Kuaishou flipped the switch on the Kling 4.0 API on 28 October. The three new endpoints are /text-to-video/kling-4.0, /image-to-video/kling-4.0, and /omni-video/kling-4.0. They replace the 3.0 paths for anyone who wants native 30-second output without stitching.
The headline number is duration. Kling 3.0 topped out at 15 seconds per generation. Kling 4.0 doubles that to 30 in a single pass. Resolution options at launch are 720p and 1080p, with 10-bit 4K listed as coming soon. Aspect ratios now include 21:9 and an adaptive mode that picks the best fit from your inputs.
Keyframes and references
You can pass up to 10 keyframe images into one task. That is enough to pin a camera move, a character pose at the midpoint, and an end frame without splitting the job. The Omni reference-to-video endpoint also accepts voice references directly, so a dialogue clip can carry its own audio target instead of relying on a separate lip-sync pass.
Input limits went up across the board. Prompts accept 26,000 English characters. A single task takes up to 10 reference images and 5 reference videos. The Omni endpoint combines multiple asset types in one call and caps total inputs at 15. Fewer mutual exclusions means fewer preflight checks in your integration code.
Billing follows duration, not credits
Kling 4.0 drops the credit-per-generation model. Billing is now based on total output duration, and the calculation includes any input video you provide. Reference videos and video character elements count toward that total. The pricing page breaks it down by resolution and mode; check it before you ship a pipeline that assumes 3.0 credit math still applies.
This matters for budgets. A 30-second 1080p clip costs more than two 15-second 3.0 clips at the old rate, but you save the seam work and the color match between generations. For sequences that need continuity, the trade favors 4.0. For social variants where 10 seconds is enough, 3.0 Turbo remains cheaper per second.
Where this fits in the current roster
In our October roundup, Kling 3.0 was the workhorse and Kling 4.0 was upside. As of today, 4.0 is the workhorse for anything over 15 seconds. Gemini Omni Flash still wins on cost for short social clips. Veo 3.1 Standard still owns 4K hero shots with native 48 kHz dialogue. Seedance 2.5 still handles the longest single takes and the heaviest reference sets. MiniMax H3 remains the open-weight option for self-hosted pipelines.
What 4.0 changes is the middle ground. Teams that were stitching two or three 3.0 generations to get a 30-second sequence can now do it in one call with shared keyframes and a single audio pass. That cuts post time and removes the most visible seams. It does not replace the other models. It replaces the workaround.
What to test first
- Continuity. Run the same scene as one 30-second 4.0 generation and as two stitched 3.0 generations. Compare the seam at the cut point and the face consistency across the full duration.
- Keyframe control. Pin a start, midpoint, and end frame. Check whether the model honors all three or drifts between them. Ten keyframes is a ceiling, not a recommendation.
- Voice reference. Use the Omni endpoint with a voice sample instead of a separate lip-sync step. Measure sync accuracy against your existing two-pass workflow.
- Cost at scale. Price a real project end-to-end under the new duration billing. Include input video duration. The old credit estimator will lie to you.
We will publish test results on our own briefs in a follow-up. Until then, treat the official API updates log as the source of truth. The community tracker on GitHub fills gaps between official posts but is not a substitute.
