Seedance is ByteDance’s video model. You can reach it at least seven different ways, and the same eight-second clip costs anywhere from $0.24 to $2.88 depending which door you walk through.
That is a twelve-fold spread on identical output, and the expensive doors are the ones most people use.
Read the config, not just the rate
Aggregators advertise Seedance 2.5 “from $0.085 per second”. That is a real price for a real
configuration — and the word doing the work is from.
The floor buys you 480p, no reference images, no audio. Plenty of work needs
nothing more than that: establishing shots, background motion, anything textural.
The moment you want a specific character to appear, or a mouth to move in time with a line, you
are buying a different product. Running 720p, reference-to-video, with audio attached
— the config a persona-driven piece actually needs — we measured roughly 20.1 credits per
second, about $0.295/s.
$0.085/s for footage. About $0.295/s for a performance. Same model,
same provider, roughly 3.5× apart — because references and synced audio are features, and features
cost.
We measured the combination rather than isolating each variable, so we cannot tell you what the
resolution bump costs versus the references versus the audio. What we can tell you is not to plan a
budget from a floor price you will not be using.
A fifteen-second talking shot lands around $3 at that rate, which matches what
working creators in this format report paying. If your storyboard is mostly talking heads, price the
whole thing at the performance rate. If it is mostly world and atmosphere, the floor is genuinely
available to you and the cheap tiers below make it cheaper still.
Every route, priced
| Route | Type | Price | Notes |
|---|---|---|---|
| BytePlus ModelArk | Official ByteDance cloud | ~$0.03/s (2.0 mini) ~$0.20/s (2.5) |
The genuine bargain. Measured $0.18 for 6 seconds. Needs per-model console activation, and parameters ride inside the prompt text. |
| Replicate | API | $0.10–0.43/s | Roughly parity with the aggregators at the likely tier; the tier structure is ambiguous. |
| Atlas Cloud | API | $0.134/s flat | Same rate across all variants, ~55% under EvoLink. Quoted, not tested — audio support unverified. |
| EvoLink | Aggregator, credits | advertised $0.085/s measured $0.295/s |
The 3.5× gap above. Result URLs expire in 24 hours and references must be public URLs, so you need your own asset host. |
| fal.ai | API | ~$0.473/s | ~60% more than EvoLink on the same config. Roughly $18 per finished minute of raw footage. |
| Dreamina / CapCut | Consumer subscription | ~$1.06–1.29 per 8s | Seedance 2.5 is subscriber-gated and costs more per second than API routes at every tier. 1080p runs ~2.5× the 720p per-second rate. |
| Higgsfield | UI subscription | $1.55 per 8s | The most expensive doors. You are paying for the interface, the presets and not having to build a pipeline — a real product, just not a cheap one. |
| Magnific | UI subscription | $1.86 per 8s | |
| Runway | UI subscription | $2.88 per 8s |
Same model, same eight seconds: $0.24 at the mini tier, $2.88 at Runway.
What that means per episode
Take a fourteen-minute video assembled from 5–8 second clips, the way this format is actually made.
- Talking shots need 720p, references and audio — roughly $3 per fifteen seconds on an aggregator, and there is no cheap version of this.
- World and crowd b-roll carries no recurring face and can run on the mini tier at three cents a second.
- Retakes are the real budget. Working creators in this format report 10–15 retakes per scene. Any per-clip figure multiplied by your shot count is wrong by that factor.
An honest planning number with retakes included is $150–300 per long episode. The creator who popularised the format states £300–800. Our own batch came in at $115 for 44 clips with no retake loop at all — which is exactly why it was not good enough.
The lesson is routing, not model choice: put face-carrying talking shots on the expensive precise route, everything else on the cheap tier, and budget retakes as the main line item rather than a contingency.
The cost nobody quotes: there is no gallery
Per-second rates make API routes look obviously cheaper. They leave out the part of the product you only notice when it is missing.
Generate through an API and the files come back to you. What does not come back is anywhere to look at them. No gallery, no side-by-side, no “show me the forty clips I made this afternoon so I can pick four”. You have a folder, and reviewing generative output means comparing many near-identical attempts — which a folder is a bad tool for.
You can build the missing layer. It is also a small software project sitting between you and your first video. So the honest comparison is not six cents against $1.55 a clip, it is:
Cheap seconds plus a pipeline you build and maintain, against expensive seconds with the pipeline included.
At volume the first wins by a distance. At low volume it does not come close.
Hidden costs that only appear in production
One to check before you subscribe to anything, because it is on no pricing page:
does what you generate through the API appear in the web interface? On some
platforms the two sides are billed separately and do not share a library, so you pay twice and can
only see half of it. Where it works properly — ElevenLabs
is the reference example — a generation lands in the same place whether you clicked or called,
and you can prototype in the interface then scale through the endpoint without losing anything.
- Expiring outputs. Some aggregator result URLs die within 24 hours. You need a permanent host of your own — both to archive, and because reference images must be publicly reachable URLs.
- Credit units differ 4× between models on the same provider. Print the reserved credits on every submit and abort above a ceiling. An $81 estimate became a $115 bill because the shot count grew and nobody re-priced it.
- Lost job IDs are lost money. Not every provider offers a “list my tasks” endpoint. Persist the ID at submit time — we lost $1.60 learning that.
- Moderation differs by provider, and it is a routing decision. The official ByteDance cloud rejected a photoreal recurring persona face as private information; another route accepted the identical image. Populated wide shots passed everywhere. If you are building around a consistent character, test this before committing.
- Silent failure. One take performed four words of a forty-word line and returned successfully. Transcript-check every talking take automatically; a 200 response is not proof of a usable clip.
Where you are changes what you can use
The consumer routes rolled out market by market and remain uneven. Seedance in CapCut expanded through 2026 to the US, Japan, Europe, Africa, South America and the Middle East — but “Europe” is not a single switch, and which models a given market gets differs from what the marketing describes.
Precision matters here because it trips people up: CapCut is the editor, Dreamina is the generator. If you open CapCut expecting Seedance and the option is absent, you have not missed a setting — it depends where you are. Dreamina directly is the workaround.
API routes do not have this problem. An endpoint behaves the same wherever you call it from, which is an underrated argument for going programmatic beyond the price.
What “free Seedance” actually means
There is no free tier of a frontier video model, because inference costs real money on every generation. What exists is trial credits, promotional windows on platforms buying users, and generation bundled into a subscription you may already hold.
If the requirement is genuinely high volume at near-zero cost, the answer is not a free route — it is the mini tier at around three cents a second, which is close enough to free for most purposes and does not run out.
Method and dates
Prices observed June–August 2026 across roughly $180 of real generation spend on a production project. Figures described as measured were metered by us in production; the rest are provider-quoted. Atlas Cloud’s rate is quoted rather than tested, and one credit-to-dollar conversion is an estimate rather than a published rate.
This market reprices monthly. Treat any undated Seedance pricing article — including this one, six months from now — as fiction.
Part of our AI video tools comparison — three categories, priced, and which one you are actually shopping in.
Confused about which ByteDance product is which? Dreamina, Jimeng, CapCut, Pippit and BytePlus, sorted — they run a Chinese product and an international twin for nearly everything.
