Seedance 2.0 API: Endpoints, Real Pricing and How to Call It

The Seedance 2.0 API is ByteDance's reference-to-video model resold by several hosts, each doing the per-second maths differently. Here is the schema and the bill.

Try Synexa → Official site

Two models wearing one name

Seedance 2.0 is a ByteDance video model, and the hosts that resell it expose two variants: Seedance 2.0 and Seedance 2.0 Fast. Kie.ai publishes the practical difference as generation time, roughly five minutes per generation for the standard model and roughly four for Fast, which is a smaller gap than the naming implies. OpenRouter's listing dates the model to April 15, 2026 and describes what it does rather than how fast it does it: text-to-video, image-to-video with first and last frame control, and multimodal reference-to-video. The strength both listings point at is the same one, preserving character identity, visual style and camera movement from material you supply. If your use case is a single character appearing across many shots, that is the reason to look at this model rather than a general text-to-video one.

The parameter list, in full

The kie.ai endpoint accepts twelve fields, and most of them are reference slots. Frames are handled by first_frame_url and last_frame_url. The prompt field takes up to 20,000 characters. Then come three arrays: reference_image_urls accepts up to nine images at 30MB each in JPEG, PNG, WebP, GIF or BMP; reference_video_urls accepts up to three videos at 50MB each with a combined length of no more than fifteen seconds; reference_audio_urls accepts up to three MPEG or WAV files at 15MB each, also capped at fifteen seconds total. The remaining switches are generate_audio, resolution at 480p or 720p, aspect_ratio from 16:9, 4:3, 1:1, 3:4, 9:16 and 21:9, a numeric duration, web_search, and nsfw_checker, which defaults to true in the playground and is not guaranteed to catch everything.

Why there are two prices for the same resolution

Kie.ai lists 480p at 6.8 credits per second, about $0.034, when you pass a video input, and 11.7 credits per second, about $0.059, when you do not. At 720p the same split runs 15 credits per second, about $0.075, with video input against 24.8 credits per second, about $0.124, without. That looks backwards until you read the formula: without a video input the charge is price multiplied by output, and with one it is price multiplied by input plus output, so the lower unit rate is applied to a larger quantity. Kie.ai also notes that high-tier top-ups carry a ten percent bonus, making effective pricing roughly ten percent below the listed figures, and flags the whole schedule as beta and subject to adjustment. A limited-time discount on Seedance 2.0 Fast was posted as running until October 7 at 06:00 UTC.

What the other listing charges

OpenRouter prices the same model differently and is worth checking against before you commit. Its model page advertises pricing from $0.06726 per second, while the provider table underneath, which shows a single upstream provider named Seed, lists $7.00 per million for video and audio and $0.1512 per second. The gap between the headline and the table is the usual difference between a floor price and a specific provider's rate, so read the table rather than the banner. OpenRouter also documents how tokens are counted for this model: output height multiplied by width, multiplied by duration, multiplied by 24, divided by 1024. That formula is the one to put in your cost estimator, because it makes resolution and duration compound rather than add.

Latency and uptime you should design around

This is not an API you call inside a request handler. OpenRouter measured end-to-end latency at 123.9 seconds at the median for the best provider, and kie.ai quotes roughly four to five minutes per generation depending on the variant. Over a three-day window OpenRouter recorded 100 percent uptime, meaning at least one provider was answering, but 93.84 percent availability, meaning some requests returned errors or empty responses. With a single upstream provider there is no failover to route around a bad patch, so build the retry yourself, queue the job, and return a job id to your user rather than holding a connection open. If you would rather not maintain one integration per model, Synexa puts FLUX, video and audio models behind one REST endpoint and a Python SDK with pay-per-run billing.

What you actually get

Seedance 2.0 hosts against a multi-model endpoint

From zero to a first generation

1. Pick the variant first
   Seedance 2.0 Fast trades roughly a minute of generation time. Decide before you write the client, because the model slug differs and swapping it later means touching your queue code.
2. Upload references, then prompt
   Put images, video and audio behind public URLs inside the documented size limits, then write a prompt that names them. Reference material is where this model earns its cost.
3. Set resolution knowing the rate
   720p costs roughly double 480p per second on kie.ai, and passing a reference video changes the formula from output-only to input plus output. Price a test batch before scaling up.
4. Queue it, do not await it
   With median latency around two minutes and availability under 94 percent over one measured window, store a job id, poll it, and retry failures rather than blocking a user request.

FAQ