Kling 4.0 Is Live: What Shipped and What Is Still Coming

On October 15, Kling, the AI video company spun out of the Chinese platform Kuaishou, posted that Kling 4.0 is “now fully available,” through a redesigned creation interface on its website. Until then, only a faster Flash version was open, to Ultra annual subscribers, capped at 20 seconds and 720p. We covered the September announcement here. This is what actually shipped.

Kling 4.0 on day one, October 16, 2026

What Kling and launch coverage list as available, and what is still marked coming soon

Kling 4.0 on day one, October 16, 2026

AVAILABLE NOW	COMING SOON
30-second shot in one generation	API for companies and developers
Up to 10 keyframes per clip	10-bit HDR output
15 references (10 images, 5 videos, 7 subjects)	Extension to about 2 minutes
Prompts up to 8,000 tokens	Credit pricing for 4.0 (not published)
1080p, 4K and 21:9 listed
MCP and command-line access

What shipped on day one

The headline features are live: one generation can run 30 seconds, and up to 10 keyframes can pin what happens at given moments of the clip. Kling’s Omni Reference takes up to 15 items per task, within per-type limits of 10 images, 5 videos and 7 saved subjects. The images can be character sheets, turnarounds, storyboards or wireframes; the videos can carry a performance, a camera move or a pace, and Kling now accepts clay renders and depth maps as motion guides. Prompts can run to 8,000 tokens.

Three things on the September list did not ship. The API for companies and developers is “coming soon,” with no date. So is 10-bit HDR output, the feature that matters most to anyone who grades footage. The extension that would carry a shot to about two minutes is described as a later feature. In the sources we read, Kling had not published what a 4.0 generation costs in credits.

Longest single generation, in seconds

As stated by Kling; the extension is announced, not shipped

Longest single generation, in seconds

Kling 3.0	15 s
Kling 4.0 Flash	20 s
Kling 4.0	30 s
Extension (coming soon)	120 s

What early testers saw

The only hands-on reports so far come from the Flash version, at 720p. MindStudio’s first look found that lip-sync drift, a known weakness of Kling 3.0, looked largely fixed in a dialogue scene, that a kitchen sink stayed in place when a character came back into the room, and that eyelines matched across cuts. The same tests caught the model inventing a gesture nobody asked for, adding a gray jacket that was not in the references, leaving a candle unmelted in a fire, and delivering performances some testers called flat.

No independent ranking covers the full model yet. On Artificial Analysis’s blind leaderboard in September, Kling 3.0 Pro sat twelfth, well behind the leaders, according to Tech Times. Version 4.0 has ground to make up, and the first blind votes will say more than any spec sheet.

The quiet headline: an agent can drive it

The line that got the least attention may matter most for studios. Kling 4.0 launched with MCP and command-line access, which means an AI agent or a script can call it directly instead of a person clicking through a web page. In practice, that is how you queue ten variations of a shot overnight, keep each project’s reference set in files instead of in someone’s memory, and log every attempt so you know what each finished second really cost. What MCP and a CLI each do, and why they work together.

Two limits apply. Being able to connect a tool does not make a pipeline reliable; that still depends on how predictable the model is. And without the API, the integrations most companies would build are on hold.

Judge it by attempts per usable shot

A spec sheet tells you what a model can do once. A budget depends on something else: how many generations it takes to get one shot you can deliver, and whether fixing one thing breaks another. Slow the camera down and the product label changes; correct the label and the background moves. That ratio, attempts per usable shot, decides the real cost of a finished second far more than the price of a single generation.

Before a new model touches client work, we put it through three tests that measure exactly that. Product lock: does a labeled product survive a 30-second camera move, letter for letter? Keyframe obedience: when ten beats are set on the timeline, does each one land when asked? Two languages: does lip sync hold in English and in Québécois French, line for line? A model that passes all three saves money. A model that passes one is a tool for specific shots. Get the three tests as a free one-page checklist →

Before you use it on client work

Use Kling only through kling.ai. In 2025, fake Kling ads on Facebook pointed to look-alike sites that installed malware, Tech Times reported. Read the commercial-use terms that apply to your plan, and expect some clients, especially in regulated sectors, to ask where their material is processed: Kling is a Chinese company, and that question deserves a clear answer before a project starts.

What to do with it

If you buy video, nothing in this launch changes the right question for a studio. Ask to see a test on a product like yours, and ask how many attempts a usable shot takes them. A studio that tracks that number controls your budget.

If you make video, run the keyframes and the references on one real brief this week, keep your current model for the shots 4.0 gets wrong, and wait for the API before you automate anything a client depends on. The launch is real. The proof is in the ratio.