Kling 4.0 opened to every fal customer on October 23, after weeks of limited access inside fal's own tools. A single generation can now run 30 seconds without a cut, steered by up to 10 keyframes.
Kling 4.0 Opens to Every fal Customer
Kling 4.0 is Kuaishou's video model, and since October 23 anyone with a fal account can call it through the fal API. Before that, it only worked inside fal's own UX and agent.
The dates are worth pinning down, because three of them circulate. Kuaishou's own release says Kling 4.0 launched on October 15. The API rollout started October 21. fal opened it to all customers on October 23, and the faster Kling 4.0 Flash variant has been in limited early access since September 28.
So this week's news isn't a new model. It's the end of the queue. If you've been waiting on a fal waitlist since September, the wait is over.
A 30-Second Single-Take Clip Changes the Storyboard
Most video models give you five or ten seconds per call and leave the joining to you. That works for a loop on social media. It falls apart the moment a shot has to hold, which is why AI video generation has looked like a slideshow with better textures.
Kling 4.0 generates 3 to 30 seconds in one native pass, so a slow push across a room or a conversation that plays out in real time can arrive as one file. For ad storyboards, that's the difference between a board with twelve panels and a board with four.
Short drama teams get the obvious win, since a 30-second take covers a full beat without a splice. Talking-head creators get a quieter one: no visible seam at second seven, where the model usually changes its mind about the lighting.
There's a catch in all that length. A model that plans 30 seconds of motion has more room to drift than one planning five, and drift is much easier to spot when there's no cut to hide it behind.
Keyframes and Omni Reference: Directing Instead of Re-Prompting
Kling 4.0 takes up to 10 keyframes, which you can assign as first frame, last frame, or checkpoints in between. Instead of describing a camera move in prose, you supply the frames you want it to hit.
Omni Reference goes further, with up to 15 reference assets in a single call: images, short video clips adding up to 30 seconds, voice references and up to seven element references. The practical use is character and product consistency across shots, which has been the weakest part of AI video for two years.
Both features move the job from prompt writing toward shot planning. That's a real skill change for anyone who has spent the last year learning to phrase motion cues.
It's also where vendor demos and API reality tend to part company, so treat the ceiling as a claim until you've run it.
Kling 4.0 Specs: What fal Actually Exposes
The published spec sheet is long. This is the part that survived contact with a third-party endpoint.
Capability | Kling 4.0 detail |
|---|---|
Clip length | 3 to 30 seconds, native single shot |
Keyframes | Up to 10, first, last or intermediate |
Omni Reference | Up to 15 assets, including 10 images and 5 clips |
Resolution | 720p, 1080p, 4K |
Audio | Native stereo, two channels |
HDR | 10-bit at 1080p and 4K, listed as coming soon |
Aspect ratios | 21:9, 16:9, 1:1, 9:16, auto |
Prompt length | Up to 8,000 tokens, multilingual on-screen text |
Video editing | Up to 5 input clips, 1 main plus 4 references, 30 seconds total |
Four items on the spec sheet are still listed as coming soon rather than shipped: 10-bit HDR, full Omni Reference, 4K output and the two-minute extension for editing. On the third-party endpoint SandBase tested, the working set was first and last frames, up to 10 images or 5 video references, and 720p or 1080p output.
That's still a capable model. It's a smaller one than the spec sheet implies.
What Kling 4.0 Costs per Second on Third-Party Endpoints
Kuaishou doesn't publish per-second API pricing, so the numbers below come from a third-party catalog and should be read as reseller rates.
Output | Price per second | 30-second clip |
|---|---|---|
720p | $0.157 | $4.71 |
1080p | $0.235 | $7.05 |
Video editing bills on input plus output seconds, which makes iterating on a clip more expensive than generating a new one from scratch. That's backwards from how most editors work, and it's the kind of line item that shows up at the end of the month rather than the start.
Seven dollars for a half-minute of 1080p is cheap next to a film crew and expensive next to a stock clip. Where it lands for you depends on how many takes you throw away. Is a usable take one in three, or one in thirty? That number decides your real cost per finished second, and nobody publishes it but you.
Reseller rates, not official prices.
The 339-Second Wait and Other Expectation Problems
SandBase ran the model once and published the timing. A five-second 1080p text-to-video generation took 339 seconds. The same prompt on Veo 3.1, with audio off, finished in 51 seconds.
One run isn't a benchmark, and queue depth on a shared endpoint can swamp model speed. Still, the gap is wide enough to plan around.
Here's the workflow consequence. A 30-second clip will take longer than a five-second one, and if you're iterating on a prompt, you're waiting minutes per attempt, not seconds. Studios that budget for 20 variations of a shot need to budget the wall-clock time too.
The other expectation problem is resolution. Marketing pages show 4K. The endpoint in front of you returns 1080p, and a 4K video model that can't hand you 4K isn't one yet.
Kling 4.0 Against Veo 3.1 and Seedance 2.5
Kling sits in the same bracket as Veo 3.1, Seedance 2.5 and Runway, and the differentiator right now is clip length plus keyframe control rather than raw realism.
Veo 3.1 wins on speed in the one comparison available, and it has had longer to settle into production pipelines. Seedance has its own following for stylized motion. Kling's angle is directable length, which matters most to teams producing narrative content on a schedule.
Kuaishou says the model has 60 million creators and 30,000 business customers. That's a company figure with no independent audit, and it covers the consumer app as much as the API.
How to Try Kling 4.0 Without Wasting Credits
Start with a five-second 720p clip and one keyframe. Confirm the output matches what the docs promise before you spend on a 30-second 1080p run, because the second one costs fifteen times as much and takes considerably longer.
Then test the thing you actually need. If your work depends on character consistency, run the same reference across three separate generations and look for drift. If it depends on a single long take, check whether the motion holds at second 25, not second 3.
Keep a written note of what the endpoint returned versus what the marketing page listed. That log will save you an argument with a client later, and it'll tell you when the coming-soon features actually arrive.
