Kling 3.0: 4K AI Video Generator, 60fps & $0.075/s
Kling 3.0 (Kuaishou, Feb 2026): 4K native video at 60fps, 15s duration, built-in audio. API at $0.075/s, ELO 1248 on Artificial Analysis. Beats Runway and Veo.
Kling 3.0 is Kuaishou's video generation model, released February 5, 2026, producing native 4K (3840x2160) video at 60fps for up to 15 seconds with an ELO score of 1,248 on Artificial Analysis. Priced from $0.075 per second via API, it supports text-to-video, image-to-video, and multi-shot storyboarding across up to 6 shots with native audio in 5 languages.
Kling 3.0, released February 5, 2026 by Kuaishou Technology, is a proprietary AI video generation model producing native 4K (3840x2160) video at 60fps for up to 15 seconds with integrated audio in 5 languages. It costs $0.075 per second via third-party API and $6.99 per month at the Standard consumer tier. It holds an ELO score of 1,248 on the Artificial Analysis video generation leaderboard, with 4 model variants in the global top 10.
Provider: Kuaishou Technology · Family: Kling
Input modalities: text, image, video · Output: video, audio
About Kling 3.0
Kling 3.0 is a video generation model developed by Kuaishou Technology, released February 5, 2026. Built on a Multi-modal Visual Language (MVL) framework using a Diffusion Transformer (DiT) architecture, it processes text, reference images, and reference video in a single representation space to output video with synchronized native audio in one generation pass. It is the third major Kling model version, following Kling 2.6 (December 2025) and Kling 2.5 Turbo (September 2025). Within days of launch, Kling 3.0 took the top position on the Artificial Analysis video generation leaderboard with an ELO score of 1,248, ahead of Google Veo 3.1 and Runway Gen-4.5. Kling 3.0 Turbo, a faster lower-cost variant, followed in June 2026. Kling 3.0 holds an ELO score of 1,248 on the Artificial Analysis video generation arena, with 4 Kling 3.0 variants appearing in the global top 10 as of mid-2026. At launch it displaced Google Veo 3.1 and Runway Gen-4.5 from the top position. ByteDance's Seedance 2.0 (launched February 12, 2026) contests the top 2 positions with Kling 3.0, alternating in first and second on the Artificial Analysis board by month. Compared to Kling 2.6, Kling 3.0 improved across every quality axis: resolution from 1080p to native 4K, maximum duration from 10 seconds to 15 seconds, frame rate from 48fps to 60fps, and audio language support from 2 to 5 languages. No standard LLM benchmarks such as SWE-bench, GPQA, or MMLU-Pro apply to Kling 3.0, as it is a dedicated video generation model. Kling 3.0 generates video at up to 15 seconds per clip at native 4K (3840x2160) resolution and 60fps. Supported aspect ratios are 16:9, 9:16, and 1:1. In May 2026, Kuaishou enabled true native 4K output without upscaling, the first AI video platform to do so. Input context consists of a text prompt combined with an optional reference image and an optional reference video clip. The MVL architecture processes all input modalities in a shared latent space, maintaining visual and audio consistency across the full 15-second output duration. Kling 3.0 accepts text prompts, reference images, and reference video clips as input modalities. Output is an MP4 video file with synchronized audio embedded. The model co-generates synchronized audio in the same model pass rather than through a separate audio synthesis step, supporting Chinese, English, Japanese, Korean, and Spanish with lip-sync. The Video 3.0 Omni variant adds multi-shot storyboarding, allowing up to 6 sequential shots within a single generation request where each shot has individually specified duration, shot size, camera angle, and narrative content. Function calling, code execution, PDF input, and web browsing are not applicable to this model. Consumer access runs through four monthly tiers: Standard at $6.99/month (first month, $8.80 renewal), Pro at $25.99/month, Premier at $64.99/month, and Ultra at $127.99/month. A free consumer tier provides approximately 66 daily credits for short watermarked test generations. The developer API charges per second of video generated: official Kuaishou rates are approximately $0.11/s for 720P (¥0.8/s) and $0.14/s for 1080P (¥1/s); third-party resellers offer access from $0.075/s. A 15-second 1080P clip costs approximately $1.65 to $2.10 at the official API rate. Consumer credits and API credits are completely separate systems that cannot be transferred. Kling 3.0 is available via the consumer web platform at klingai.com and through the Kling Open Platform developer API at kling.ai/dev. The model is API-only with no self-hosting or open-weights option. Authentication uses API key with prepaid credit packages. Third-party integrations include ComfyUI plugins and API resellers such as EvoLink. The service runs from Kuaishou's China-based infrastructure in Inner Mongolia. AWS Bedrock, Google Vertex AI, and Azure marketplace listings are not available for Kling 3.0 as of June 2026. Kling 3.0 applies content moderation required under China's Generative AI Service Management Provisions, mandating filtering of content deemed harmful or politically sensitive under PRC law. The model watermarks all generated video to signal AI origin per Chinese regulatory requirements. No published system card, independent red-team evaluation, or Western-style safety assessment exists for Kling 3.0 as of June 2026. The model declines to generate explicit adult content, graphic violence, and harmful depictions of real individuals. No safety audits from METR, Apollo Research, or AISI have been published. Kling 3.0 is best for creative professionals, social media marketers, and film studios that need high-quality 4K video at a cost-effective API price. Multi-shot storyboarding in Video 3.0 Omni is particularly useful for generating coherent multi-scene ad sequences in a single API call. Teams requiring GDPR compliance documentation, SOC 2 Type II, or HIPAA eligibility should not use Kling 3.0 for production workloads involving customer personal data; Runway ML or Google Veo 3 via Google Cloud are better alternatives for regulated industries. Real-time video applications are also poorly served, as generation takes 2 to 5 minutes per clip even via API.
Pricing
Video generation is priced per second of output, not per token. Official Kuaishou API: ~$0.11/s (¥0.8/s) for 720P, ~$0.14/s (¥1/s) for 1080P. Third-party resellers from $0.075/s. Consumer plans: Standard $6.99/mo, Pro $25.99/mo, Premier $64.99/mo, Ultra $127.99/mo. Consumer and API credits are fully separate systems.
Key Features
- Native 4K at 60fps: Generates 3840x2160 video at up to 60fps without upscaling — the highest native resolution and frame rate of any AI video model at its February 2026 launch.
- Integrated Audio in 5 Languages: Co-generates synchronized audio with lip-sync in Chinese, English, Japanese, Korean, and Spanish in the same model pass, removing the need for a separate audio synthesis tool.
- Multi-Shot Storyboarding: Video 3.0 Omni accepts up to 6 sequential shots per API call, each with individually specified duration, shot size, camera angle, and narrative, generating coherent multi-scene video in one request.
- Unified MVL Architecture: Multi-modal Visual Language framework on a Diffusion Transformer backbone processes text, image, and audio in a shared latent space, producing higher visual-audio consistency than pipeline-based models.
- Affordable API Access: Official API costs $0.11/s (720P) to $0.14/s (1080P); third-party resellers offer access from $0.075/s, making a 15-second clip cost $1.13 to $2.10 per generation.
Pros
- Native 4K at 60fps leads the AI video generation market on resolution and frame rate as of the February 2026 launch.
- Audio in 5 languages generated natively in one pass eliminates a separate dubbing step from AI video workflows.
- API pricing from $0.075/s to $0.14/s undercuts Runway Gen-4.5 per second while delivering equal or higher resolution.
- Free consumer tier (66 daily credits) allows no-cost testing of the model before committing to a subscription.
Cons
- 2 to 5 minute generation time per clip rules out real-time or low-latency video applications entirely.
- No GDPR DPA, SOC 2 Type II, or HIPAA eligibility; enterprise teams in regulated US and EU industries cannot use Kling 3.0 for production data.
- Consumer credits and API credits are separate systems; a $127.99/mo Ultra subscription grants zero API access.
- Non-English lip-sync quality is noticeably lower than Chinese and English for Korean and Spanish, particularly on longer clips.
Benchmarks
- lmarena elo: 1248
Frequently Asked Questions
What is Kling 3.0 and who built it?
Kling 3.0 is a video generation model built by Kuaishou Technology (HKEX:1024), a publicly listed Chinese tech company, released on February 5, 2026. It runs on a Multi-modal Visual Language (MVL) framework using a Diffusion Transformer (DiT) architecture that processes text, reference images, and reference video in one shared generation pipeline. Kling 3.0 produces native 4K (3840x2160) video at up to 60fps and up to 15 seconds per clip, with synchronized audio in Chinese, English, Japanese, Korean, and Spanish generated in the same model pass. At launch it reached an ELO score of 1,248 on the Artificial Analysis video generation leaderboard, displacing Google Veo 3.1 and Runway Gen-4.5 from the top position. Consumer access starts at $6.99/month (Standard tier) at klingai.com; API access from $0.075/s at kling.ai/dev. Kuaishou reported Kling AI Q1 2026 revenue of CN¥650 million+, more than 3x year-over-year. A faster variant, Kling 3.0 Turbo, launched in June 2026.
How much does Kling 3.0 cost in 2026?
Kling 3.0 offers two separate pricing tracks: consumer subscriptions and developer API credits, which cannot be mixed. Consumer tiers are Standard at $6.99/month (first month, $8.80 renewal), Pro at $25.99/month, Premier at $64.99/month, and Ultra at $127.99/month; each tier provides a monthly credit allowance sized to the subscription level. There is a free tier with approximately 66 daily credits for short watermarked test videos. Developer API pricing is per second of video generated: official Kuaishou rates are approximately $0.11/s for 720P (¥0.8/s) and $0.14/s for 1080P (¥1/s); third-party resellers offer access from $0.075/s. A 15-second 1080P clip costs approximately $1.65 to $2.10 via the official API. Failed API generation tasks do not consume credits, reducing wasted cost on bad prompts. Annual plans for consumer tiers offer slight discounts (e.g. Standard at $6.60/month billed annually).
What is Kling 3.0's maximum video duration and output quality?
Kling 3.0 generates video at up to 15 seconds per clip, upgraded from the 10-second maximum of Kling 2.6. Output resolution is native 4K (3840x2160), the first AI video model to offer this without upscaling, a capability enabled in May 2026. Frame rate is up to 60fps, compared to 48fps in Kling 2.6. Supported aspect ratios are 16:9, 9:16, and 1:1. The Video 3.0 Omni variant extends effective output via multi-shot storyboarding: up to 6 sequential shots in a single API call, each with individually specified duration, so total output can exceed 15 seconds through shot stitching. Unlike language models, Kling 3.0 does not use a token context window; input context is a text prompt plus optional reference image and reference video. Compared to Runway Gen-4.5 (max 16s clips), Kling 3.0's resolution advantage is the main differentiator.
How does Kling 3.0 compare on benchmarks vs Runway Gen-4.5?
On the Artificial Analysis video generation leaderboard, Kling 3.0 holds an ELO of 1,248, placing it in the global top 5 as of mid-2026, while Runway Gen-4.5 dropped out of the Artificial Analysis top 10 after its late 2025 ELO peak of 1,247. Kling 3.0 beats Runway Gen-4.5 on resolution (native 4K vs. 1080p), frame rate (60fps vs. lower), and integrated audio (Kling includes 5 languages natively; Runway requires a separate audio step). API pricing also favors Kling: $0.075 to $0.14/s vs. Runway's higher per-second rates at comparable quality. Runway wins on Western enterprise compliance: it is accessible through cloud platforms with SOC 2 documentation, while Kling 3.0 has no published Western compliance certifications. Google Veo 3.1 (#3 on the Artificial Analysis with-audio board) and ByteDance Seedance 2.0 (#1-#2 alternating with Kling) are stronger quality competitors to Kling 3.0 than Runway is as of June 2026. Practical conclusion: Kling 3.0 leads on cost and 4K resolution; Runway leads on Western enterprise trust; Google Veo leads on enterprise cloud integration depth.
Is Kling 3.0 open source or proprietary?
Kling 3.0 is fully proprietary. The model weights are not publicly available and no open-source license (Apache, MIT, or otherwise) has been applied to any Kling model version. Access is exclusively through the consumer platform at klingai.com or the Kling developer API at kling.ai/dev. There is no self-hosting path, no Docker image, no VRAM requirement, and no Hugging Face weights repository for Kling 3.0. The model is built on a proprietary MVL (Multi-modal Visual Language) Diffusion Transformer framework developed internally by Kuaishou's KwaiVGI team. Third-party API resellers (such as EvoLink) route requests through Kuaishou's own infrastructure. Teams needing self-hosted video generation can consider CogVideoX (Zhipu AI, Apache 2.0) or Wan (Alibaba, open weights) as alternatives, though both trail Kling 3.0 on output quality benchmarks.
What modalities does Kling 3.0 support?
Kling 3.0 accepts three input modalities: text prompts describing the scene and camera movement, reference images for image-to-video conversion or visual style reference, and reference video clips for motion transfer or subject consistency. Output is a video file with synchronized audio embedded as a single MP4. The model natively co-generates audio in Chinese, English, Japanese, Korean, and Spanish with lip-sync in the same model pass, without a separate audio pipeline. The Video 3.0 Omni variant adds multi-shot storyboarding input: users define up to 6 sequential shots each with duration, shot size, camera angle, and narrative content. Output resolution is up to 4K (3840x2160) at up to 60fps and up to 15 seconds per clip. Function calling, code execution, PDF input, and web browsing are not applicable; Kling 3.0 is a dedicated video generation model, not a general-purpose language model.
Does Kling 3.0 train on user data?
Kuaishou Technology has not published a public API data retention policy or training-on-inputs disclosure for Kling AI as of June 2026. The consumer privacy policy is at klingai.com/privacy but does not specify whether API inputs (text prompts, reference images, reference video) are retained for model training. No GDPR data processing agreement (DPA), standard contractual clauses, or enterprise zero-retention option has been publicly documented for the Kling API, making it impossible for EU-based companies to confirm GDPR-compliant processing. For Chinese users, data handling is governed by China's PIPL and Cybersecurity Law, which mandate data minimization and user consent but do not map to Western compliance frameworks. Teams handling personal data of EU residents should not use Kling 3.0 as a production data processor without explicit written agreements. US federal teams or organizations with FedRAMP, ITAR, or HIPAA requirements should also avoid Kling 3.0 for regulated workloads. For comparison, Runway ML and Google Veo 3 (via Google Cloud) provide explicit compliance documentation for enterprise customers.
Who is Kling 3.0 best for and who should avoid it?
Kling 3.0 is best for creative professionals and agencies producing high-quality video content at scale: social media marketers who need cinematic 4K ad sequences at $0.075 to $0.14/s, film studios using AI for B-roll and pre-visualization at 60fps quality, and multilingual content teams who need audio in 5 languages without a separate dubbing step. The Video 3.0 Omni multi-shot storyboarding feature makes it uniquely strong for agencies generating complete 6-shot ad sequences in a single API call. Teams that should avoid Kling 3.0: any organization with GDPR compliance requirements (no DPA available), real-time video applications (2 to 5 minute generation time per clip), US federal or defense teams (no FedRAMP authorization), and developers who need self-hosting or open weights (Kling 3.0 is API-only). For regulated industries in the US and EU, Runway ML or Google Veo 3 via Vertex AI are better-suited alternatives with published compliance documentation.