Seedance 2.5 4K Is Here for Film Delivery — and the API You Call Still Stops at 1080p

Volcano Engine, ByteDance's cloud arm, began putting Seedance 2.5 4K into the hands of film and television customers on 23 October 2026. It is a native 4K delivery tier for Seedance 2.5 — the video model family that has been generally available since 31 July 2026 — and it carries 10-bit color plus a two-stage Draft workflow that previews a shot at 480p before committing to a 4K master. The detail that matters more than any of those specs is the one buried in the availability line: the Seedance 2.5 endpoint most developers can actually call still lists 480p, 720p and 1080p, and no 4K bracket at all. Two products now share one name, and choosing between them is the decision this update is about.

What ByteDance actually shipped on 23 October

The release is the top tier of the Seedance 2.5 generation rather than a new generation. ByteDance says it trained this version specifically on facial expressions, subtle movement and fine visual detail, and positions it for big-screen delivery — the kind of work where a visible gradient artifact or a soft edge on small on-set text gets a shot rejected. The company also says the tier preserves the color, quality and consistency of reference material better than the standard model, and that small text on posters and packaging survives generation more legibly. Those are vendor claims; no independent test of the 4K tier has been published, and the coverage so far traces back to ByteDance's own announcement.

Availability is narrow. Volcano Engine describes the tier as open to trials by selected enterprise customers, with some Chinese and overseas film and post-production houses already testing it, and says broader access will be expanded gradually. No general-availability date, no API date and no price were announced. If you are not already in that trial conversation, Seedance 2.5 4K is not something you can buy today.

It is also worth being precise about what is new, because 4K has been in this model's marketing since the beginning. ByteDance unveiled Seedance 2.5 at its FORCE conference on 23 June 2026 as a 30-second, native 4K model accepting up to 50 multimodal references, and 10-bit color was part of that pitch. What shipped to developers in July and August was not that: the callable API topped out at 1080p. The October announcement is the first time Volcano Engine has packaged the 4K capability as a named, film-facing tier with a stated 10-bit delivery spec and an announced trial programme — but it remains a tier you are invited into rather than one you can sign up for, and the API still has not caught up with the June pitch.

10-bit color is the part that changes a delivery, not the pixel count

Going from 8-bit to 10-bit color takes each channel from 256 steps to 1,024, which is four times the gradation. In practice that buys two things on a film pipeline: smoother ramps where a scene has large, slow gradients — skies, fog, practical light falloff, dark interiors — and more headroom to grade without the image falling apart into banding or posterized shadows. ByteDance frames it as richer color layering and smoother transitions, which is the same claim in vendor language.

The practical caveat is that 10-bit only helps if the rest of the chain carries it. A 10-bit master squeezed into an 8-bit H.264 deliverable gives most of the benefit back, and the value shows up in the grade and the master, not in the social cut. ByteDance has not published the color primaries, transfer function or container details for the 4K tier, so whether the 10-bit output is genuinely end-to-end rather than a deeper container around an 8-bit-ish image is an open question. Treat the gradation improvement as vendor-stated until someone outside the trial measures it.

Draft mode: cheap previews, then a real render — not an upscale

The Draft workflow is the more useful half of this release, and it is the part most secondhand summaries get wrong. The shorthand going around — generate a cheap 480p preview, then upscale it — describes the cost shape correctly and the mechanism incorrectly. There is no super-resolution pass sitting between the preview and the master. The final is the same shot re-rendered at the target specification, with the prompt, reference assets, duration, aspect ratio and seed carried over from the draft, which is why composition and motion survive while fine texture, small text and crowd counts can shift slightly between the two.

The mechanics are documented in ByteDance's API. A draft request sets draft: true with resolution fixed at 480p — 720p and 1080p are rejected as draft resolutions. That returns a draft task ID. The final request then contains exactly one draft_task object and nothing else: no prompt, no reference images, no duration, no seed, because all of those are inherited and resending them is an error. The final currently renders at 1080p only through the public API. The draft stays renderable for seven days from its creation, and only from the same account that made it. ByteDance bills the draft and the final as two separate tasks. This is the machinery the 4K tier extends, with the final landing at 4K instead of 1080p.

Whether that saves money is arithmetic, not marketing. At BytePlus ModelArk list rates a 480p render runs about $0.10 per second and a 1080p render about $0.57 per second, so one draft plus one final costs roughly 1.18 times a single 1080p attempt. Draft mode therefore pays for itself only from the second take onward — precisely the situation it was built for. ByteDance's own worked examples put the saving at about 57% on a shot that needs four takes and up to 77% on one that needs twenty, and those figures check out against the list prices. If you tend to land a shot first try, Draft mode costs you about 18% more.

The spec sheet, with the tiers kept apart

• Max clip length — 30 seconds in a single pass, with a long-video mode reaching roughly three minutes in the standard tier.

• Public API resolution — 480p, 720p and 1080p on dreamina-seedance-2-5-260628; no 4K bracket is listed in BytePlus ModelArk's capability or pricing tables.

• 4K tier resolution — native 4K, 10-bit color, selected enterprise trials only.

• Color depth — 10-bit on the 4K tier; no 10-bit option is documented at 1080p.

• References — up to 50 multimodal inputs, including images, video, audio, 3D whitebox models and greenscreen plates.

• Draft mode — 480p previews, seven-day validity, final re-rendered at 1080p today and at 4K on the new tier.

• Weights — closed. No checkpoint, no licence, no repository.

• Price — $11.70 per million video tokens at 1080p, about $0.57 per second for 16:9 at 24fps; $10.70 per million at 480p and 720p, about $0.10 per second at 480p. No 4K price exists yet.

What the 4K tier does not change

Everything that made Seedance 2.5 awkward to plan around in July is still true. It is closed-weight, so there is no self-hosting path and no way to fine-tune on your own footage. Access runs through ByteDance's own surfaces — the Dreamina platform for hands-on work and BytePlus ModelArk for the API — and promotional credit pricing has moved around enough this year that any budget should be re-checked against the current list rather than carried forward from a launch post. Resolution is now split across two tiers with different gates, which makes "Seedance 2.5" a less precise phrase than it used to be: ask which tier before anyone quotes you a spec.

The 4K tier also does not settle the quality question. ByteDance publishes no benchmark for this model, and we found no published human-preference arena score for the 4K tier. Claims that it looks better than a rival at 4K are unfalsifiable right now, in either direction.

Who should act on this, and who should wait

If you deliver film or high-end episodic work and you already have a ByteDance commercial relationship, the trial is worth asking about — a 4K master with 10-bit color and a 480p iteration loop is aimed exactly at your bottleneck, and the cost of finding out is a conversation rather than a contract. But do not restructure a delivery pipeline on it. There is no price, no API date and no general availability, and a tier in invited trials can stay in invited trials for a long time.

If you make advertising, e-commerce or social video, 4K is close to irrelevant: the platforms you deliver to will recompress it anyway. The change that matters to you landed in late September, when Draft mode arrived at 1080p, and it is available now. That is the version of this model with a measurable effect on your per-approved-shot cost.

Where this fits in a multi-model stack

Seedance 2.5 is not in OrcaRouter's catalogue — we route text, multimodal and some video models, and this is not one of them, so you keep a ByteDance account and key for the generation itself. What a router is good for here is the part of the pipeline around the video model: the script treatment, the shot list, the prompt variants you will throw at Draft mode, and the review pass that decides which draft gets the 4K render. OrcaRouter puts 200+ models behind one OpenAI-compatible endpoint with 0% markup — provider list price passed through unchanged — so comparing two LLMs on the same prompt costs one key instead of two contracts, and automatic failover keeps that side of the pipeline up while you are iterating on shots.

That separation is worth keeping deliberate. The video model is the expensive, opinionated part of the stack and it will change every few months; the text layer around it should be the cheap, swappable part. Keeping the two on separate keys means a new Seedance tier is a procurement question rather than a rewrite.

What would change this answer

Two announcements would make this piece out of date, and both are worth watching if 4K delivery is on your roadmap. The first is a 4K API date with a price attached — until Volcano Engine publishes a model ID and a per-second rate for the 4K tier, no one outside the trial can budget against it, and the draft-final arithmetic that makes the 1080p tier affordable cannot be repeated at 4K. The second is an independent measurement of the 10-bit output, ideally a side-by-side grade of the same shot delivered at 8-bit and 10-bit. ByteDance's gradation claim is plausible and entirely untested by anyone outside the programme.

Until one of those lands, the planning assumption is the unglamorous one. Seedance 2.5 at 1080p with Draft mode is the version of this model that a production team can use today, and the draft loop is where the money is: roughly $0.10 per second to find out whether a shot works, against $0.57 to render it for real. The 4K tier is a genuine step for film delivery and a non-event for everyone else — which is a narrower claim than the launch coverage makes, and the accurate one.