Where to run Seedance 2.5: platforms compared

Where should you run Seedance 2.5?

Start with the resolution you need, because that's the axis platforms actually differ on and several of them quietly cap it. Seedance 2.5 supports 480p, 720p and 1080p, and some of the places selling it expose only the first two. If 720p is fine, you have a lot of good options and the cheapest ones are very cheap. If you need 1080p, the list shortens immediately.

The second thing to settle is whether you want an endpoint or a workspace. The per-second price spread from the cheapest API to the most expensive is only about 2x, which is small enough that it rarely decides anything. What differs by much more is whether the platform gives you somewhere for the takes to live, other models to compare against, and a way to run the same shot again next week.

What to know before you pick

• Seedance 2.5 does not do 4K. ByteDance's own capability table caps it at 1080p and lists 4K only for the base Seedance 2.0, yet several platforms advertise native 4K on 2.5. Where you see 4K delivered, it's an upscaler running after the fact.
• Audio is free on the token-priced platforms. It's generated in the same pass and the billing formula has no audio term, so switching it off saves you nothing. That's specific to the 2.x line: Seedance 1.5 Pro does price audio separately.
• It's slow. Measured end-to-end latency runs to several minutes for a single generation, so budget in minutes, not seconds, and don't design an interactive loop around it.
• Feeding in a reference video is metered on input duration plus output duration, so a 30-second reference with a 5-second result can cost several times a text-only take of the same length.

What is new in Seedance 2.5

ByteDance's Seed team launched Seedance 2.5 on 31 July 2026, and it reached the public developer API on BytePlus ModelArk in early August. Against Seedance 2.0 it doubles the maximum single generation from 15 seconds to 30, and takes the reference budget from 15 files to 50: up to 30 images, 10 videos and 10 audio clips. It's also the first in the family to accept an audio-only reference, where the whole 2.0 line required audio to be paired with an image or a video.

The trade is real and it's rarely mentioned. The base Seedance 2.0 generates 4K with 10-bit output and Seedance 2.5 does not, topping out at 1080p in 10-bit H.265. So 2.5 is the longer, more referenceable model and 2.0 is still the higher-resolution one, which means the upgrade isn't automatic for every job.

Worth knowing before you plan around benchmarks: Seedance 2.5 has no leaderboard standing yet. Any Elo figure you see attached to Seedance is the 2.0 entry, which as of early September sits fifth on text-to-video with audio and second on image-to-video with audio. ByteDance has published no benchmark numbers for 2.5, only qualitative claims, so the widely repeated prompt-adherence improvement figure traces to platform marketing rather than to a ByteDance source.

What separates a good place to run it?

Four things, and the first one is where most of the confusion in this category lives.

The resolution it actually exposes

A platform can carry Seedance 2.5 and still not give you all of it. 1080p is available on Melius, on ByteDance's own surfaces, on fal, Segmind and Krea, and on several of the smaller API resellers. It isn't available on OpenRouter, on Higgsfield, or on Magnific, all of which cap at 720p.

Then there's the 4K question, which is worth being blunt about because so much of the coverage is wrong. Pollo, Kie.ai, CapCut and Dreamina pages all describe native 4K on this model. The model doesn't have it. Magnific is the one platform that states the mechanism honestly: it renders at 480p and 720p, and its separate video upscaler takes the finished take to 4K afterwards. That's a legitimate way to get a 4K deliverable, and it isn't the same claim.

What a take actually costs

The token-priced platforms all bill the same underlying formula, which multiplies input and output duration by width, height and frame rate, and only successful generations are charged, so a run killed by content moderation costs nothing. That makes the comparison unusually clean: for a 5-second 720p clip with audio, ByteDance's international surface and OpenRouter both land near $1.16, Segmind is a fraction under $1.19, and fal is roughly $2.31, which is exactly double ByteDance's token rate.

So the spread is about 2x across the whole field, and it's small in absolute terms. Ten takes at 720p costs somewhere between $12 and $23 depending where you run them, which isn't usually the number that decides a production. Where cost does bite is 1080p, roughly two and a half times the 720p rate, and reference video, which adds the input clip's duration to the bill.

Whether it takes the references you have

The 50-file reference budget is the headline feature and it comes with a restriction almost nobody mentions: reference images and videos containing real human faces cannot be uploaded directly on ByteDance's own platform. The documented workarounds are reusing outputs the model itself produced, drawing from a preset virtual-portrait library, or supplying licensed footage.

That constraint propagates unevenly. Some resellers surface it as a liability notice, a couple sell relaxed-moderation tiers that require faces to be verified in an asset library first, and at least one API reseller's documentation still carries the old 2.0 audio limits of three files and 15 seconds paired with an image. If holding a specific real person across a 30-second take is the job, verify that path on your chosen platform before you plan around it.

Whether there is anywhere to work

Most of this list is an endpoint. You send a request, you get a URL back, and what happens next is your problem, which on ByteDance's own platform is more pressing than it sounds: task records are kept for seven days and the output video URL expires after 24 hours with a download cap. Fetch and re-host immediately or you'll lose the take.

The alternative is a platform that holds the work: the take, the version before it, the reference set that produced it, and the other models you compared against. That's the difference between running a model and having a process, and it's the reason the price spread matters less than it looks.

The six compared

Capabilities and prices as of September 2026, from each platform's own documentation and pricing pages. Two ByteDance promotions on 1080p rates are live and end mid-September, so the figures here are list rates where a promotion applies.

1. Melius

Melius is the Agentic OS for creative work: you describe what you want and the Melius agent Mel builds the workflow, routes each step to a model and leaves every step editable. Seedance 2.5 is one node on a canvas that also holds Veo 3.1, Kling 3 and Wan 3.0, part of a library of leading models sharing one subscription and one credit pool.

Melius has published its own usage data on this, which is the most interesting number available about Seedance adoption anywhere: over a 30-day window the Seedance family accounted for 54.6% of every video generation on the platform, up from 26.7% when it was previously measured. Nobody's routed to Seedance by default. People run it beside Kling, Veo and Wan on the same canvas and keep choosing it, which is a stronger signal than any benchmark on a model with no leaderboard standing.

What it does well

• Seedance 2.5 runs at 480p, 720p and 1080p on the same credit pool as every other model, so comparing it against Veo, Kling or Wan on one shot costs nothing extra in setup or subscriptions.
• The canvas holds the takes, the references and the versions, with per-node history and an active-version selector, so a good take does not depend on you saving a URL.
• Reference and character nodes carry a look across every clip in a set, and changing an upstream node re-runs everything below it.
• Branching is the natural way to test two reference sets against the same prompt without rebuilding the graph.
• Every canvas is real-time and collaborative by default, so a director and an editor watch the same takes arrive.
• It's drivable from Slack, the Model Context Protocol (MCP), a CLI and an API, which matters on a model this slow because you can start a batch and go elsewhere.

The canvas accepts image, video and audio uploads, so you can bring your own reference assets into the workflow.

The trade-off

The choice is about workflow, not resolution. Melius supports 1080p directly, with the takes, references and versions held together on the canvas. If you already have storage, review and orchestration in your own stack, a standalone API may fit that setup better.

The catch: you are choosing a shared creative workspace as well as model access. Its value is in comparing models and keeping the work together across a project.

Best for: teams weighing Seedance 2.5 against the rest of the field, or using it as one step inside a longer piece of work. Plans start at $18 a month as of September 2026, with current numbers on the pricing page.

2. ByteDance direct

Going to the source means one of two clouds, and they aren't interchangeable. Volcano Engine's Ark platform is the China-domestic surface: prices in yuan, Chinese-first documentation, a Beijing endpoint, and model activation gated behind a balance or savings plan of at least ¥200. BytePlus ModelArk is the international surface, with dollar pricing, English documentation and the same resolution tiers.

For a team outside China, BytePlus is the route, and it's the cheapest first-party access there is. The model IDs differ between the two clouds, sharing a suffix but not a prefix, so code written against one won't run on the other unchanged.

What it does well

• The lowest published rate anywhere, at roughly $1.16 for a 5-second 720p take with audio on BytePlus.
• Full 1080p in 10-bit H.265, plus mov output with PCM audio, which no reseller in this comparison mentions.
• The complete mode set, including the omni-modal reference path that splits by prompt intent into reference, edit and extend.
• The only published rate limits in the category: 600 requests a minute and 10 concurrent tasks for verified business accounts, 180 and 3 for individuals, counted per primary account and per model rather than per plan.
• Eleven prompt languages, and an optional watermark that's off by default.

The trade-off

Asset handling is the sharp edge. Task records last seven days and the output URL expires after 24 hours with a 100-download cap, so anything you want to keep has to be pulled down and stored by you, immediately. There's no library, no versioning and no collaboration.

The two-cloud split is also a genuine trap for teams that read the wrong documentation, and the Ark side's yuan pricing and balance threshold make it impractical for most non-China buyers regardless of whether they could register.

The catch: the cheapest and most complete access, with a 24-hour memory.

Best for: engineering teams outside China who want first-party access at the lowest rate and already have somewhere to put the files.

3. fal.ai

fal is the developer platform most teams reach for, and it carries Seedance 2.5 across three endpoints: text-to-video, image-to-video and reference-to-video. Its API schema exposes 1080p, though its own marketing copy still says up to 720p, so trust the schema. The documentation is the clearest of any reseller here, and the playground means a non-developer can at least check output before anyone writes code.

The reason to know the numbers is that fal is the expensive option. Its token rate is exactly double ByteDance's, which works out to roughly $2.31 for a 5-second 720p take and around $13.87 for 30 seconds.

What it does well

• Three separate endpoints, so the mode is explicit, not inferred from what you passed.
• Full 1080p, with per-token pricing published for each resolution tier.
• A playground alongside the API, which shortens evaluation considerably.
• Documentation states plainly that enabling audio does not change the token count.
• Wan, Kling and much else sit behind the same API, so swapping models is a parameter change.

The trade-off

You are paying twice ByteDance's rate for the same model, and unlike the Veo situation where resellers match Google's list price, here the markup is real and it's 100%. On a project of any volume that's the whole conversation.

It's also an API and nothing more: no canvas, no asset library, no versions, no collaboration. Nothing wrong with that, but it means fal solves the access problem and none of the production problem.

The catch: the best developer experience at the worst unit price.

Best for: teams already building on fal who value one integration over the cost difference.

4. Segmind

Segmind is the cheapest way to reach 1080p, at roughly $0.5876 a second, and its 720p rate of about $0.2389 is within a rounding error of ByteDance's own. It publishes per-second prices by resolution instead of making you compute tokens, which is the most usable pricing presentation in this comparison.

It carries text-to-video, image-to-video with first and last frame handles, and reference-to-video, and it states outright that audio is free.

What it does well

• Per-second prices published by resolution: about $0.1065 at 480p, $0.2389 at 720p and $0.5876 at 1080p.
• The cheapest 1080p access of any platform here.
• First and last frame handles on image-to-video, exposed as explicit parameters.
• A playground alongside the API.
• Pricing is stated in dollars per second, so a 10-second take costs what you would expect without arithmetic.

The trade-off

Segmind markets itself on running Seedance 2.5 without a moderation filter, which is its own claim rather than a documented platform feature, and it's a positioning choice worth understanding before you route client work through it. Rate limits aren't published, so throughput on a real batch is unknown until you hit it.

Like fal it's an endpoint with a playground, so there's no library, no versions and nowhere for a team to work.

The catch: the best price for 1080p, from the platform with the least published about its limits.

Best for: cost-sensitive teams who need 1080p and are handling storage and review elsewhere.

5. OpenRouter

OpenRouter passes Seedance 2.5 through at ByteDance's own token rate with no markup, which makes it the cheapest route that isn't ByteDance itself. It's also the only platform in this comparison that publishes measured performance, and those numbers are the most useful thing on this page for anyone planning a batch.

As of early September it reported a median end-to-end latency for this model of about 316 seconds, a 90th percentile around 781 and a 99th near 1,700, on roughly 90% three-day availability. Those are rolling windows that move, so treat them as an order of magnitude: this is a model you queue, not one you sit and wait for.

What it does well

• Pass-through pricing at ByteDance's rate, so no markup on top of first-party cost.
• Published latency percentiles and availability, which nobody else provides and which change how you design a job.
• Text-to-video and first and last frame modes, with the full 50-file reference budget.
• One API key across a very large number of models from different providers.
• A playground for evaluation.

The trade-off

It caps at 720p. Its own documentation confirms 480p and 720p only, so anything needing 1080p has to go elsewhere, and that rules it out for a lot of delivery work.

The published availability figure is worth reading twice as well: roughly 90% over three days means a batch will see failures, so retry logic isn't optional. And as with the other API entries, there's no workspace of any kind.

The catch: honest pricing and honest numbers, at two thirds of the resolution.

Best for: teams who want first-party economics and real latency data, working at 720p.

6. Krea

Krea is the easiest way to try Seedance 2.5 properly without deciding anything first. It carries text-to-video, image-to-video and reference-to-video with native audio at up to 1080p, in an interface built for people rather than for integrations, and its free tier includes daily credits.

It's also a genuine multi-model suite, not a Seedance page, so comparing this model against Veo variants, Vidu and the rest of the Seedance family happens in the same place.

What it does well

• Full 1080p in a UI-first product, which is a rarer combination than it sounds.
• Free daily credits, so evaluation costs nothing.
• All three main modes including reference-to-video, with audio.
• A commercial license starts on the entry paid tier.
• Nodes on the Pro tier for chaining steps into something repeatable.

The trade-off

Credits burn quickly on video, and the entry tiers include only limited video access, so the cheap plan isn't the plan that does the work. Krea's published pricing was also hard to read reliably, with figures differing between sources, so confirm the tier before committing.

Collaboration is thin below the Business tier, and there's no real asset or version management for a team, so it's a strong individual tool rather than a place a production lives.

The catch: the friendliest way in, on credits that go fast.

Best for: individuals evaluating the model, or doing occasional 1080p work without an engineering conversation.

Where Seedance 2.5 is not available

Some absences are worth stating, because a few widely-cited roundups get them wrong. Adobe Firefly carries no Seedance model at all: its video partner roster is Runway, Kling, Luma and Topaz, with no ByteDance model of any version. Novita lists no Seedance model either, despite carrying 65 other video models. Canva publishes no Seedance integration.

Several platforms do have it but publish too little to plan against: Replicate and Together both carry it without a resolution breakdown or a usable rate card, and one API reseller's 2.5 documentation still shows the 2.0 reference limits of three audio files and 15 seconds requiring a companion image, which is either a genuine platform restriction or a documentation bug and is worth one test call before you commit.

The general lesson from checking all of this is that the model moved faster than the pages describing it. Version numbers, resolution ceilings and reference limits are all being quoted from the previous release across a lot of the web, so read the platform's own current documentation instead of a comparison post, this one included.

Frequently asked questions

Run the same shot through Seedance 2.5, Veo, Kling and Wan on one canvas and one credit pool, then keep the take that works with the references that produced it. Open Melius.