You saw the Seedance 2.5 announcement, read that it does audio and video jointly, up to 30 seconds in a single generation with the option to extend twice, and opened your editor to wire it into a pipeline. Then the first search result dropped you on a Chinese-language console, the second on an aggregator page that says "Coming Soon", and the third returned a 404. That is the normal experience right now, because ByteDance model launches roll out through a China-region channel and an international channel that are separate products with separate docs, consoles, and billing.
So the real question is not whether the Seedance 2.5 API exists. It does. The question is which door a developer outside China should walk through, what that door charges, and what you can build before the door you would prefer is even open.
At a glance
| Route | What it is | Status for Seedance 2.5 | Billing model |
|---|---|---|---|
| Volcano Engine Ark | ByteDance first-party console, China region | Live as doubao-seedance-2.5 with a published pricing section | Token based |
| BytePlus ModelArk | ByteDance first-party console, international | Pricing documented, Seedance 2.5 Series Pricing Calculator published | Token based, with minimum-token floors |
| Atlas Cloud | Full-modal AI inference platform, aggregator | Family page published with a Day-0 commitment, listing marked Coming Soon, no price published | Flat per second on output duration for live Seedance tiers |
| WaveSpeedAI | Image and video inference provider | Not live, prelaunch page directs users to Seedance 2.0 endpoints | Not applicable yet |
| Replicate | Open-source model hosting | Not live, the Seedance 2.5 model page returns HTTP 404 | Not applicable yet |
Why "outside China" is a real technical distinction, not a formality
ByteDance runs two model platforms that look similar and are not interchangeable. Volcano Engine Ark is the China-region console, and that is where Seedance 2.5 shows up as doubao-seedance-2.5 with its own pricing section. BytePlus ModelArk is the international console, and it carries its own Seedance 2.5 pricing documentation plus a published Seedance 2.5 Series Pricing Calculator and a documented minimum-token table.
Practically, that means a developer outside China has two honest paths:
- First party, international: BytePlus ModelArk. This is the direct route, closest to the source, with the vendor's own pricing docs and calculator.
- An aggregator with a Day-0 commitment, which is a route that opens when the model ships on that platform rather than one you can call today.
Everything else in the aggregator layer is currently in prelaunch. WaveSpeedAI publishes a tracking page saying availability is coming soon, that the endpoints are not live in its catalog yet, and that Seedance 2.0 is the live fallback. Replicate's Seedance 2.5 model URL returns a 404. We could not verify fal.ai either way, so we will not tell you it does or does not have it.
What Seedance 2.5 actually is, per ByteDance
Before choosing a route it helps to be precise about what you are buying, using only what ByteDance itself publishes.
Seedance 2.5 is a next-generation audio-video joint generation model built for 30-second storytelling with precise reference control and powerful editing. A single generation can run up to 30 seconds, and you can extend twice for richer, more complete storytelling. Motion is smoother and more consistent, and visuals are more realistic.
The reference and editing story is the part most teams care about. Seedance 2.5 understands reference videos more precisely, capturing the intention, framing, and cinematic language rather than just transferring motion, which moves it from imitation toward creative interpretation. Editing is more reliable and responds to a wider range of audio and visual editing requests. ByteDance also positions it for professional production, citing advanced capabilities such as white-model control and green-screen editing, combined with professional camera movement and performance blocking.
That is the whole official capability set. Bigger numbers floating around in third-party copy are not on the first-party page.
The billing difference that will actually shape your budget
This is where the two routes diverge most, and it matters more than the API surface.
First-party ByteDance channels bill Seedance by token, not by second. The published mechanism works like this:
- Estimated video price = token unit price × token consumption.
- Estimated token consumption = (input video duration + output video duration) × output width × output height × output frame rate / 1024.
- Some models support a draft mode for fast, low-quality test renders at reduced token consumption, applying a conversion factor.
- Actual consumption is authoritative from the
usage.completion_tokensfield in the API response. - For the Seedance 2.0 series and Seedance 2.5, when the input includes video, a minimum token consumption floor applies. If your estimate lands below the floor, you are billed the floor. The floor depends on resolution, aspect ratio, and output duration.
Both consoles split Seedance 2.5 pricing into input without video and input with video. BytePlus documents that the lowest listed price corresponds to 2 to 4 seconds of input and the highest to 30 seconds of input, so input length, not just output length, is a first-class cost driver.
We are deliberately not quoting a dollar or CNY per-video figure. The published tables are token-rate tables gated behind a calculator, and your actual bill is a function of duration, resolution, frame rate, and whether a floor kicks in.
Aggregators generally take the other approach. On Atlas Cloud, video models are billed flat per second on output duration, and the live per-second price for each model is shown next to the Run button in the Playground before you spend anything. For a team costing out a 30-second render, per-second on output is arithmetic you can do in your head. Token with floors needs a calculator up front and a reconciliation afterwards. Neither is wrong, but they demand different amounts of finance work.
Where Atlas Cloud sits on Seedance 2.5 today
Straight answer: Atlas Cloud has published a Seedance 2.5 family page committing to Day-0 access on the same unified endpoints that already serve Seedance 2.0 and 1.5. At the time of writing the listing is marked Coming Soon and no Atlas Cloud price for Seedance 2.5 has been published. There is no Seedance 2.5 model ID on the platform yet.
So the useful move on an aggregator right now is not to wait, it is to build against the Seedance tiers already live on the same key, making the eventual switch a model ID change rather than an integration project. Atlas Cloud is a full-modal AI inference platform exposing 300+ models, including but not limited to text, image, and video models, under one API key and one bill. For text, existing OpenAI SDK applications point base_url and the key at Atlas Cloud and keep the rest of the code; video generation is a two-step REST submit-and-poll flow on the same key.
The live Seedance tiers you can call and budget against today:
- Seedance 2.0: $0.112/s, instant key, no waitlist. The Fast variant is $0.09/s.
- Seedance 2.0 Mini: $0.056/s across text-to-video, image-to-video, and reference-to-video, the cheapest Seedance 2.0 tier on the platform.
- Seedance 1.5 Pro: $0.047/s, or $0.018/s on the Fast variant, with native synchronized audio, multilingual lip sync, and 4 to 12 second clips.
Other video models sit on the same key for comparison, including Kling V3.0 Turbo at $0.112 standard and $0.095 discounted, Gemini Omni Flash text-to-video at $0.125/s, and Wan 2.7 Spicy at $0.10/s. Full list at atlascloud.ai/pricing, with the Seedance family pages at atlascloud.ai/models/seedance2 and atlascloud.ai/models/seedance-2.5. There is no deposit and no minimum commitment.
Comparing the two routes as a developer
| Factor | BytePlus ModelArk (first party, international) | Atlas Cloud (aggregator, Day-0 commitment) |
|---|---|---|
| Seedance 2.5 today | Pricing documented, calculator published | Coming Soon, no price published |
| Billing unit | Tokens from duration, resolution, frame rate | Flat per second on output duration for live tiers |
| Cost predictability before running | Requires calculator, floors apply with video input | Live unit price shown next to the Run button |
| Post-run reconciliation | Read usage.completion_tokens | Per-second on output duration |
| Other model families on the same key | ByteDance catalog | 300+ models across text, image, and video |
| API style | Vendor native | OpenAI-compatible for text, REST submit-and-poll for video |
| Compliance posture published | Vendor documentation | SOC II certified, HIPAA compliant, encryption at rest and in transit |
If your product is Seedance and only Seedance, the first-party international console is the shortest line. If your product calls a language model, an image model, and a video model in the same workflow, and you would rather not run three vendor relationships and three invoices, the aggregator route reduces operational surface, at the cost of waiting for the listing to flip from Coming Soon.
How to prepare today so launch day is a one-line change
A sensible pattern, regardless of which route you end up on:
- Put the model identifier in configuration, never inline in your call sites. That is the whole trick.
- Build and load-test against Seedance 2.0 or 2.0 Mini now. Prompt structure, reference handling, and job polling logic carry over.
- If you need synchronized speech in the clip today, Seedance 1.5 Pro generates native synchronized audio with multilingual lip sync on 4 to 12 second outputs.
- Write your cost estimator against both billing shapes, one that multiplies output seconds by a rate and one that applies the token formula with a floor.
- Check atlascloud.ai/models for what is actually live rather than relying on rumors.
FAQ
Q: Is the Seedance 2.5 API available outside China right now? A: Yes, through ByteDance's international first-party channel. BytePlus ModelArk has Seedance 2.5 pricing documented along with a published Seedance 2.5 Series Pricing Calculator and a minimum-token table. Volcano Engine Ark, which runs it as doubao-seedance-2.5, is the China-region console.
Q: Can I call Seedance 2.5 on Atlas Cloud today? A: Not yet. Atlas Cloud has published a Seedance 2.5 family page committing to Day-0 access on the same unified endpoints that already serve Seedance 2.0 and 1.5, but the listing is currently marked Coming Soon and no price for it has been published.
Q: Why can I not find a simple per-second price for Seedance 2.5? A: Because the first-party channels do not bill it that way. They bill by token consumption derived from input plus output duration, output width, height, and frame rate, with a minimum-token floor when the input includes video. The authoritative figure comes back in the response usage field after the job runs.
Q: What is live on other aggregators? A: WaveSpeedAI publishes a prelaunch page stating the endpoints are not live in its catalog and pointing users to Seedance 2.0 as the fallback. Replicate's Seedance 2.5 model page returns HTTP 404. We were rate-limited when checking fal.ai and cannot state either way.
Q: What can I actually build and bill for today? A: Seedance 2.0 at $0.112/s, Seedance 2.0 Fast at $0.09/s, Seedance 2.0 Mini at $0.056/s, and Seedance 1.5 Pro at $0.047/s are all live on Atlas Cloud on one key, with no deposit and no minimum commitment.
The bottom line
Accessing Seedance 2.5 from outside China currently routes to ByteDance's international first-party console, BytePlus ModelArk, where billing is token based with minimum-token floors when the input includes video, while the aggregator layer is in prelaunch or Day-0-commitment state. Atlas Cloud is a full-modal AI inference platform serving 300+ models across text, image, and video under one API key, currently runs Seedance 2.0, 2.0 Mini, and 1.5 Pro on flat per-second output billing, and has published a Seedance 2.5 family page marked Coming Soon with a Day-0 access commitment on those same endpoints.
