Published Updated
In one or two sentences, before anything else.
Kling 4.0 is Kuaishou's video model for clips of up to 30 seconds with stereo sound in one generation, fully released on Kling AI on October 15, 2026. Public tests credit its long takes, camera moves and lip sync; it is slow, adds details nobody asked for, garbles on-screen text and has no public leaderboard score yet.
What it can do, one sentence each, and whether we actually tested it.
Generates 3 to 30 seconds in one pass, adjustable in 1-second steps, and splits the clip into shots itself unless the prompt says otherwise.
SandBase's 30-second test came back at 30.04 seconds with no cuts and all four timed beats in order.
Places up to 10 frames anywhere on the timeline, including first, middle and last frames, for the clip to pass through.
Kling's user guide says keyframes cannot be combined with Omni Reference in the same task.
Takes up to 15 references in one task: 10 images, 5 videos of up to 30 seconds in total, 7 saved subjects and 3 voice profiles.
MindStudio saw it add a grey jacket that was in neither the reference nor the prompt.
Generates dialogue, effects and ambience in two-channel stereo, with lip sync in Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French and Hindi, plus dialects such as Cantonese.
Sound is always on: Kling's guide lists no switch to turn it off, and the API has none either.
Edits up to 5 clips of 30 seconds combined: adds, swaps or removes people and objects, and changes expressions, camera, style or background.
In 光子星球's test all four edits ran; the man it added looked stiff and the background jumped during the style change.
A faster, cheaper version at 720p and 3 to 20 seconds, without first-and-last-frame or multi-keyframe control.
Users report 6 credits a second for Flash, against about 15 for the full model at 1080p.
Lets AI agents start Omni Reference and keyframe generations through Kling's MCP server and command-line tool.
Available from launch day, per Kling's release note.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
A night-market wok stall filmed as one handheld take, with four actions at set moments; one run at 720p by SandBase, a reseller of the model, on October 23, 2026.
What happened
The clip came back at 30.04 seconds with no cuts and all four beats in order, the camera arriving behind the cook a little early. It cost 4.71 USD and took about 498 seconds. The neon signs came out as unreadable pseudo-Chinese.
Case 2
Input
Dialogue in Northeastern Mandarin and in Cantonese, tested by the Chinese tech outlet 光子星球 on October 16, 2026.
What happened
Both stayed in dialect and the lips matched the words one by one. In a separate 5-second test the shouting on the left and the television on the right were barely audible, so background sound still needs work in post.
Case 3
Input
One prompt, one run each, 720p and 5 seconds where possible: Kling 4.0, Kling 3.0 Pro, Seedance 2.5, Wan 3.0 and Veo 3.1 (SandBase, October 23, 2026).
What happened
Kling 4.0 was the only model to perform the half orbit during the ollie. It also took longest, 339 seconds against 51 to 175 for the others, and added two unexplained human shadows near the end.
Case 4
Input
Two 10-second clips, one guided by 10 keyframes and one by a single photo of the 16-year-old (光子星球, October 16, 2026).
What happened
With 10 keyframes every age appeared in order with smooth transitions; from one photo the ageing lagged and then jumped. The more frames it is given, the more closely it follows them.
What it does badly, with the evidence, and what we could not settle.
Weak
Slow to generate.
Evidence: SandBase waited 339 seconds for a 5-second text-to-video clip and about 498 seconds for a 30-second one. In the same test Veo 3.1 returned a 4-second clip in 51 seconds, and Seedance 2.5 and Kling 3.0 Pro took about 150 seconds for 5.
Adds things nobody asked for and slips on continuity.
Evidence: MindStudio saw an unprompted gesture and a grey jacket that was in neither reference nor prompt; SandBase saw the sea switch sides halfway through an orbit; Atlas Cloud saw a shoe's tread flatten across cuts in Kling 4.0 Flash.
Text inside the scene is unreliable.
Evidence: Both SandBase tests rendered sign and neon text as unreadable pseudo-characters, although Kling lists on-screen text in nine languages as a feature. The 21st Century Business Herald also collected user complaints about garbled text.
Simplifies hard actions to keep the shot together.
Evidence: In 光子星球's 30-second chase, vaulting a bicycle became stepping over it and breaking through a metal door became walking through an open one, so the chase lost its tension.
Not for anyone who needs 4K, HDR or legible signs and captions in the picture today.
Evidence: Kling marks 10-bit HDR and 2-minute extensions as coming soon, the API and the hosts that publish prices stop at 1080p, and no public test has produced readable sign text.
A clip that comes out wrong is still paid for.
Evidence: Kling's credits policy refunds credits only when a generation fails. On Trustpilot, where Kling AI has a TrustScore of 1.2 from 420 reviews, recent reviewers complain of retries charged again and credits expiring, though none of them names Kling 4.0.
Uncertain
Where it ranks against other video models.
Why unsure: Kling 4.0 has no row on any Artificial Analysis video leaderboard yet, which promptowy puts down to it having no API at launch. The closest stand-in, Kling 3.0 Omni, sits 138 Elo points below Seedance 2.5 on text to video.
Whether it beats Seedance 2.5.
Why unsure: Every public head-to-head so far is a single run or a set of impressions, and they disagree: Atlas Cloud and AI Samson saw Seedance ahead on realism and detail, while SandBase saw only Kling 4.0 perform the camera move it was asked for.
What a usable clip really costs.
Why unsure: Nobody has published how many tries a usable Kling 4.0 clip takes, and Kling shows the per-second credit cost only on its membership page.
When 4K and the full API arrive.
Why unsure: Kling's release note lists 4K while its Kling 4.0 page calls 4K coming soon. The API opens to everyone on October 28, 2026, and its price is still marked as a preview.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Released | October 15, 2026 for everyone; Kling 4.0 Flash in early access from September 28, 2026 | Source: kling.aichecked |
| Clip length | 3 to 30 seconds per generation in 1-second steps (Flash: 3 to 20 seconds) | Source: kling.aichecked |
| Resolution | 720p and 1080p; 4K listed, with 4K and 10-bit HDR marked coming soon | The API and the hosts that publish prices offer 720p and 1080p only. Source: kling.aichecked |
| Frame rate | 24 | Stated in the API documentation; the app does not state one. Source: kling.aichecked |
| Aspect ratios | 21:9, 16:9, 1:1 and 9:16 | Source: kling.aichecked |
| Sound | Two-channel stereo with lip sync, always on | Source: kling.aichecked |
| Keyframes | Up to 10 per clip; cannot be combined with Omni Reference | Source: kling.aichecked |
| Omni Reference | Up to 15 inputs: 10 images, 5 videos (30 seconds in total), 7 subjects, 3 voice profiles | Source: kling.aichecked |
| Prompt length | Up to 8,000 tokens in the app | Source: kling.aichecked |
| API price (preview) | 0.14 USD per second at 720p, 0.21 USD at 1080p | Without video input. Kling marks the price as a preview, not final. Source: kling.aichecked |
| API availability | Early access now; open to everyone on October 28, 2026 at 23:00 Beijing time | Source: kling.aichecked |
| Credits in the Kling app | About 15 credits per second at 1080p (450 for 30 seconds) and 10 at 720p | Reported by resellers and users; Kling shows the rate only on its membership page. Source: kie.aichecked |
| Credit price | 66 credits per USD at the standard top-up price; refunded if a generation fails | Source: kling.aichecked |
| Commercial use | Paying members only; free users' videos must carry Kling branding | Source: kling.aichecked |
Answer first, then why, an example, the limits and what to do next.
A 30-second 1080p clip takes about 450 credits in the Kling app, roughly 6.82 USD at the standard top-up price of 66 credits per dollar. Users and resellers report 15 credits a second at 1080p and 10 at 720p, because Kling shows the rate only on its membership page. Through the API the preview price is 0.21 USD a second at 1080p, or 6.30 USD for 30 seconds.
Kling has not said whether free accounts can use Kling 4.0 or Kling 4.0 Flash. Free accounts get no monthly credits, cannot use their videos commercially and must keep Kling's branding on them, while 1080p and watermark removal are for paying members.
Neither model wins everywhere in the public tests so far. Kling 4.0 makes longer single takes, follows camera moves and keyframes more closely and costs about a quarter as much per second through the same reseller. Seedance 2.5 kept product shapes and fine detail better in Atlas Cloud's and AI Samson's comparisons, and sits near the top of Artificial Analysis's leaderboards, where Kling 4.0 is not listed yet.
Kling's official API is in early access and opens to everyone on October 28, 2026 at 23:00 Beijing time. Its preview price is 0.14 USD a second at 720p and 0.21 USD at 1080p. WaveSpeedAI and Kie.ai already sell Kling 4.0, and fal shows it in its playground with API access coming soon.
Kling has announced 4K but nothing shows it working yet. Its release note lists 4K, its Kling 4.0 page calls 4K and 10-bit HDR coming soon, and the API and the hosts that publish prices offer only 720p and 1080p.
Kling 4.0 Flash is a faster, cheaper draft version at 720p and 3 to 20 seconds, without first-and-last-frame or multi-keyframe control. Creators who ran the same prompt on both found the full model smoother and sharper, with better lip sync, and users report Flash at 6 credits a second against 15 for the full model at 1080p.
Paying members can use their videos commercially without limits, except to build a competing product, under Kling's paid-service terms of October 10, 2026. Free users need Kling's written permission and must keep its branding on the video.
This review is built from what can be checked on October 23, 2026: Kling's own release note, user guide and API documentation, the Artificial Analysis video leaderboards, about a dozen public tests in English and Chinese, and what creators posted on X since Kling 4.0 Flash opened on September 28. Each piece of evidence is labelled by who produced it, because much of it is not neutral: the most-watched English videos were sponsored by Kling, and the most measured tests come from resellers who sell the model.
Kling 4.0 doubles the length of a single generation and replaces start-and-end frames with up to 10 keyframes. Kling's own comparison lists the rest:
| Kling 3.0 | Kling 4.0 | |
|---|---|---|
| Length per generation | 3 to 15 seconds | 3 to 30 seconds |
| Frame control | First and last frame | Up to 10 keyframes |
| References | Up to 7 images or subjects, or 1 video of up to 15 seconds plus 4 | Up to 15: 10 images, 5 videos, 7 subjects |
| Sound | Single channel | Two-channel stereo |
| Speech languages | Chinese, English, Japanese, Korean, Spanish | Adds Portuguese, German, French, Hindi and more |
| Dynamic range | 8-bit SDR | 10-bit HDR announced, not live yet |
| API price at 1080p with sound | 0.168 USD a second | 0.21 USD a second (preview) |
Sources: Kling 4.0 vs Kling 3.0 and Kling's API pricing page. Kling 3.0 already outputs native 4K in the app; for Kling 4.0, 4K is listed in the release note but marked coming soon on the Kling 4.0 page.
Since October 15 Kling 4.0 has been open on kling.ai and klingai.com, and AI agents can drive it through Kling's MCP server and command-line tool (release note, Kling AI on X). Live today: 3 to 30 seconds, 720p and 1080p, 21:9 among the aspect ratios, 10 keyframes, 15 references, editing of existing clips and stereo sound with lip sync.
Still coming: 10-bit HDR, extending a video several times to 2 minutes, 4K (Kling's pages contradict each other), and the full API, which Kling says opens on October 28, 2026 at 23:00 Beijing time (API documentation). On October 23 the API offered no switch to turn sound off, no seed or negative prompt and no 4K or HDR, according to SandBase's API test.
Kling 4.0 is not on any public video leaderboard yet. Artificial Analysis ranks models by blind votes, and an entry needs an API and several days of votes; Kling 4.0 had no API when it launched (promptowy). The board still shows where Kling stood before 4.0:
| Model | Text to video with sound (Elo) | Rank range | Image to video (Elo) |
|---|---|---|---|
| Wan 3.0 | 1175 | 1 to 2 | 1127 |
| Seedance 2.5 | 1158 | 2 to 8 | 1138 |
| Kling 3.0 Omni (1080p) | 1020 | 16 to 17 | 1004 |
| Kling 3.0 (1080p) | 1000 (the baseline) | 18 to 19 | 1000 |
| Veo 3.1 | 963 | 21 to 24 | 1028 |
| Kling 4.0 | Not listed | Not listed |
Sources: text-to-video leaderboard and image-to-video leaderboard, read October 23, 2026. Kling 4.0 starts from 138 points behind Seedance 2.5 if it scores like Kling 3.0 Omni; whether it closes that gap will only show once it is on the board.
Long takes hold together. SandBase asked for a 30-second handheld take at a night-market wok stall with four actions at set moments; the clip came back at 30.04 seconds with no cuts and all four beats in order. 光子星球's 30-second snowy-night chase through Harbin kept the runner's face, hair and clothes the same from start to end, and the camera swung around a pillar from behind to in front without a break (光子星球).
Camera moves are followed. On one prompt asking for a half orbit during a skateboard ollie, Kling 4.0 was the only one of five models to do the orbit; Kling 3.0 Pro, Seedance 2.5, Wan 3.0 and Veo 3.1 all kept the camera on one side. 数字生命卡兹克, who reports spending tens of thousands of credits, praised a figure-skating jump filmed from a very low angle and a rooftop basketball step-back with a clean swish (数字生命卡兹克); APPSO saw no jitter or flicker in push, pan, tilt and tracking shots (APPSO). Neither article reports a failure, so treat them as best cases.
Dialogue and lip sync improved. MindStudio found the lip-sync drift of Kling 3.0 largely fixed in a full diner conversation (MindStudio). 光子星球's Northeastern Mandarin and Cantonese lines matched word for word. Background sound is weaker: in a 5-second scene the shouting on the left and the television on the right were barely audible, and the stereo placement was unclear.
More frames, more control. Aging a character from 16 to 80 with 10 keyframes produced every stage in order; from a single photo the change lagged and then jumped. MindStudio also got better results from keyframes plus references than from references alone. One catch from Kling's own guide: keyframes and Omni Reference cannot be used in the same task.
Where it slips. Kling 4.0 adds things: MindStudio saw an unprompted gesture and a grey jacket in neither reference nor prompt, SandBase saw two human shadows appear from nowhere and the sea switch sides mid-orbit. It simplifies hard actions to keep the shot intact: in 光子星球's chase, vaulting a bicycle became stepping over it and breaking through a metal door became walking through an open one. Text in the scene is unreliable, with sign and neon lettering garbled in both SandBase tests. The 21st Century Business Herald also collected complaints of overdone skin sheen, props from uploads not reproduced and odd body movement.
It is slow. SandBase waited 339 seconds for a 5-second text-to-video clip, 248 seconds for image to video and about 498 seconds for 30 seconds. On the same day Seedance 2.5 and Kling 3.0 Pro took about 150 seconds for 5 seconds and Veo 3.1 took 51 seconds for 4. These are single readings from launch week, not an average.
The comparison people search for most has no clear winner yet, and no blind or repeated test.
| Who ran it | Version | Setup | Result |
|---|---|---|---|
| SandBase, sells both | Kling 4.0 | Same prompt, one run, 720p, 5 seconds | Only Kling 4.0 did the orbit; Seedance 2.5 landed the ollie late with heavy motion blur |
| Atlas Cloud, sells Seedance 2.5 | Kling 4.0 Flash | Same prompts, one run, three tests | Seedance kept a shoe's shape and won the dialogue scene; neither handled the keyframe walk |
| AI Samson, sponsored by Artlist, which sells Kling 4.0 | Kling 4.0 | Side-by-side tests in a YouTube video | Kling ahead on 30-second length, keyframes and price; Seedance ahead on realism and detail in several tests |
Sources: SandBase, Atlas Cloud, AI Samson. Kie.ai, which sells Kling 4.0, relays testers who disagree: one Japanese creator found Kling stronger on motion and impact and Seedance steadier overall (Kie.ai). On X the AI filmmaker Dave Clark posted first tests cutting Kling 4.0 and Seedance 2.5 into one piece rather than choosing between them (@Diesol).
Price is the clearest difference. Through the same reseller on October 23, Kling 4.0 cost 0.235 USD a second at 1080p and Seedance 2.5 cost 0.90 USD, so a 30-second 1080p clip is about 7 USD against 27 USD. Pick Kling 4.0 for long takes, camera moves and dialogue on a budget; pick Seedance 2.5 when product shapes and fine detail must survive every cut.
Kling does not publish the credit cost of Kling 4.0 anywhere but its membership page. Resellers and users report 15 credits a second at 1080p and 10 at 720p, so 450 credits for a 30-second 1080p clip (Kie.ai, 数字生命卡兹克); a creator in early access saw 540 for the same clip, and Kingy AI's sponsored video showed 468 or 540 depending on settings (@onofumi_AI, Kingy AI). Kling 4.0 Flash is reported at 6 credits a second, 120 for 20 seconds at 720p.
Kling sells credits at 66 per US dollar at the standard top-up price (credits policy). Its paid plans start at Standard, 8.80 USD a month with 660 credits, and Pro, 32.56 USD with 3,000 credits (Kling's price list, unchanged in TechPilot's October 23 record). Pro's monthly credits cover six 30-second 1080p clips, about 4.88 USD each, before any rerolls. Credits come back only when a generation fails, not when it comes out wrong.
One 30-second 1080p clip by route, checked October 23, 2026:
| Route | Price per second at 1080p | 30 seconds |
|---|---|---|
| Kling app, standard top-up credits | about 15 credits | 450 credits, about 6.82 USD |
| Kling API, preview price | 0.21 USD | 6.30 USD |
| Kie.ai | 0.2125 USD | 6.38 USD |
| WaveSpeedAI | 0.235 USD | 7.05 USD |
| fal, playground only so far | 0.235 USD | 7.05 USD |
| Higgsfield, Plus plan | 7.5 credits | 225 credits, about 11.03 USD |
Sources: Kling API pricing, Kie.ai, WaveSpeedAI, fal, Higgsfield. Kling 4.0 is also available inside Adobe Firefly since October 22 (Kling AI on X).
Flash is the draft model: 720p, 3 to 20 seconds, 8-bit SDR, and no first-and-last-frame or multi-keyframe control. On X, a Japanese creator ran one prompt on both and found the full model smoother, while Flash's 20-second cap made the same action look fast-forwarded (@towya_aillust); an invited tester found the full model clearly sharper with more precise lip sync (@hq4ai). 数字生命卡兹克 gives the same advice: find the shot with Flash, then render it with the full model.
Kling opened Kling 4.0 Flash to Ultra yearly subscribers on September 28 and announced full availability on October 15, with HDR and the API still coming (Kling AI). Posts from people who tried it agree on a few things. Lip sync is the most praised: one Chinese user muted a clip and could read about 90% of the words from the actor's mouth (@huangyun_122), and a Japanese creator built two speaking characters from one image and a voice and found the voice less artificial (@akiyoshisan). Seedance 2.5 is the model everyone compares against, as in a same-prompt test at 720p and 20 seconds that left the verdict to followers (@sebatheepan).
Many launch-week posts shared Kling's campaign wording or carried #ad, so this page counts them as promotion, not tests. An hour after the launch one creator remarked on seeing almost no Kling 4.0 demos and asked whether people were tired of video generation (Adam Holter). Hosts were slower than the app: fal announced Kling 4 as coming soon on launch day (fal on X).
Paying members can use Kling 4.0 videos commercially, except to build a competing product, and can remove the watermark (Terms of Paid Service). Free users cannot use output commercially without written permission and must keep Kling's branding on it (Terms of Service). The consumer terms also let Kling use what you upload to train its models; you can withdraw that by emailing support@kling.ai. Kling's API privacy policy says API customers' personal information is not used for model training.
Skip Kling 4.0 for now if you need 4K, HDR, readable text in the picture or product shots that keep their exact shape across cuts; Seedance 2.5 did better on the last in the one test that checked it. Developers should check what the full API includes when it opens on October 28: in launch week it had no switch to turn sound off and no 4K or HDR. If you iterate fast, the wait per clip matters: a 5-second clip took over five minutes in the one test that timed it.
Sources
Changelog
Higgsfield AI is a paid web platform from Higgsfield Inc. that runs more than 30 image and video models from other companies, including Veo 3.1, Kling 3.0 and Seedance, beside its own tools, with a limited free tier and paid credit plans from US$19 a month. In February 2026 it drew public complaints over 'unlimited' plans and creator payouts.
VideoEnhancer is a web tool that upscales short MP4 clips to 1080p, 2K or 4K with the FlashVSR model and also compresses, face-swaps and generates video. After a 5-second free trial, 1080p costs 3 credits a second, $0.03 to $0.09 depending on how credits are bought. VideoEnhancer is our own product: it is made by the team behind GenGrades.
LayerGrab splits a flat image into up to 16 named, transparent layers over a filled-in background, on the web, in Figma or in Chrome, or pulls out only what you name. Each layer costs $0.10 on its $12 plan, $0.08 on Premium. No independent review exists yet. LayerGrab is our own product: it is made by the team behind GenGrades.
VideoTextRemover is a web tool that erases burned-in subtitles, captions and other on-screen text from MP4 and MOV videos and repaints what was behind them, up to 1080p. It bills per second processed, about $0.025 to $0.049, with 30 free seconds a month in some countries. VideoTextRemover is our own product: it is made by the team behind GenGrades.
MagicRemover is a browser tool that erases people, objects, text, logos and watermarks from photos, and subtitles from videos. Photos get 2 free removals a day, 2 more after sign-in, at about 1 megapixel; beyond that, credits start at $9.90 for 330. MagicRemover is our own product: it is made by the team behind GenGrades.
Hermes Agent is the best OpenClaw alternative for most people: free, open source, safer by default and able to import an OpenClaw setup in one command. NanoClaw puts every session in a container, ZeroClaw is a tiny sandboxed binary, Claude Cowork suits people who would rather pay than host, and Hostinger is the cheapest way to keep OpenClaw itself.