Kling 4.0 and the New Controllability Race in AI Video

The biggest complaint about AI video has never really been quality. It has been control. You write a prompt, the model produces something gorgeous, and then the client asks for one small change, and you have to regenerate the whole thing and hope the character's face and the lighting come out the same. That gap is where Kling 4.0 is aiming.
Kuaishou's Kling announced the model on September 28. The full version launches sometime in October; a lighter "Flash" variant, limited to 720p and 20-second clips, went into early access the same day for Ultra yearly subscribers. The full model extends single-shot generation to 30 seconds and adds 1080p and 4K output, with 10-bit HDR planned at the higher resolutions.
The specifications that matter are not the resolution numbers. They are the editing features. Three of them stand out.
The first is local editing. You can take a finished video and change one thing, an expression, a motion, a camera move, a style, or the background, while everything else stays locked. You can feed up to five clips including one master. That is the difference between "the client changed one detail" being a full redo and it being a single edit. It is also the feature that closes the gap between AI video as a demo and AI video as a revision loop.
The second is keyframes. Image-to-video now accepts up to ten keyframe images plus first and last frames. You fix the product shot, the usage scene, and the packaging closeup in order, and let the model fill in the motion between them. For anyone doing product video, this turns generation into storyboarding. The storyboard gets confirmed before you spend credits, which is where the real time savings live. A storyboard is also a contract: the client agrees to the sequence before the expensive part starts.

The third is Omni Reference, which accepts up to 15 inputs at once: ten images, five videos totaling 30 seconds, and up to seven subjects with voice references. You address them in the prompt with @image1 and @video1, telling each reference what job it has. One reference holds identity, another holds style, another holds pose. It is the same principle people already use in image editing, extended to moving footage.
These three features share a single through-line. They are all about reducing the number of times you have to start over. The early AI video era was a slot machine: you pulled the lever, got a clip, and pulled again if you disliked it. Kling 4.0 is pushing toward a deterministic pipeline where the parts you care about are pinned down and only the motion is left to chance.
Pricing is aggressive in a way that says Kuaishou is playing for volume. Yearly billing puts the entry tier at about $21 a month for 180 credits, roughly $1.90 per video, and the top tier at $90 a month for about $1.10 per video. No watermarks at any tier, so finished clips ship directly to clients. For context, Google's Veo 3.1 standard tier lists around $0.40 per second, while Alibaba's Wan 3.0 runs $0.05 to $0.20. Chinese video models are buying market share with price, and Kling is not hiding it.
The controls land on top of a model that has already proven it can hold a character across scenes. That combination, control plus consistency plus a low per-second price, is what makes Kling 4.0 feel like a different category from the early Sora-era "wow, look what it generated" demos. The demos are over. The work is now about iteration.
This is the quiet theme running through the whole video generation field right now. Runway's Gen-4.5 pushes reference-driven consistency, letting a single reference image hold a subject across multiple shots. Veo 3.1 sells audio and picture generated together, so footsteps land where the feet land. Wan 3.0 competes on price and open weights. Kling 4.0 is staking its claim on the unglamorous problem of changing your mind after you have seen the result. That has always been the real bottleneck in creative work, and it is the one a benchmark never measures.
The reason this matters commercially is that the people who pay for AI video are not the ones making one viral clip. They are the ones making a hundred product videos a month, and they lose more time to revision than to generation. A tool that makes revision cheap attacks the actual cost center. That is why Kling's local editing and keyframes, not its 4K spec, are the real announcement.
There is a useful way to measure the shift. In the old workflow, a single "regenerate" was the unit of failure. Every unwanted variation cost you a full generation, a full credit, and a full wait. Kling 4.0's local editing turns that unit of failure into a small operation. The emotional difference matters as much as the economic one: creators stop treating the model as a slot machine and start treating it as a collaborator they can direct. That is the moment a tool crosses from novelty to workflow, and it is the moment the AI video market actually started to mature.
The audio side of Kling 4.0 is easy to overlook but worth a mention. The full model generates two-channel stereo audio in sync with the picture, with improved lip sync and support for nine languages including multiple Chinese dialects. For product videos and short drama, where a voiceover that matches the mouth is the difference between believable and jarring, that is a bigger deal than the resolution bump. Early AI video produced silent footage and left you to dub it later; the models that generate sound and picture together remove an entire post-production pass.
The strategic picture is worth stepping back to see. Kuaishou is a short-video platform, and Kling is its AI arm. That means Kling 4.0 is not a tool company trying to sell you a subscription. It is a platform feeding its own content ecosystem, and the aggressive pricing reflects a different endgame than a standalone vendor's. The same is true of ByteDance's Seedance and Dreamina. The Chinese video models are subsidized by platforms that benefit from the content they generate, which is a structural advantage the Western labs do not have, and it is why the price war keeps tilting in one direction.
The one thing the announcement did not fully answer is the detail everyone in production cares about next: credit pricing and API access for the full model. Kling has published the subscription tiers but not the per-credit cost for Kling 4.0 itself, and its API packages do not support the new model yet. That gap will close when the full model launches in October, but it is worth flagging because the economics of a revision-friendly model are different from a one-shot model. If a local edit costs a fraction of a full regeneration, the value proposition is obvious. If it costs nearly the same, the feature is a convenience rather than a cost breakthrough. Either way, the direction is clear, and the pricing details will determine how fast teams adopt it.
For anyone shipping product video, marketing clips, or short drama, the practical takeaway is simple. Look past the resolution spec. Ask what happens when the client asks for a different angle. If the answer is "regenerate everything," you are still living in 2025. Kling 4.0 is betting the answer in 2026 is "edit one thing and keep the rest." Whichever vendor wins that bet will own the next phase of the market, because revision is where the time, and the money, actually go.
Related articles
Meitu now has 19.83 million paying users, and that says more than any benchmark
One side adds 1.39 million subscribers in a quarter and books $220 million a month. The other raised the price of a 15-second clip from 0.65 to 5.03 yuan. Same cohort of Chinese imaging apps, completely different arithmetic.
They Rebuilt a Proprietary Vision Model With Public Data and Beat It on One Benchmark
The paper's best contribution isn't the score. It's that the data pipeline, the training code and the intermediate checkpoints all shipped.
4LC Caches the Scene So a Video Model Stops Redrawing It
Every other method in the comparison wins one axis and loses the other. 4LC's paper is candid about which half it still can't solve.
Eleven Skills That Clone an App, Down to Writing the Store Listing
The pipeline ends with a skill called the Entrepreneur, which reads what the original's users complain about. That's where cloning stops being a copy.