Kling 4.0 Release: 30-Second Single-Shot Video
Quick answer: the Kling 4.0 release on October 16, 2026. The model generates up to 30 seconds in one unbroken shot, accepts up to 10 reference frames, can copy the camera work from a reference video while replacing the scene and characters, and holds motion physics noticeably better than the previous generation. It is the first widely available video model where "one scene beat, one shot" is a realistic production unit rather than a demo trick.
Want to try these techniques today? The video models inside Oimi AI Studio Canvas — Seedance 2.0, Grok Imagine Video, and Storyboard to Video, alongside 30+ image models — run the same long-take structure, free to start.
Kling 4.0 release date and timeline
| Announcement | October 16, 2026, early morning Beijing time |
| Preview window | Late September 2026 — Kling 4.0 Flash opened to premium annual subscribers |
| Maximum shot length | 30 seconds, one continuous take, no cuts |
| Reference frames | Up to 10 (previously a first-frame/last-frame pair) |
| Reference video | Yes — camera work, framing, and rhythm can be copied from a clip and re-cast with new scenes and subjects |
| Pricing | 450 credits / 30s @ 1080p; Flash tier 120 credits / 20s @ 720p |
| Announced, not live | 10-bit 4K HDR output |
The Flash-first release order matters more than it looks. Anyone who spent the last two weeks drafting on Flash has already been choosing compositions, camera moves, and reference sets in the 4.0 environment — which means the interesting results will start appearing immediately, not two weeks from now.
The four things that actually changed
1. Thirty seconds in one shot. Previous generations made you generate five seconds at a time, then hide the seams in an edit. A 30-second take can carry an entire scene beat — a chase, a full performance, a fight — without a cut. This is the difference between generating clips and generating scenes.
2. Ten reference frames, not two. The old first-frame/last-frame pair told the model where a shot started and ended. A sequence of up to ten frames defines the path in between, which turns a set of stills into something much closer to a storyboard the model can walk through.
3. Video reference. You can hand it an existing clip and ask for its camera work back — the framing, the movement, the rhythm — while the scene, the location, and the people are replaced. This is the feature that matters most for anyone who can see the shot they want but cannot describe it in words.
4. Motion that respects weight. The published examples focus on contact and deformation: a crate flattening under a planted foot, bodies meeting a metal wall hard enough to shake it. Those are the details that have separated "impressive AI video" from "footage" for the last two years, and they are visibly better here.
Why 30 seconds is the real story
Duration is usually a spec-sheet number. Here it changes the unit of work.
At five seconds, you are shooting close-ups you will assemble later — a face, a hand, a door. At 30 seconds, you are shooting a scene: an actor enters, something goes wrong, the camera repositions, and the shot resolves. The creative decisions shift upstream, from "which clip do I keep" to "what happens in this scene, and in what order." Anyone who has written for a camera rather than an editor will recognize this as the point where AI video starts behaving like production instead of sampling.
It also raises the bar on prompt writing. A five-second clip can survive a vague prompt; a 30-second take cannot, because the model has 30 seconds in which to lose the plot. Three beats, a camera timeline, and a stated physical outcome will do more for the result than any number of quality keywords — and that is true whether you run the prompt on Kling 4.0 or on any other long-take model.
What the reference system enables
Reference frames and video reference are the features most likely to change how people work, because they solve the oldest problem in AI video: you can picture the shot but you cannot describe it.
- Storyboards that move. Feed four stills as the first, second, third, and fourth frames of one shot and the model fills the motion between them — a street that shifts from flat daylight to rain to dusk to wet night, camera locked in place the whole time.
- Camera work without a camera plan. Point the model at a clip you love, ask for its framing and movement back, and replace the scene and the subject. The camera direction you could not articulate is already in the reference.
- Characters that stay themselves. A reference image for identity plus a line telling the model to hold the face, hair, and clothing for the entire shot is the difference between a 30-second take and a 30-second face swap.
The workflow the pricing suggests
Cost per finished take is the number that actually matters, and the credit ratio between the tiers points at one obvious workflow:
- Draft cheap, finish expensive. At 120 credits for a 20-second 720p Flash generation against 450 credits for 30 seconds in 1080p, a bad composition costs a fraction of a final render. Block the scene on Flash, then spend the full model on the take you are keeping.
- Shorten before you downgrade. A single-event idea works better at 10 to 15 seconds than padded to 30. Cutting duration costs nothing in quality and halves the bill.
- Price the shot, not the session. A 30-second take earns its cost when the scene itself has several moving parts — a crowd, a flock, weather, a second performer. One action stretched over 30 seconds is where credits go to die.
What is not there yet
Worth being clear about, because the launch posts are not:
- 4K HDR is announced, not available. Plan on 1080p today; the 10-bit 4K mode is future-tense.
- Thirty seconds still needs content. Duration is capacity, not quality. A take with one event in it will feel longer than a good 12-second one.
- Physics improved, it did not become predictable. Contact and weight read better; that is not the same as every object behaving. Keep the physical beats explicit in the prompt, and expect to re-roll the ones that carry the most weight.
- Contact sheets come at a price. Anyone generating daily will feel 450 credits per finished 1080p take. The tier split is the mitigation, not a solution.
FAQ
What is the Kling 4.0 release date?
October 16, 2026 — Kling 4.0 officially launched in the early hours of that morning, Beijing time. A limited preview build, Kling 4.0 Flash, had been running since late September for premium annual subscribers, which is why so many test results appeared within hours of the public launch rather than weeks after it.
How long can Kling 4.0 generate in a single shot?
Up to 30 seconds without a cut. The launch material leans on that number heavily, and it is the change with the widest consequences: at 30 seconds a shot can contain several beats, a camera reposition, and a resolution, which is the shape of a scene rather than a clip. The practical limit is content, not the clock — if you only have one event to show, a shorter take usually looks better and costs less.
What is new compared to the previous Kling generation?
Three things: maximum shot length (up to 30 seconds in one take), the reference system (up to 10 reference frames instead of a first-and-last pair, plus a video-reference mode for camera work), and motion handling — contact, weight, and prop deformation hold up better, particularly in dense action. Pricing is now quoted at 450 credits for 30 seconds at 1080p, with a 120-credit Flash tier for 20 seconds at 720p.
Is Kling 4.0 free, and what does a video cost?
Generations are paid in credits rather than being free to run at length. The announced rates are 450 credits for a 30-second 1080p take and 120 credits for a 20-second 720p take on the Flash tier — which makes Flash the sensible drafting model and the full 4.0 build the final pass. For anyone producing daily, the cost per finished take depends as much on how many re-rolls a prompt needs as on the headline price.
Do I have to use Kling 4.0 to get results like these?
No. The structure of a good prompt transfers — a stated shot length, numbered references, action as a chain of cause and effect, a camera timeline, and an explicit physics line all respond well on other long-take models, including Seedance 2.0, Veo, Sora, and Grok Imagine Video. What changes between them is the duration ceiling and how literally each one reads camera instructions. If you would rather compare than re-paste a prompt into five tabs, Oimi AI Studio Canvas runs the video models — Seedance 2.0, Grok Imagine Video, Storyboard to Video — next to 30+ image models in one workspace, free to start, so you can put the same prompt on two engines before committing to a final take.
Keep reading
-
AI Portrait Prompts: 10 Copy-Paste Portrait Styles That Look Shot, Not Rendered (2026) — the still-image side of the same problem: light, lens, and texture instead of vibes
-
Oimi Prompt Library — Find a look you love, copy the prompt, and make it your own