Kling 4 and Seedance 4K: Can You Use Their APIs Yet?

As of October 26, 2026, neither Kling 4.0 nor Seedance 2.5 4K can be called through a public API. Kuaishou's developer pricing page states that Kling 4.0 API access is not yet available, that the listed unit rates are for preview purposes only, and that the launch date and pricing will be announced separately. BytePlus's October 23 announcement of Seedance 2.5 4K says the model is available to selected enterprise customers and asks developers to wait for news on API availability. Both are real releases; neither is a line item a pipeline can budget for today.

That distinction matters because video-generation headlines travel faster than endpoints. A creative team reads that a model is launched, a product manager asks engineering to wire it in, and the engineer discovers the model identifier does not exist on any API. We saw the same gap when Seedance 2.5 itself launched on consumer apps in July with its API still to come. The useful question is not which model is best; it is which door is open, what it costs per second at each resolution, and what a build should do while the other doors stay shut.

This post compares four options from their vendors' own pages: the two headline models, the callable Seedance 2.5 API on BytePlus ModelArk, and xAI's Grok Imagine Video 1.5, with Kling 3.0's callable rates as the fallback for anyone who needs 4K from Kling now. It prices one hypothetical clip across all of them, then turns to what a pipeline should assume about preview prices, enterprise trials and model identifiers that do not exist yet.

The table sorts the four by one question: can a developer with a credit card send a request today? Kling 4.0 and Seedance 2.5 4K fail that test in different ways. Kling's is an undated promise with a preview price list; BytePlus's is a closed trial with no price at all. The other two pass it with published per-second or per-token rates. Kling 3.0 appears as a fifth row because it is Kling's current callable flagship and the only one of the five with a 4K rate on its vendor's page, which makes it the practical answer for a team that was planning on Kling 4.0 at 4K.

An example of why the sorting matters. A studio quoting a client for a batch of 1080p product clips in November needs a per-second rate it can defend. A preview rate from a page that says the price will be announced separately is not that; a trial it is not part of is not that either. The two callable rows give defensible numbers, and the differences between them are large enough to decide the quote.

The failure case is reading the table as a quality ranking. It is an availability ledger with prices, and nothing on it says which model produces the better clip for a given brief. Vendor demos exist for all four; none of them is a controlled comparison, and we have run none ourselves. Quality is a test you run on your own briefs through the doors that are open.

Kuaishou's press release of October 19 says Kling 4.0 officially launched on October 15, with three headline upgrades: steadier large-scale motion in fast action and complex camera moves, native generation of up to 30 seconds in a single pass, and an Omni Reference feature that accepts up to 15 reference assets and up to 10 keyframes. Those are app capabilities. The developer pricing page tells the API story in one paragraph: access is not yet available, the prices shown are for preview purposes only and cannot be used to submit requests before launch, and the official launch date and pricing will be announced separately.

The preview table is still worth reading, because it shows the shape of the billing Kuaishou intends. Without video input, Kling 4.0 is listed at 1.0 unit per second at 720p and 1.5 units at 1080p, with no 4K column. With video input the per-second rate drops to 0.6 and 0.9 units, but the billable duration changes: output duration doubled when the input is no longer than the output, or input plus output when the input is longer. At the page's list price of $0.14 per unit, the no-input rates would be $0.14 and $0.21 per second. Kling 4.0 Flash is listed as coming soon with every price marked to be announced.

The failure case is buying a prepaid pack for it. The page says the standard packs, from $700 for 5,000 units at $0.14 up to $7,560 for 60,000 units at a discounted $0.126, are not compatible with Kling 4.0 or 4.0 Flash, and that early access runs through the business inquiries form. A team that buys units expecting to spend them on 4.0 holds credit with 180-day validity and no rollover that can be spent only on the models the packs support, from V1.0 to 3.0 Omni. Our Kling 3.0 guide covers what those units buy today.

BytePlus's announcement thread of October 23, 2026 calls Seedance 2.5 4K the most capable model of the Seedance 2.5 generation, with more nuanced facial expression and movement and richer detail. The workflow it describes is a two-step one: a Draft mode renders lower-cost 480p previews for testing storyboards and refining shots, and the final render goes out at 4K with 10-bit colour depth. The access line is the one that matters for this post. The model is available to selected enterprise customers, and developers are told to wait for updates on API availability.

What is not published is everything a budget needs: a price, a maximum clip length, a frame rate, or a date. It is tempting to fill those gaps from the base model, since the 2.5 API supports clips of 4 to 30 seconds, and we have declined to. The 4K model is described as a distinct model, not a resolution flag on the existing one, and its limits are unknown until BytePlus publishes them. The ModelArk pricing page, updated October 23, is explicit on the current state: the 4K column for the Seedance 2.5 model reads not supported, and the only Seedance model with a 4K rate on that page is the older 2.0 model at $4.00 per million tokens.

The failure case is treating the 480p draft as a free preview of the 4K result. On ModelArk, draft videos for Seedance 2.5 are generated at a fixed 480p and billed at the normal 480p token rate, so drafting is cheaper than finals, not free, and the 4K trial's own draft economics are unpublished. A team modelling a draft-then-final workflow should count every draft as a paid 480p clip until a price list for the 4K tier says otherwise.

The base Seedance 2.5 model is the callable ByteDance option, sold through BytePlus with 30-second narratives, up to 50 multimodal inputs, and precision editing that replaces products, faces, lighting or backgrounds while keeping the original motion. It bills in tokens rather than seconds. The ModelArk pricing page lists the 2.5 model at $10.70 per million tokens for 480p and 720p output and $11.70 for 1080p, and gives the formula: input plus output duration, times width, times height, times frame rate, divided by 1,024. Resolution therefore scales cost twice, through the rate and through the pixel count.

BytePlus publishes its own worked examples for a 5-second 16:9 clip with no video input: $0.514 at 480p, $1.156 at 720p and $2.843 at 1080p, which it also states per second as $0.103, $0.231 and $0.569. Those are the figures to quote, because they come from the vendor's calculator rather than ours. Two billing rules sit underneath them. When the input includes video, a minimum token consumption applies by resolution, aspect ratio and duration, so a very short edit can cost more than its length suggests. And BytePlus charges only for successful generations; a clip that fails moderation is not billed.

The failure case is buying tokens for seconds you will not get. The product page sells packages such as $32 for 5 million tokens, valid three months, and describes them as roughly 120 to 500 seconds of 480p video. The spread is wide because frame rate, aspect ratio and video input all change the token count, and the package card lists only 480p and 720p as resolution options, so confirm whether a package covers 1080p before buying one for 1080p work. Size a package from your own expected resolution mix, read the usage field the API returns on every job, and reconcile against the package before the three months expire.

xAI's pricing page is the simplest of the four to budget from, because it bills video by the output second at a fixed rate per resolution. Grok Imagine Video 1.5 takes text, image and audio input and lists $0.08 per second at 480p, $0.14 at 720p and $0.25 at 1080p, plus $0.01 per input image. The Lite variant, which we covered in its own business guide, lists $0.02, $0.03 and $0.14 at the same three resolutions. The older Grok Imagine Video model remains at $0.05 and $0.07 for 480p and 720p, with video input charged at $0.01 per second. There is no 4K tier on any of the three.

Two notes on the page change the arithmetic for batch work. xAI's Batch API accepts image and video generation but bills them at standard rates; the batch discount applies to text models only. Priority Processing is not available for video generation at all. So a nightly batch of a thousand clips pays the same per-second rate as an interactive request and cannot pay extra to jump the queue. For a pipeline that is fine, as long as the cost model does not assume a batch discount that the page says is not there.

The failure case is the 1080p jump on Lite. At 480p and 720p the Lite model costs a quarter or less of the full 1.5 model, but at 1080p it costs $0.14 against $0.25, a much smaller gap. A workflow that drafts on Lite at 720p and finals on Lite at 1080p is paying nearly five times the draft rate for the final, and should compare that against finalising on the full model or on another vendor before assuming Lite is the cheap route at every resolution.

To put the rows on one line, take a hypothetical ten-second 1080p clip with no video input and no audio, and multiply each vendor's list rate by ten. Seedance 2.5 on ModelArk comes to about $5.69, using BytePlus's own $0.569 per second estimate, which assumes 16:9 and the vendor's default frame rate. Grok Imagine Video 1.5 is $2.50 and Lite is $1.40. Kling 3.0 without native audio is 0.8 units per second at $0.14, or $1.12; with audio it is 1.2 units, or $1.68. Kling 4.0's preview rate of 1.5 units would be $2.10, and that figure exists only on a page that says it cannot be used.

The spread is about five to one between the cheapest callable option and the most expensive, before any quality judgement, and the ordering changes with resolution. At 480p the same clip is about $1.03 on Seedance 2.5, $0.80 on Grok 1.5 and $0.20 on Lite, while Kling 3.0 has no 480p tier and its 720p rate without audio gives $0.84, so Seedance's premium narrows and Lite's advantage widens. At 4K only Kling 3.0 has a published rate among the callable rows: 3.0 units per second, or $4.20 for the ten seconds, with or without audio.

The failure case is quoting the cheapest number as the cost of the job. Every vendor here bills per generated second or token, and a real brief generates more than one take. If a ten-second final typically needs four attempts and two drafts, the Seedance figure becomes roughly four finals at $5.69 plus two 480p drafts at about $1.03, near $25, while the Grok 1.5 equivalent is four at $2.50 plus two at $0.80, near $12. The multiplier is yours to measure, and it moves the answer more than the list rates do.

A pipeline that treats video models as interchangeable tools, the pattern in our guide to wiring video APIs into agent workflows, has three places where this month's gap between launched and callable shows up. The model registry should hold only identifiers that resolve on the vendor's API today, with the Kling 4.0 and Seedance 4K entries kept as disabled rows carrying the date they were last checked. The cost guard should carry the vendor's list rate and its basis, per second or per token, and refuse a job whose model has no rate rather than falling back to a guessed one. And the review gate should record which model and resolution produced each delivered clip, because a client who approved a 1080p Seedance render in October will ask why November's looks different.

An example of the registry rule in practice. A content team builds a product-video generator in October and, reading the launch coverage, defaults it to Kling 4.0 with Seedance 2.5 as the fallback. The first job fails on an unknown model identifier, the fallback fires, and every clip for the next month comes from the fallback at a cost the team never budgeted. The fix is the one above: no identifier in the registry that has not returned a successful job, and an alert, not a silent fallback, when a primary is unavailable. This is the kind of guard our content engine builds carry by default.

The failure case on the cost side is the token model. A guard written for per-second vendors multiplies duration by a rate and stops. Seedance's token formula multiplies duration by width, height and frame rate, applies a minimum when video is supplied, and only settles the true figure in the usage field of the completed job. A guard that estimates Seedance in seconds will under-count 1080p clips with video input by a margin that depends on the aspect ratio. Estimate with the vendor's formula, then reconcile every job against the returned token count.

The decision above expires when specific text on specific pages changes, and it is worth naming the text so a team can check it in a minute rather than re-reading launch coverage. For Kling, the paragraph on the developer pricing page that says 4.0 API access is not yet available, and the preview-only note above its table, are the gate; when they go, the page will carry a launch date, a unit rate that may differ from the preview, and a note on whether existing packs apply. For Seedance 4K, the signals are a row for the 4K model on the ModelArk pricing page with a token rate, and a BytePlus notice that replaces "selected enterprise customers" with general availability.

An example of a signal that does not count: a third-party router or reseller adding a route named after either model before the vendor's own page changes. Such routes set their own prices and terms, and they tell you nothing about the vendor's API terms, limits or moderation. Our October catalogue snapshot showed no new image or video routes at all, which is consistent with the vendor pages and is the only thing it establishes.

The failure case is checking once. Kling's page says pricing will be announced separately, which means the preview rate is also provisional; a team that budgets November on $0.21 per second for 1080p may find the launch rate higher or lower. BytePlus updated its ModelArk pricing page on October 23, the same day as the 4K announcement, without adding a 4K row for 2.5, which suggests the two are on separate tracks. Put the four pages on a weekly check and re-run the equal-clip table when any of them moves.