Seedance 2.5: Modes, Limits and Price per Second

Seedance 2.5 is ByteDance’s video model from July 31, 2026. It makes up to 30 seconds of video with sound in one pass, takes up to 50 references at once, and can edit or extend a clip you already have. Through an API it costs roughly $0.10 to $1.14 a second depending on the channel and resolution, and there is no free API tier.

This guide covers the modes, the limits that decide whether a job is accepted, the price on each channel we could check, and what changed from Seedance 2.0.

What Seedance 2.5 is

ByteDance calls it an audio-video joint generation model: picture and sound come out of the same pass. Its own page says it generates “up to 30 seconds in a single generation” with the option to extend twice, reads reference videos more precisely (framing, intent, camera language), and supports white-model control and green-screen editing. TechNode’s launch report adds the reference counts: up to 30 images, 10 videos and 10 audio clips in one request.

At launch it rolled out in ByteDance’s own apps, with the API “expected soon”. It is now on BytePlus ModelArk as dreamina-seedance-2-5-260628, on Higgsfield’s API, as a Comfy partner node, and in Dreamina.

The modes

Higgsfield’s API and Comfy’s workflow guide list the same jobs under slightly different names:

• Text to video. A prompt in, a clip of 4 to 30 seconds out, with sound.
• Image to video. One image animated into a clip. Higgsfield’s endpoint also takes an end image, and Comfy has a separate first-last-frame workflow that locks both endpoints.
• Reference to video. Images, video clips or audio steer the character, style and motion rhythm, with lip-synced dialogue.
• Video edit. You send a clip and describe the change: replace an object, swap the background, use a green screen. This is the mode character-swap tools build on.
• Video extend. Continues a clip forward or backward. In Comfy you connect the clip as the reference video and set task_type to extend; the output is only the new segment.
• Draft mode. Renders a 480p preview, then turns that same draft into a 1080p final. Comfy says a draft stays renderable for 7 days. It works for text, first-last-frame and reference jobs.

Limits

Limit	Value	Source
Output length	4–30 s per generation	Higgsfield API pages
Resolutions	480p, 720p, 1080p	Higgsfield, Comfy
References per request	30 images, 10 videos, 10 audio clips	Comfy, TechNode
Reference videos	at least 2 s each, 30 s in total	Comfy pricing notes
Extension	twice (ByteDance) or “multiple rounds” (Comfy)	ByteDance, Comfy
Draft lifetime	7 days	Comfy draft docs

Frame rate on the rate cards is 24 fps, which is what the token formulas below assume. Dreamina’s Seedance 2.5 pages advertise “30-second videos from up to 50 references”, the same 30 + 10 + 10.

What Seedance 2.5 costs

Every channel bills by tokens, and tokens grow with width × height × seconds × 24 fps. If you send a video, its seconds count too, at a lower rate per token. So the per-second figures below are for a 16:9 clip without a video input.

Channel	480p	720p	1080p	Basis
Higgsfield API	$0.2056	$0.4622	$1.1372	List price per second, checked October 15, 2026
Comfy partner node	about $0.147	about $0.331	about $0.813	Comfy’s 5 s estimate (155.1 / 348.8 / 857.8 credits) at 211 credits = $1
BytePlus ModelArk	about $0.103	about $0.231	about $0.569	BytePlus’s 5 s examples ($0.514 / $1.156 / $2.843)

Higgsfield’s numbers come from Jutsu Studio’s Higgsfield price table and are rates before any discount. Comfy prices in credits (3.229 per 1,000 tokens at 480p and 720p, 3.530 at 1080p) and says the API’s actual usage is what you pay. BytePlus lists $10.70 per million tokens at 480p and 720p and $11.70 at 1080p, and charges only for videos that generate successfully.

Editing a clip costs more, because the input seconds are billed as well. For a 10-second edit of a 10-second clip:

Channel	480p	720p	1080p
Higgsfield API (video edit)	$2.47	$5.55	$13.65
Comfy partner node (Swapjutsu’s measured figures)	$1.71	$3.96	about $9.75

BytePlus doesn’t give a 10-second edit example. Its 5-second examples with a 2 to 30 second video input run from $0.553 to $2.152 at 480p and $3.062 to $11.907 at 1080p, and a minimum token charge applies whenever the input includes video. Extending a clip on Higgsfield is listed at $0.12336 a second at 480p and $0.68232 at 1080p.

Dreamina, ByteDance’s consumer app, sells Seedance 2.5 on its own credit plans. Its Seedance 2.5 pages don’t publish a credit cost per clip, and they describe free access as a trial for “eligible users” whose amounts “can vary”. Nothing we checked offers free Seedance 2.5 through an API, and Comfy states that partner nodes have no free usage.

Seedance 2.0 vs 2.5

Only differences that ByteDance’s channels publish:

	Seedance 2.0	Seedance 2.5
Longest single clip	15 s	30 s
Output	480p to 4K	480p to 1080p
Variants	Standard, Fast, Mini	Standard, Draft (480p preview)
Video input length on BytePlus	2–15 s	2–30 s
BytePlus, 480p/720p, no video input	$7.00 per million tokens	$10.70 per million tokens
5 s at 480p on BytePlus	$0.35	$0.514
Comfy credits per 1,000 tokens, 480p/720p	2.112	3.229

So 2.5 is the longer, more referenced and more expensive model, and 2.0 is still the only one of the two with 4K. ByteDance hasn’t published a side-by-side quality comparison, and third-party benchmark claims vary, so we leave quality out. If a 15-second clip at 720p is enough, 2.0 Fast or Mini costs a fraction of 2.5.

Picking a channel

• Lowest list price per token: BytePlus, if you can create the endpoint in your region and are happy calling a raw API.
• No coding, inside a node graph: Comfy’s partner node, with prepaid credits. See Seedance in ComfyUI for setup and cost per clip.
• One account for many models: Higgsfield’s API, which also runs Genjutsu, Kling and Wan at the rates on the price table.

Jutsu Studio is a free, open-source Mac app for Higgsfield. It runs Seedance 2.5 through your Higgsfield account or API key, shows the price on the Generate button before you send, and saves every result with its prompt and cost to your disk. If the job is replacing a person in a clip, Swapjutsu builds the control video on your Mac and renders it with Seedance 2.5 through Comfy.

For the prompt itself, the open-source Seedance 2.5 prompt skill lets Claude Code, Codex or Cursor write it from your idea, photos or reference clip.