Kling 4.0 Is Live for Everyone — the API Is Still Early Access

Kling 4.0 became available to everyone on 15 October 2026, according to K​uaishou's own release note, and that sentence is doing more work than it looks like it is. The consumer product and the developer product did not ship on the same clock. In the K​ling AI app and on the website, the full model is live and selectable in the video workspace today. On the open platform, the banner above every page of the API documentation still reads "Kling 4.0 API Is Now Available for Early Access" — early access, not general availability, which in K​uaishou's case has historically meant an application and a wait rather than a rate card you can read. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

That split is the whole story for anyone building rather than browsing. The previous generation, Kling 3.0, is the one with a published price and an endpoint that has been in production since February; the new model is the one with the better spec sheet and a queue. This piece is about what actually changed on the 15th, which of the headline numbers are shipping today and which are still labelled "coming soon" on the vendor's own page, and what the gap between the two clocks costs you if you were planning to put Kling 4.0 in a pipeline this month.

What the 15 October release actually contains

The release note is unusually specific for a launch post, and three things in it are new rather than restated. The first is that the model family now has two tiers that are genuinely different products rather than one model at two speeds. Kling 4.0 proper does text-to-video, image-to-video, first-and-last-frame, multi-keyframe, omni-reference generation and targeted video editing, at 3 to 30 seconds, 720p to 4K. Kling 4.0 Flash does text-to-video and first-frame image-to-video only, at 3 to 20 seconds, 720p, 8-bit SDR. Flash is the tier the release note describes as "built for frequent creation," which is vendor language for cheaper and faster with a lower ceiling.

The second is the reference budget, which is the part that changes what a prompt can carry. The release note puts it at up to 15 items across four categories in a single generation: up to 10 images, up to 5 videos whose combined duration cannot exceed 30 seconds, voice references, and up to 7 elements of which up to 3 may be video-based. Those are per-category caps inside a combined total, not four independent allowances — the same structure the preview pages used in September, now attached to a shipped model.

The third is the delivery surface. K​uaishou states that MCP and CLI now support Kling 4.0, including omni-reference and keyframe generation. That is a smaller-sounding line than it is: it means an agent or a build script can drive the model without going through the web editor, which is the difference between a tool a person uses and a component a system calls.

The spec sheet, read closely

Everything below is from K​uaishou's release note and feature page, read on 22 October 2026. None of it has been independently measured — there is no third-party Elo for Kling 4.0 on the video boards as this is written, and the vendor has published no benchmark of its own.

• Duration — 3 to 30 seconds in a single generation for the full model; 3 to 20 seconds for Flash. Extension in multiple passes up to 2 minutes is listed, but marked coming soon. • Resolution — 720p, 1080p and 4K for the full model; 720p for Flash. 10-bit HDR at 1080p and 4K is explicitly "coming soon" on both the release note and the feature page, so the HDR claim is a roadmap item, not a shipped capability. • Aspect ratio — 21:9, 16:9, 1:1, 9:16 and Auto on the full model. • Keyframes — up to 10 keyframe images as input, placeable to define character states, scene changes and story beats. • References — up to 15 combined assets as described above, plus omni-reference for characters, layouts, styles, composition and motion. • Audio — two-channel stereo, improved lip sync, dialogue across nine languages plus named accents and dialects (American, British and Indian English; Beijing, Taiwanese, Northeastern and Sichuanese Mandarin among others), and on-screen text generation in the same languages plus emoji. • Prompt length — up to 8,000 tokens of input text. • Editing — up to 5 input clips including one primary, each 3 to 30 seconds, combined duration capped at 30 seconds; targeted edits to expression, camera position, background or style while the rest of the scene holds.

Two of those rows deserve to be read twice. The 4K row is real as an output setting but the 10-bit HDR qualifier attached to it is not shipped, and a colour-managed pipeline that assumes HDR will find out the hard way. The reference row is the one that most summaries get wrong: 15 is a budget you allocate across images, video seconds, voice and elements, and a prompt that spends it all on stills cannot also carry five reference videos.

What is still missing, by the vendor's own labelling

Three capabilities appear in the launch material with a future tense attached, and it is worth separating them from the parts of the model you can use today. 10-bit HDR output at 1080p and 4K is coming soon on both the release note and the feature page. Video extension beyond 30 seconds, up to the advertised 2 minutes, is coming soon. And native 4K with 21:9 is described on the feature page as coming soon for the full model in the same breath as the resolution itself, which is a softer commitment than "4K is supported" implies.

None of that makes the launch hollow. A 30-second single-pass generation with ten keyframes and a 15-asset reference budget is a real capability, and it is the reason the model is worth a look. It does mean the shopping comparison should be run on 1080p, where the model definitely ships, rather than on 4K HDR, where it does not yet.

Cost: what K​uaishou publishes, and what it does not

The consumer side runs on credits through K​ling AI's plans, and the API side has a pricing section in the developer documentation. What K​uaishou has not published, as of 22 October, is a per-second rate for Kling 4.0 that a developer can budget against — the model is behind the early-access banner, and early access in this API has previously meant an application rather than a public price. Several third-party platforms list Kling 4.0 endpoints with per-second figures, but those are reseller rates, not the vendor's, and they will move independently of whatever K​uaishou eventually publishes.

For contrast, Kling 3.0 has a price that has been stable for months. On OrcaRouter, the kling/kling-v3 entry is listed at $0.084 per request — the flagship text-to-video and image-to-video model with multi-shot, subject and motion control, 3 to 15 second clips, up to native 4K. Kling 3.0 Omni is at the same $0.084, Kling 3.0 Turbo at $0.112, and the older 2.1 and 2.5 tiers run from $0.042 up to $0.28. That is the number to compare a new model's rate against when it arrives, not the launch narrative.

We do not route Kling 4.0 yet. There is no identifier for it in our catalogue, so there is nothing to put in a config file and no endpoint to point at, and saying otherwise would be a fiction. What we do route is the generation it replaces, on the same OpenAI-compatible endpoint as 200+ other models, at provider list price with 0% markup — which matters here because when K​uaishou does publish a Kling 4.0 rate, a pass-through catalogue reflects it the same day rather than after a contract cycle. If you are building against the family now and want the new model the week it opens up, the practical shape is to keep the call behind a model name you can change, and to route it through a gateway that can fall through to Kling 3.0 when the new endpoint is unavailable rather than hard-coding either.

What to watch over the next month

The first datable event is the API coming out of early access, because that is when a price appears and the model becomes something you can put in a budget instead of a slide. The second is the first independent score: Kling 4.0 is absent from the Artificial Analysis video boards today, which is not a bad score but is an absence, and the previous generation's numbers on those boards are mid-table at best — Kling 3.0 1080p Pro sits around Elo 1,055 on image-to-video, well below the current leaders. The third is whether the "coming soon" labels actually clear: HDR, 2-minute extension, and native 4K with 21:9 are three separate promises, and launches that ship with three of them outstanding have a habit of clearing two.

Until the first two happen, the honest position on Kling 4.0 is that the product shipped and the platform did not, and the gap between them is where most of the risk lives for anyone who is not editing by hand in a browser.