Kling 3.0 vs Seedance 2.0: inputs, limits and price

Kling 3.0 and Seedance 2.0 differ most in what can steer a clip and in how portable a request is between them. Through Sume's video API, both make 4–15 second clips from a prompt or from a first and last frame, with optional audio. Seedance 2.0 also takes image, video and audio references, six aspect ratios and three resolutions, and is priced per 1,000 video tokens; Kling 3.0 takes no references and three aspect ratios, and is priced per second, with a higher rate when audio is on.

The limits are those of kling-3 and seedance-2 on Sume's POST /v1/videos, from the Sume API reference and the catalog and pricing code, read on 2026-09-28; how the shots look is for your own prompts to settle. Kling's own API offers Kling 3.0 with other options; Kling 3.0 API lists them in a dated table.

How do Kling 3.0 and Seedance 2.0 differ?

Length, the text and frame modes, and optional audio are the same on both, so the table keeps the rows that separate them:

Which one takes reference images, video or audio?

Only Seedance 2.0. It honors input_references of type image_url, video_url and audio_url, so you can guide a clip with a product photo, a motion clip or a sound track without pinning any frame. kling-3 lists no reference types and refuses them. If a request carries both frames and references, frames win and it is treated as image-to-video. Reference-to-video API covers the reference rules.

Can I switch between them without changing my code?

Mostly. Both are bare catalog ids on the same POST /v1/videos request, so switching is the model field. Three things break when you move a Seedance request to Kling: input_references, aspect ratios other than 16:9, 9:16 and 1:1, and a 480p resolution (the kling-3 catalog lists no 480p). All three are refused with 400 unsupported_capability in current code. With a first frame, current code passes aspect_ratio to Seedance but not to Kling, so crop the image for Kling. Going the other way, the ratios, resolutions and lengths a kling-3 request can carry all appear in the seedance-2 lists.

What does a 10-second clip cost on each?

Kling's meter ignores size; Seedance counts video tokens, so its price grows with resolution and changes with aspect ratio. In current code kling-3 does not pass resolution on to the model, so don't compare them on resolution alone. Amounts are reserved on submit, plus a 5.5% agent fee by default; kling-3 bills the audio rate unless you send generate_audio: false.

What about Seedance 2.5 vs Kling 3.0?

seedance-2.5 takes the same inputs as seedance-2 but runs 4–30 seconds, twice Kling's 15-second cap, at a higher token rate. Seedance 2.5 API lists every Seedance id.

Which should I pick?

• Pick Seedance 2.0 if you need references, a 21:9, 4:3 or 3:4 frame, or a 480p option.
• Pick Kling 3.0 if you want a per-second price that doesn't change with size, and a silent option at a lower rate.
• Build on the Seedance request shape if you expect to move between the two, since the refusals above run in the Seedance-to-Kling direction.