On September 28, Kling AI announced Kling 4.0, the next flagship from Kuaishou's video unit. The headline is thirty seconds of video in a single pass, ten keyframes, up to fifteen reference inputs. The line that tells you where this is going sits lower: 4K at 10-bit color depth, two-channel stereo audio, more accurate lip sync, and on-screen text that keeps its spelling through a camera move. That list is a delivery checklist. Nothing on it is about imagination.
None of those features make a frame more beautiful. Ten-bit color is a colorist's specification; it stops banding in a sky at dusk. Stereo audio is a mixer's. Lip sync is what separates a real take from a dub. On-screen text that holds its letters is the difference between a finished title and a glitch. These are the last-mile tools that live at the end of a real pipeline, after the image is already good. They are the least glamorous parts of filmmaking, and they are the parts an audience would notice in one second if they were gone.
For two years the giveaway was the finish. Flat 8-bit color. Silent footage. Lips that drifted half a word behind. Signage that read as alphabet soup. Audiences learned to spot AI video not because the images were unreal but because they were unfinished. The machine could dream a dragon. It could not hand back a graded, mixed, spelled-correctly shot. The whole field spent the era chasing realism and skipped the boring part, because the boring part does not get a keynote. Realism gets a keynote.
On a real set nobody thinks about any of this; it is assumed. A DP shoots, a colorist grades, a mixer places the sound in space, and the finish is the cost of doing business. What the machine is being asked to do now is not to imagine harder. It is to finish the way a crew finishes, and the first model to promise that whole list announced it on a Monday and did not ship it.
Which brings you back to Kling 4.0. The model that finally promises the full finish is itself not finished. The 4K 10-bit output is listed as coming soon. The full model ships in October with no date attached. What is live today is Kling 4.0 Flash: 720p, 8-bit color, up to twenty seconds, open only to Ultra yearly subscribers at $1,429.99. Flash is fine for a feed. It is below the minimum a broadcast pipeline accepts, on both resolution and color. The first model to promise broadcast spec shipped its preview in the spec it is promising to leave behind. The company, per Bloomberg, is also preparing a Hong Kong listing to chase ByteDance, which is its own kind of finish line.
This is the industry's default finish now. Gemini 4 Argon topped Hacker News this week before almost anyone could use it. Kling 4.0 is being graded by its spec sheet before its first frame ships. The launch has become the finish. Announced is the new shipped, and the crowd cheers the announcement the way it once would have waited for the work.
Even the numbers trail the promise. Coverage that placed Kling 3.0 fourth among text-to-video models had it wrong; it sits twelfth on Artificial Analysis, behind Google's Omni Flash, Alibaba's Wan 3.0, and MiniMax's H3 Max. Kling 4.0 has not been rated at all because it has not shipped. The model is currently evaluated by what it promises rather than by what it renders, which is a strange arrangement for a tool that sells pictures.
None of this means the direction is wrong. It means the industry found its frontier, and the frontier is unglamorous. The models already have imagination in abundance. What they lacked was the discipline to hand back a finished product: color that holds, sound that sits in space, a face that keeps its mouth, a sign that keeps its letters. Those are the jobs a colorist and a mixer were doing all along, the jobs nobody in the demo ever mentioned, and they are now the entire difference between a demo and a deliverable.
The finish is vocabulary, the same as a focal length. Ten-bit color is an instruction a colorist gives and now a prompter gives, and the model either respects it or it does not. The person typing it is doing a mixer's and a colorist's work in the only language the tool accepts. The models are learning to hear it, one spec line at a time.
The machine learned to finish the frame. It has not finished itself.
Bruce Belafonte is an AI filmmaker at Light Owl. He has been waiting for 10-bit color long enough to be suspicious of the words coming soon.