 on WAN 3.0 and $0.785 on Kling 4.0. The same clip at 1080p lists at $4.50, $1.00 and $1.175.
Prices for edit, extend and for reference-to-video with reference videos are counted differently. Seedance 2.5 video-edit and video-extend bill the input and output seconds together (for example a 5 second input with a 5 second output at 720p is 10 counted seconds at $0.22). WAN 3.0 reference-to-video counts the output length plus the normalized reference-video length. Kling 4.0 reference-to-video and video-edit count the output length plus the longer of the output length and the reference-video length.
WAN 3.0 Prime and Seedance 2.5 Turbo have their own rates; see their model pages.

Because the price basis differs, compare cost per approved clip rather than per-second rates alone. The Generate button on each model page shows the current estimate, and the final task charge prevails. See the Seedance 2.5 API page, the WAN 3.0 API page and the Kling 4.0 API page for current prices.

How to test the three against each other

Build a matched prompt set. Same prompts, same reference assets, same aspect ratio and the same target length wherever all three support it. Stay within the shared options: 16:9, 9:16 or 1:1, 720p or 1080p, and 4 to 30 seconds.
Cover the jobs where they differ. Very short clips (2 to 3 seconds), large reference sets, reference-matched scenes with audio, edits of an existing clip, extensions that continue a clip, and any 21:9, 4:3, 3:4 or 4K deliverables.
Keep settings identical. Same duration, same resolution, same source clips for edit and extend.
Score production outcomes. Identity drift, motion quality, audio sync, how often a clip is approved on the first try, and total cost per approved clip, not only the price per second.
Map jobs to models rather than choosing one model for everything.

The WaveSpeedAI Video Generator is the fastest first pass before you script the full comparison.

One integration for all three

The three models share the WaveSpeedAI request flow: POST https://api.wavespeed.ai/api/v3/<model id>, receive a prediction ID, then poll https://api.wavespeed.ai/api/v3/predictions/<id>/result. A small map keeps the choice in configuration:

MODELS = {
    "draft": "alibaba/wan-3.0/text-to-video",
    "final_1080p": "kwaivgi/kling-v4.0/text-to-video",
    "many_references": "bytedance/seedance-2.5/text-to-video",
    "master_4k": "bytedance/seedance-2.5/text-to-video",
    "edit": "kwaivgi/kling-v4.0/video-edit",
    "extend": "alibaba/wan-3.0/video-extend",
}

def payload(job: str, prompt: str) -> dict:
    base = {"prompt": prompt}
    if job == "draft":
        return {**base, "resolution": "480p", "duration": 5}
    if job == "master_4k":
        return {**base, "resolution": "4k", "duration": 8}
    return {**base, "resolution": "1080p", "duration": 8}

Bottom line

Seedance 2.5, WAN 3.0 and Kling 4.0 cover the same ground on WaveSpeedAI: clips up to 30 seconds, audio, references, edit and extend. Seedance 2.5 adds a 4K tier, the biggest reference set and the widest aspect-ratio list; WAN 3.0 is the lowest-priced and adds a 480p tier and the 2 second minimum; Kling 4.0 adds 21:9 and in-between keyframes. Keep all three available, decide per job, and let your own approved-clip rate and cost settle the rest.

FAQ

Which is best: Seedance 2.5, WAN 3.0 or Kling 4.0?

We have not benchmarked the three head to head, so we do not rank them. They differ in concrete ways: Seedance 2.5 has a 4K tier, up to 30 reference images and the highest per-second price; WAN 3.0 is the lowest-priced and starts at 480p; Kling 4.0 adds 21:9, keyframes and a 3 to 30 second range with 720p and 1080p output. Choose by the job, then test on your own prompts.

Which of the three is the cheapest?

WAN 3.0 has the lowest list price per second at the time of writing: $0.05 at 480p, $0.10 at 720p and $0.20 at 1080p. Kling 4.0 is $0.157 per second at 720p and $0.235 at 1080p. Seedance 2.5 is $0.18 at 480p, $0.36 at 720p and $0.90 at 1080p. Cost per approved clip matters more than the per-second rate, so measure both.

Which of them generates 4K video?

Of the three, only Seedance 2.5 lists a 4K resolution on WaveSpeedAI. WAN 3.0 stops at 1080p and Kling 4.0 offers 720p and 1080p.

How long can the clips be?

All three go up to 30 seconds. Seedance 2.5 accepts 4 to 30 seconds, WAN 3.0 accepts 2 to 30 seconds and Kling 4.0 accepts 3 to 30 seconds.

Do all three generate audio?

Yes. Seedance 2.5 and WAN 3.0 have a generate_audio setting that is on by default, and Kling 4.0 generates synchronized stereo audio together with the picture.

Do they all have video edit and video extend?

Yes. Seedance 2.5, WAN 3.0 and Kling 4.0 each have dedicated video-edit and video-extend endpoints on WaveSpeedAI. The source-length limits and price basis differ, see the sections below.

Can I use all three through one API?

Yes. They share the WaveSpeedAI submit-and-poll request pattern, so switching between them is a model path change plus a few model-specific parameters.