Kling 4.0 vs Kandinsky 6.0 Video: The Head-to-Head Nobody Has Run

The only published human evaluation that puts Kling 4.0 and Kandinsky 6.0 Video in the same room does not exist. What exists is a paper from Kandinsky's own lab, submitted to arXiv on 4 October 2026, that compares Kandinsky 6.0 Video Pro against Kling 2.6 — two generations behind the model Kuaishou shipped to everyone on 15 October. That gap is the honest starting point for this matchup, and it cuts in a direction most comparison pages ignore: on the columns the paper does measure, the older Kling won several outright. Kuaishou's claims for 4.0 on exactly those columns are the ones nobody has verified.

So this is not a page that will tell you which model is better, because the evidence to say so does not exist yet. It is a page about what each one is actually built to do, what the one real head-to-head found, and which of the two you can put in a pipeline this week.

Two products that only look alike

Kling 4.0 is a closed, credit-metered generator from Kuaishou. The full model does text-to-video, image-to-video, first-and-last-frame, multi-keyframe, omni-reference and targeted editing, at 3 to 30 seconds and 720p to 4K. It is sold through the Kling app and website, with an API that Kuaishou began opening in waves from 21 October and whose developer price sheet still labels the model "Coming Soon" with preview rates only.

Kandinsky 6.0 Video is the opposite kind of artifact. Sber's Kandinsky Lab released it on 6 October 2026 as two checkpoints — Pro at 29B parameters and Lite at 3B, plus a separate 1.4B super-resolution model — under MIT, with weights on Hugging Face, code on GitHub, and day-0 integration in Diffusers, ComfyUI, FastVideo, SGLang and vLLM-Omni. It generates five-second clips at 121 frames and 24 fps, with synchronized 44.1 kHz audio and lip-sync, from a text prompt or a reference image.

One is a service with a duration advantage. The other is a download with a licence advantage. Almost every practical difference between them falls out of that sentence.

The one head-to-head, read honestly

The Kandinsky paper's human evaluation is the closest thing to a shared scoreboard, and it is worth reading closely because it is a lab grading its own new model against a competitor it chose.

• Overall visual quality — Kling 2.6 ahead, per the paper's own raters.

• Visual aesthetics — Kling 2.6 ahead.

• Audio quality — Kling 2.6 ahead.

• Audio-video synchronization — Kling 2.6 ahead.

• Sound prompt following — Kandinsky 6.0 Video Pro ahead.

• Speech quality — Kandinsky 6.0 Video Pro ahead, consistent with its reported 47% reduction in generated-speech word error rate after reinforcement-learning post-training.

• Camera control in image-to-video-with-audio — Kandinsky 6.0 Video Pro ahead.

• Visual prompt following — roughly tied.

Three things follow. First, the newer model did not sweep the older one, which is unusual for a lab's own evaluation and worth crediting. Second, Kandinsky's wins cluster where the architecture is unusual — it denoises video and audio latents jointly through one transformer rather than running two models in sequence — and that shows up in speech and sound-prompt adherence more than in raw picture quality. Third, the comparison tells you nothing about Kling 4.0, whose entire pitch is better realism, steadier large-scale motion and cleaner lip-sync. Those are precisely the columns where its predecessor already led.

If you want the comparison to be real rather than inferred, the missing artifact is an independent arena run with both models in it. Neither has one against the other today.

The spec gap that decides most jobs

• Clip length — Kling 4.0 runs 3 to 30 seconds in a single pass, with extension toward 2 minutes listed as coming soon. Kandinsky 6.0 Video is trained at 121 frames and generates 5 seconds. Anything longer is a stitching problem you own.

• Resolution — Kling 4.0 ships 720p, 1080p and 4K, with 10-bit HDR at 1080p and 4K labelled coming soon on Kuaishou's own release note. Kandinsky generates at SD and upscales through its 1.4B super-resolution model to Full HD, 1920×1080, with ×2, ×2.25 and ×4 factors.

• Control inputs — Kling 4.0 takes up to 10 keyframes and a 15-item reference budget split across 10 images, 5 videos capped at 30 seconds combined, voice references and up to 7 elements. Kandinsky 6.0 Video takes a text prompt or a single reference image, applied as a masked tail frame.

• Audio — Kling 4.0 outputs two-channel stereo with lip-sync across nine languages and named accents. Kandinsky outputs 44.1 kHz AAC with audio on by default; the pipeline returns synchronized speech, music or ambience even when you pass no audio parameters.

• Prompt length — Kling 4.0 accepts up to 8,000 tokens. Kandinsky's serving recipe is built around the caption style the model was trained on, long, medium or short, not an open-ended script.

• Licence — Kling 4.0 is closed, governed by Kuaishou's terms. Kandinsky 6.0 Video is MIT for code, weights and the Diffusers integration.

That last row is doing more work than its length suggests. A 30-second closed model and a 5-second MIT model are not interchangeable even when their picture quality is close, because one constrains what you can build and the other constrains what you can generate.

What each costs, in the unit that actually bills

Kuaishou's developer pricing page lists Kling 4.0 at 1.5 units per second at 1080p, which it converts to $0.235 per second, and 1.0 unit, $0.157 per second, at 720p. There is no 4K rate — the column reads "-". With video input the rate drops to $0.094 per second at 720p and $0.141 at 1080p, but billing counts output duration times two when the input is shorter than the output. The page also says the figures are "for preview purposes only" and that API access is "not yet available," which sits awkwardly beside Kuaishou's 21 October announcement that the API had gone live for enterprise clients and developers. Read together, the fair statement is that the API is opening in stages and no public, stable rate card exists on 23 October.

At the preview rate, a five-second 720p clip costs about $0.79, and a 30-second one about $4.71. That is the number to sanity-check any third-party reseller quote against, because reseller rates move independently of Kuaishou's.

Kandinsky 6.0 Video has no per-second price because there is no vendor endpoint to buy. Its cost is GPU time, and the repository publishes measured figures for one five-second clip after warmup, excluding model loading and encoding: Pro at Full HD takes 402 seconds on an H100, 854 on an RTX 5090, 1,247 on an RTX 4090 and 3,530 on an RTX 5060 Ti. Lite at Full HD takes 284 seconds on an H100. At any H100 rate you happen to pay, 402 seconds is 6.7 minutes of that card per five-second clip — multiply through and the arithmetic is yours to run, because the vendor does not sell you the output and therefore publishes no price for it.

The memory ceiling is the other number that decides whether self-hosting is real for you. A Pro SD run peaks at 72.8 GiB, past every consumer card. The repository's 16 GB preset streams transformer blocks from host memory and quantizes the text encoder to NF4 — and the paper is explicit that this is the only preset change that alters the clip itself. If you reproduce someone's output on a 16 GB card and it does not match, that setting is the first thing to check.

What you can call today, and how that shapes the decision

Here is the asymmetry that matters more than any benchmark: Kling 4.0 is reachable through Kuaishou's own app and, increasingly, its API, while Kandinsky 6.0 Video is reachable by downloading it. Neither is a drop-in call from a routing layer.

What OrcaRouter does route is the generation Kling 4.0 replaces. Kling 3.0 is on the catalogue at $0.084 per request, and MiniMax H3 is there at $0.08 per second of generated video, both on the same OpenAI-compatible endpoint as more than 200 models, at provider list price with 0% markup. That matters for a self-hosted model in a way it does not for a hosted one: the prompt-expansion pass that turns a shot description into the long caption Kandinsky was trained on, the caption and transcript work that feeds an image input, the review summaries that decide which of four generations to keep — those are ordinary text calls, and they can share one key and one failover policy while the video heavy lifting runs wherever you put it. Kandinsky itself is not an OrcaRouter route, and nothing here should be read as claiming it is.

Which one belongs in your stack

The decision does not hinge on quality, because nobody has measured it fairly. It hinges on three questions you can answer yourself.

Do you need a shot longer than five seconds? Then Kandinsky 6.0 Video is out of the running for that shot, and no amount of MIT licensing changes it — you would be stitching, and stitching costs you the temporal coherence the model is good at.

Do you need to own the weights, run offline, fine-tune, or ship inside a product where a per-call vendor bill is unacceptable? Then Kling 4.0 is out of the running regardless of its spec sheet, because a closed API is a dependency, not an asset.

Do you need to make the call this week with a budget someone will sign? Then neither is clean. Kling 4.0's API is mid-rollout with preview pricing, and Kandinsky's true cost is a GPU-hour figure you have to translate into your own infrastructure. The models you can price and call today are the previous generation on both sides — Kling 3.0 through Kuaishou or a router, Kandinsky 6.0 Video from the Hub.

The thing to watch is not a spec. It is whether anyone runs both models on the same prompts and publishes the result — an independent arena entry, a lab that does not own either model, or a credible set of side-by-side clips with the prompts attached. Until that exists, any page that ranks Kling 4.0 against Kandinsky 6.0 Video with a winner is extrapolating from a comparison against Kling 2.6 and hoping you do not notice. The useful work in the meantime is narrower: pick the constraint you cannot move, and let it pick the model.