Seedance 2.5 Pricing and the 'AI UGC for $0' Claim

"i made this AI UGC for almost $0." That was the whole pitch(opens in new tab) on 2026-09-22, over a clip made with Seedance 2.5: download a photo from Pinterest, feed it to the model as a reference, swap one thing for another. "damn you don't need prompts anymore."

Half of that is true. What this post won't do is tell you what it costs us, because we haven't measured it. What we can do is read ByteDance's own billing rules, and those count something the $0 leaves out: every take you throw away.

What Seedance 2.5 actually is

Seedance 2.5 is ByteDance's video model that generates picture and sound together in one pass, up to 30 seconds per generation. The Seed team launched it on 2026-07-31(opens in new tab), describing it as built on "the unified multimodal audio-video joint-generation architecture of Seedance 2.0" and able to "generate high-quality, 30-second audio-video clips in a single pass."

Three things from that announcement matter for UGC:

References instead of prompts. "Users can now input up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single pass." That's the "replace X with Y using references" step from the post.
Voices come with the picture. The model "can preserve the appearances and voices of multiple characters." Dreamina, ByteDance's consumer app, lists "Lip sync audio generations" in its plans for the model(opens in new tab).
Editing after the fact. Timestamp-level edits to "audio and video content", which ByteDance frames as "reducing the cost of repetitive generation." So the vendor itself counts retakes as a cost.

The API price is public. On BytePlus ModelArk's pricing page(opens in new tab), read on 2026-09-23, the model is billed in tokens, and ByteDance converts its own 5-second examples into per-second prices.

ByteDance's own worked examples: a 5-second 16:9 clip lists at $0.514 (480p), $1.156 (720p) and $2.843 (1080p). These are list prices, not our measurements.

The part of the $0 that is true

For a single take, the price really did collapse, and one whole step disappeared. That deserves saying before any criticism.

A talking-head AI video used to be a chain: a voice from one model, a face animated to that voice by another, then a render. Each link was a separate bill and a separate place to fail. Seedance 2.5 generates the speech and the moving face in the same pass, so the voice and lip-sync steps aren't separate purchases here.

And the take itself can cost nothing to you. Dreamina's FAQ says it "provides free daily credits to test out Seedance 2.5 before you commit to a subscription plan." Make one clip inside that allowance, keep it, and "almost $0" is an accurate description of what you paid.

It's an accurate description of one take, though. The reviewers quoted below didn't ship their first one.

What "almost $0" counts, and what it doesn't

The $0 counts one kept take inside a free allowance. It doesn't count the takes you rejected, the script, or the face.

The claim is true for the left column. Every item on the right has a cost that shows up somewhere else.

Retakes are billed like keepers

On the API, a take you throw away costs the same as one you keep. BytePlus spells out the billing rule: "You are only charged for successfully generated videos. No fee is charged if generation fails due to reasons such as content moderation."

"Failed" there means the model refused or broke. A take where the actor's accent is wrong, or the product morphs mid-sentence, succeeded as far as billing is concerned. You decide it failed. The invoice doesn't.

We don't have our own number for how many takes a kept Seedance clip costs, so here is a documented one. MindStudio made a short film almost entirely in its 30-second mode(opens in new tab) (published 2026-08-06) and reported "about 15 generations at 30 seconds each, totaling roughly 450 seconds of raw footage for a final 2-minute-22-second cut, a 3.2-to-1 shoot ratio," at an estimated $90 to $355 depending on API or platform rates.

About 3.2 seconds generated for every second that made the cut. At a flat per-second rate, that is the multiplier on your bill.

A short film isn't a UGC ad, and a practiced creator will beat that ratio. But the direction doesn't change. A second MindStudio test(opens in new tab) (2026-08-12) found that "a multi-shot intro requires a dozen or more reroll attempts to land on a usable output." ByteDance's own answer is partly in the product: draft videos at a fixed 480p(opens in new tab), billed at the 480p rate, so you can throw away cheaper takes before you commit to the expensive one.

So the useful number is never "price per take." It is cost per usable clip: everything you were billed, divided by the seconds you kept.

What a reference photo doesn't replace

References replaced the prompt. They didn't replace the script, the face, or the voice you meant to get.

The script. The video that went viral the same day, @maxxmalist's "The race for AGI"(opens in new tab), had 14,937 likes and about 1.8 million views when we read it on 2026-09-23. Its caption credits two tools: "Script: Sherpa by Pocket FM / Video: Seedance 2.5." The words came from somewhere else. "No prompts" isn't "no writing." Deciding what the person on screen says is still the part that makes an ad sell or not.

The face. Step one of the $0 recipe is "download a pic from pinterest." A pin is someone's photo, and possibly someone's face. Pinterest's own terms(opens in new tab) say that "any use of Pinterest not expressly permitted by these Terms is a breach of these Terms and may violate copyright, trademark, and other laws." Using a stranger's face in an ad costs nothing until someone recognises it.

The voice. Voices come with the picture, which also means the model picks them. In one of MindStudio's tests, a character got a British accent nobody asked for; prompting "American" fixed it, and "American English" didn't. Each fix like that is another take, and the section above covers what takes cost.

What we measured, and what we haven't

We have no measurement of Seedance 2.5. The only place we've seen it in a product we track is a "SEEDANCE 2.5 LATEST" badge on Arcads' pricing page, on 2026-09-15. At ViralFaceless(opens in new tab) we're evaluating Seedance as a second clip option on the more expensive tier. It's still an evaluation, so there are no numbers from it here.

What we have measured is a different model doing a different job. We build Clipwright(opens in new tab), which turns a script into a talking actor, and one step in it is HeyGen's Avatar IV. On 2026-09-10 we checked that step's list price against the change in our wallet: $0.0374 a second effective against $0.0385 listed, within 3%. In the one step we measured, the price list was honest about the price of a second.

The price list doesn't tell you how many takes you'll need. We learned how far a bill can drift from the output when we charged 900 credits for a video that didn't exist, and our rule since is that the bill is authorised by delivery, not by spend. That rule has a limit: a take you reject was still delivered, so billing by delivery, ours included, charges for it.

Before you believe the next $0

You can check any "AI UGC for $0" claim, including your own, in about ten minutes:

Count every take. Open your generation history for the last clip you shipped and count the takes, not the keepers.
Divide. Credits or dollars spent, divided by seconds you actually used. That's your cost per usable clip. For credit-based apps, find out what a credit buys first.
Draft cheap. If your tool offers a draft or low-resolution mode, lock the take there and render the keeper once at full resolution.
Own the face and the words. Use a photo you have rights to, and write the script before you touch the model.

If the $0 survives those four steps, it's real, and you should keep doing exactly what you're doing.

FAQ

How much does Seedance 2.5 cost?

On BytePlus ModelArk, read 2026-09-23, it lists at $10.70 per million tokens for 480p and 720p without video input. ByteDance's own examples work out to $0.103 a second at 480p, $0.231 at 720p and $0.569 at 1080p. Adding a reference video lowers the token rate but adds the reference's length to the billed duration. Consumer apps like Dreamina sell credits instead and give free daily credits to try it.

Does Seedance 2.5 do lip-sync and voice?

It generates audio and video together. ByteDance says it can "preserve the appearances and voices of multiple characters," and Dreamina's plans list lip-sync audio generations. Reviewers report the voice is inferred from references unless you direct it.

Is AI UGC really almost free now?

One take can be, inside a free allowance. A usable clip usually takes several takes, and on ByteDance's API every take that renders is billed, whether you keep it or not.

Sources

ByteDance Seed, "One-take Creation, Flexible Referencing: Introducing Seedance 2.5", 2026-07-31: https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5(opens in new tab)
BytePlus ModelArk pricing (Seedance 2.5 rows, billing notes, draft mode), read 2026-09-23: https://docs.byteplus.com/en/docs/ModelArk/1099320(opens in new tab)
Dreamina, Seedance 2.5 page (free daily credits, plans), read 2026-09-23: https://dreamina.capcut.com/seedance/seedance-2-5(opens in new tab)
MindStudio, "Seedance 2.5 Review: 30-Second Clips, Voice Casting, and Morphing Bugs", 2026-08-06: https://www.mindstudio.ai/blog/seedance-2-5-review-guide(opens in new tab)
MindStudio, "Seedance 2.5 Review: Video Quality Gains, But at 23 Cents a Second", 2026-08-12: https://www.mindstudio.ai/blog/seedance-2-5-review-pricing(opens in new tab)
@simonecanciello on X, 2026-09-22: https://x.com/simonecanciello/status/2102425012225401329(opens in new tab)
@maxxmalist on X, 2026-09-22: https://x.com/maxxmalist/status/2102429352650514714(opens in new tab)
Pinterest Terms of Service: https://policy.pinterest.com/en/terms-of-service(opens in new tab)

AI UGC for $0? Seedance 2.5 Still Bills the Retakes