Kling 3.0, Turbo, and Omni API: The Missing Manual (Auth, Model IDs, Real Per-Second Pricing)

Kling 3.0 is a serious production video API with a documentation problem: the docs portal is a JavaScript app that serves an error to plain fetchers, the purchase pages sit behind a login, and most tutorials in circulation describe an authentication scheme that is now legacy. The result is a market where resellers quote anywhere from $0.10 to $0.22 per second for the same model and nobody can easily check.

We operate video APIs in production for a living, so we assembled the manual we wanted: every number below comes from Kuaishou's official documentation, and the full set was re-verified by an independent second pass before publishing. Where something is only implied or sits behind the login wall, we say so instead of guessing.

First, figure out which API you are on

Two API generations coexist, and mixing their conventions is the classic first bug.

The current API ("API 2.0", released June 17, 2026): auth is a plain API key sent as Authorization: Bearer <key>, created in the console at kling.ai/dev. The model ID lives in the URL path (/image-to-video/kling-3.0), the international base domain is api-singapore.klingai.com, images can be up to 50MB, and a successful task reports succeeded.

The legacy API (/v1/videos/*): auth is an Access-Key/Secret-Key self-signed JWT, the model goes in a model_name body field whose enum stops at kling-v3 — Turbo and Omni do not exist there — images cap at 10MB, Base64 images must be raw strings without the data:image/... prefix, and success is spelled succeed. Kuaishou says there are no current plans to deprecate it, which means five years of JWT tutorials will keep ranking. If you are starting today, start on the current API.

The model picker

Model ID	Duration	Resolutions	Audio
kling-3.0-turbo	3–15s	720p / 1080p	Native, included
kling-v3	3–15s	720p / 1080p / 4K	native / off
kling-v3-omni	3–15s	720p / 1080p / 4K	native / off / original
kling-video-o1	3–10s	up to 1080p	—
kling-v2-6	3–10s	up to 1080p	—
kling-v2-5-turbo	5 or 10s	up to 1080p	—

Native audio on the 3.0 family speaks Chinese, English, Japanese, Korean and Spanish; ask for any other language and the speech is silently rendered in English. For other markets, generate with audio off and dub separately.

Turbo is deliberately the cut-down tier: first-frame input only, no last frame, no elements, no 4K, and audio always included at one price row. Do not build a frame-interpolation workflow on it.

The real prices, decoded

Official pricing is per second of output, in Units, at a printed $0.14 per Unit:

Model	720p	1080p	4K
3.0 Turbo (audio incl.)	$0.112	$0.14	—
3.0, audio off	$0.084	$0.112	$0.42
3.0, native audio	$0.126	$0.168	$0.42
3.0 Omni, no video ref	$0.084–0.112	$0.112–0.14	$0.42
3.0 Omni, video ref, audio off	$0.126	$0.168	$0.42
Motion Control	$0.126 (STD)	$0.168 (PRO)	—

That band explains the reseller confusion: fal.ai passes the official USD equivalents through unchanged ($0.084/s for 720p image-to-video with audio off), while other wrappers mark up or bundle. Quotes outside the band are paying for the wrapper.

Two billing traps. The consumer app charges in Credits, a different currency with different numbers (12 Credits/s for 1080p with audio in-app, versus 1.2 Units/s on the API) — blog posts mix the two tables constantly. And whether failed generations are billed is not stated anywhere in the public docs; the docs do confirm a failed task occupies a concurrency slot until it ends. Budget as if failures cost you throughput at minimum.

The rules that actually 400 your requests

Aspect ratio only exists where nothing anchors it. aspect_ratio (16:9 / 9:16 / 1:1) applies to text-to-video, and to Omni only when there is no frame and no reference video — where it becomes required. Image-to-video has no ratio setting at all: output follows the input image. To ship 9:16 ads from product stills, crop to 9:16 before the call. The same rule holds on every engine we run; Kling is just stricter about admitting it.

Multi-shot grammar is exact. multi_shot expects shot n, m, words; segments — half-width semicolons, one to six shots, each at least 1 second, durations summing exactly to the total, 512 characters per shot, 3,072 overall. And @element names must not be substrings of each other or of prompt words.

Omni's modes are mutually exclusive. A feature_video (style reference) forces multi-shot on and native audio off; a base_video (edit mode) forbids first/last frames, multi-shot, and native audio. Element counts have their own arithmetic: up to 7 image elements without a reference video, at most 4 with one. Combining the wrong pair is a 400, not a warning.

The limiter is concurrency, not QPS. The docs state there are no QPS limits; the real ceiling is simultaneous tasks per account, shared across all API keys, per model version and package type. Bursting past it returns HTTP 429 with code 1303 rather than a queue. Moderation rejections surface as 1300/1301 at submit, and content-risk failures can also land later in the task's message field.

Outputs are purged after 30 days, and URLs are hotlink-protected. Copy results to your own storage on success. (Seedance gives you 24 hours; Kling gives you 30 days; the pipeline pattern is identical.)

Date-check anything written in spring 2026. 4K excluded reference-video workflows on April 23, then the June 17 update extended Omni reference video to 15 seconds with 4K. Articles written between those dates contradict current behavior.

For product video specifically

The economics suggest a split the vendor does not spell out: iterate on Turbo (cheapest audio-included second in the family) and reserve kling-v3 at 4K for the hero cut that earned it. Kling 3.0's other e-commerce-relevant trick is what the official guide calls native-level text output — it preserves signs, labels and logos from input images and can generate legible new lettering, which is genuinely rare in this model class. Add-on APIs (lip sync at 0.5 Units per 5 seconds, avatars at 0.4–0.8 Units/s, TTS at 0.05 Units per call) cover the talking-head cases.

If you are choosing between engine families rather than tuning this one, the same missing-manual treatment exists for MiniMax H3, and our Sora 2 shutdown migration map puts Kling's spec sheet in context against Seedance and H3.

Or skip the operations entirely

Everything above — auth generations, ratio rules, exclusion matrices, 30-day purges, concurrency budgeting — is the overhead of operating a raw video API. It is exactly the layer Riffkit exists to absorb: hand it a short video that already worked, and it rebuilds that video's formula around your product on the engines we operate, with the per-second rate shown before you render on the pricing page. We keep manuals like this one because we read them so you don't have to.