Kling 4.0: What It Changes for Brands Commissioning AI Video — and What We Don't Know Yet

What was announced, what was reported, and what's still unknown

Launch week isn't short on information. It's short on sorting. An official post, a tech blog's spec sheet and a lookalike site's pricing table all look the same size in search results. So before asking what Kling 4 changes, separate three layers.

The announcement went out on Kling's official X account on September 27–28, 2026 (sources differ by a day), under the line "You Call the Shots". The sentence that matters: "Kling 4.0 is coming this October. Kling 4.0 Flash is live now for Ultra Yearly subscribers."

What Kling announced officially

• Native 30-second generation, plus video extension.
• Up to 10 keyframes per generation.
• Omni Reference with up to 15 multimodal references, more consistent results and enhanced video editing.
• High-quality stereo audio, more accurate lip sync, and support for multiple languages, accents and dialects.
• Output up to 4K with 10-bit HDR, and more stable dynamic motion.

What's been reported but isn't in the official post

• According to Morphic and GPTunneL, the full model generates 3 to 30 seconds at 720p, 1080p or 4K, and adds a 21:9 aspect ratio. Both list 10-bit HDR at 1080p and 4K as "coming soon".
• According to GPTunneL, the 15 references break down into up to 10 images, up to 5 videos (30 seconds in total) and up to 7 saved characters or objects. Morphic and GPTunneL both say voice samples can be attached too.
• According to Morphic and GPTunneL, video extension up to two minutes is marked "coming soon". According to Morphic, prompts can run to 8,000 tokens.
• According to HuggingNews, "enhanced video editing" means changing camera movement, style or background in an existing clip without regenerating all of it. Kling itself only says "enhanced video editing".
• Kling 4.0 Flash, according to Morphic and GPTunneL: 3 to 20 seconds, 720p, 8-bit SDR, from text or an image, aimed at fast, high-volume work.

What's still unknown

• An exact release date. "October" is all there is.
• Pricing and credit costs. Not published, for 4.0 or for Flash.
• API access. Not confirmed.
• A language list. Kling hasn't said which languages 4.0 speaks.
• Per-plan limits. Length, resolution and generation caps for each subscription tier haven't been published.

Kling 4.0 vs Kling 3.0, Seedance 2.5 and Veo 3.1

This isn't a ranking. It's what each company has published, so you can see what carries over from Kling 3 to Kling 4 and where the competition stands. Where there's no official figure, the table says so. Kling 3.0 figures come from Kuaishou's launch release, Veo 3.1 from Google's Gemini API docs, and Seedance 2.5 from a HowAIWorks report citing ByteDance Seed's launch post. The Flash specs and the 3–30 second range for the full model are reported by Morphic and GPTunneL, not announced by Kling.

Sora isn't in the table because it no longer exists as a product: OpenAI closed the Sora app on April 26, 2026 and discontinued the Sora API on September 24, 2026.

What changes in the production process

I haven't run the full Kling 4.0. It isn't out. What I can do is read the three changes against the workflow we already run at A&E Creations: a base frame from an image model, animation in a video model, and identity blocks that hold a character across shots. Each change in Kling 4.0 lands on a different step of that workflow.

10 keyframes: the storyboard becomes the prompt

Right now the most precise way to control motion is a start frame and an end frame. You define where the shot begins and where it ends, and the model fills in the middle. That works, but two anchors aren't enough for a shot in which three things happen. Kling says 4.0 takes up to ten.

In practice, the storyboard stops being something you show the client before work starts. It becomes the instruction. Every frame on the board has to be precise enough for the model to get from it to the next one, so each frame is built in an image model and approved before anyone hits generate.

Those frames can come out of ChatGPT Images 2.5 and go into Kling as keyframes. I cover how to get consistent frames for animation in Animation with ChatGPT and Codex.

15 references: keeping faces and products consistent

The most expensive problem in brand AI video isn't image quality. It's the product, the face or the logo drifting slightly from shot to shot. It's the same problem behind building a virtual influencer who stays the same person. Omni Reference with up to 15 references moves part of that job inside the model.

But the model can only hold what you give it. A brand that sends three photos from its website gets the consistency of three photos from its website. A brand that sends the product from several angles, a clean character sheet and defined brand colours gets something else entirely. Each reference also needs one job: in the Better production file we assigned references by role, so the character reference never governs the product. The reference pack becomes a brand asset, the same way a brand book is.

30 seconds in one pass: fewer seams, same brief

Until now, a 30-second scene meant several short clips stitched together in the edit, and every seam is a risk: lighting that jumps, a character who shifts position, audio that cuts out. Native 30-second generation means fewer seams.

It doesn't remove the need to write in shots. Thirty seconds is a scene, not a film, and someone still has to decide what happens in each second. There's also a flip side: when one second fails inside a 30-second generation, you may have to regenerate the whole thing. If the enhanced editing works the way it has been reported, some of that becomes fixable without starting over. Until we've tested it, I plan as if it doesn't.

Non-English dialogue: test before you commit

Kling lists "multiple languages, accents and dialects" for 4.0 but hasn't published which ones. The Kling 3.0 launch release said much the same, also without a list. If your market's language matters (for us, that's Hebrew), test it on the full model before you build a campaign around it, and plan a separate voice layer until it passes. We cover the Hebrew case in a separate guide (Hebrew article).

What the four changes have in common: the work moves upstream, into the brief, the board and the references. That's what our AI consulting is built around. Not picking the newest model, but building a process that can swap models without starting over. For what that looks like from brief to frame, see the CRINIÈRE production file.

Want to learn how to direct a result with AI, not just write a prompt? Open the directing guide in OMER METHOD →

Commissioning AI video: what to decide now

1. Don't wait for the model. The brief, the frame board and the reference pack carry over to Kling 4 as they are. The model is the last layer of the process, not the first.
2. Ask for a frame board before animation. Approving a frame is cheap. Approving a video and then finding out the product is wrong is expensive. A studio that won't show you frames before it animates is working for itself, not for you.
3. Write the brief in shots. Not "a 30-second video about the product", but a list: what happens in each shot, what the viewer needs to take from it, and where the product is.
4. Put together a reference pack. Product shots from several angles, faces you have the rights to use, your logo and colours. That's how 15 references end up working for you.
5. Budget for iterations. No model delivers the final cut on the first try. Ask how many revision rounds are included, and what happens after them.
6. Settle rights and brand safety. Whose face and voice are in the references, and who owns the files. Check the platform's commercial terms for your plan.
7. Ask which model, and why. "Because it's the newest" isn't an answer. The answer should follow from what the video has to do.

Warning: lookalike Kling 4 sites are flooding search

When we checked on September 29, 2026, a search for "Kling 4" in English and in Hebrew returned a large share of unofficial sites with near-identical names: kling4.org, kling-4.com, kling4.pro, kling4.app and kling3.io. Some of them show "plans" and version names the site made up. EvoLink, for example, points out that the "Flash/Pro" tiers on kling3.io are that site's own naming.

Kling has two official sites: kling.ai and app.klingai.com. The rule is simple: no credit card, no brand assets, no product shots and no one's face goes into a domain that isn't one of those two.

What we'll update when the full model ships

• Release date, pricing and credit costs, once they're official.
• Whether there's an API, and what each plan's limits are.
• The published language list, and our own test of non-English dialogue.
• A like-for-like comparison: the same shot through Kling 3.0 and Kling 4.0, in the same process.

Every update will be logged here, with a date.