Kling O3 4K Adds Video Edit and Video Reference

Kling O3 4K on WaveSpeedAI gained two video-to-video endpoints on September 1, 2026: kwaivgi/kling-video-o3-4k/video-edit and kwaivgi/kling-video-o3-4k/video-reference. Both take a video plus a prompt, both output 4K, and both are priced at $0.462 per second. The difference is what the input video is for. Video-edit changes an existing clip and keeps its length. Video-reference makes a new clip and treats the input as guidance.

This post is the reference for the whole Kling O3 family as it stands today: three tiers, five task types, and the price of each. All figures come from the live catalog formulas on September 5, 2026.

The Kling O3 family on WaveSpeedAI

Kling O3 (also marketed as Kling 3.0 Omni) is Kuaishou’s unified multimodal video model. On WaveSpeedAI it ships in three tiers, each a separate set of model_uuid values.

Task	std	pro	4K
text-to-video	kwaivgi/kling-video-o3-std/text-to-video	kwaivgi/kling-video-o3-pro/text-to-video	kwaivgi/kling-video-o3-4k/text-to-video
image-to-video	kwaivgi/kling-video-o3-std/image-to-video	kwaivgi/kling-video-o3-pro/image-to-video	kwaivgi/kling-video-o3-4k/image-to-video
reference-to-video	kwaivgi/kling-video-o3-std/reference-to-video	kwaivgi/kling-video-o3-pro/reference-to-video	kwaivgi/kling-video-o3-4k/reference-to-video
video-edit	kwaivgi/kling-video-o3-std/video-edit	kwaivgi/kling-video-o3-pro/video-edit	kwaivgi/kling-video-o3-4k/video-edit
video-reference	via video on reference-to-video	via video on reference-to-video	kwaivgi/kling-video-o3-4k/video-reference

The std and pro tiers arrived in February 2026. The three 4K generation endpoints followed in April. The two new 4K video-to-video endpoints close the gap so that every task available at std and pro is now available at 4K.

Every endpoint uses the same request pattern: POST https://api.wavespeed.ai/api/v3/<model_uuid> with a JSON body, then poll GET /api/v3/predictions/<id>/result until the status is completed.

Video-edit vs video-reference

The two new endpoints share the same required inputs (prompt and video) and the same optional guidance (images, up to 4, and element_list, up to 3 Kling element IDs from kwaivgi/kling-elements). They differ in what they produce.

Video-edit uses the input video as the base context. The output follows the input’s duration, and the model applies the change you describe: restyle the scene, swap a product, change lighting or weather, replace a background. Motion and camera behavior from the source are preserved unless you say otherwise. The model page’s own advice is to state what should stay the same as clearly as what should change.

Video-reference uses the input video as a reference for motion, appearance, or scene, and generates a new clip. You choose the output duration (3 to 15 seconds) and aspect_ratio (16:9, 9:16, or 1:1). Use it when you want the pacing, gesture, or camera move from a piece of footage applied to a different subject or setting.

Neither endpoint generates new audio. Both keep the source soundtrack by default (keep_original_sound: true); set it to false for a silent output.

Parameters

Parameter	video-edit	video-reference
prompt	required	required
video	required, billed 3-16s	required, guidance only
images	up to 4	up to 4
element_list	up to 3	up to 3
keep_original_sound	default true	default true
duration	follows input	3-15, default 5
aspect_ratio	follows input	16:9, 9:16, 1:1

A minimal video-edit call:

curl -X POST https://api.wavespeed.ai/api/v3/kwaivgi/kling-video-o3-4k/video-edit \
  -H "Authorization: Bearer $WAVESPEED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "video": "https://example.com/product-shot.mp4",
    "prompt": "Change the daytime street to dusk with warm shop lights. Keep the camera move and the product exactly as they are.",
    "keep_original_sound": true
  }'

How the lower tiers handle the same jobs

The std and pro tiers already had video-edit, with one constraint the 4K endpoint relaxes: the std and pro input video cannot exceed 10 seconds, while 4K bills up to 16 seconds of input.

There is no std or pro video-reference endpoint. Instead, the std and pro reference-to-video endpoints accept an optional video URL alongside up to 4 reference images (up to 7 without a video). That is the same idea at lower resolution, and it is priced with a 1.5x multiplier. The 4K reference-to-video endpoint has no video field, so at 4K the video-guided path is the dedicated endpoint.

Pricing by tier

All Kling O3 prices are per second of billed duration. Generation endpoints bill the requested duration (3 to 15 seconds). Video-edit bills the input length, rounded up to a whole second and clamped to the tier’s range. The 4K tier is the only one with no audio surcharge.

Generation endpoints (text, image, reference to video)

Tier	Per second	Per 5s clip	With sound	Reference with video
std	$0.084	$0.42	$0.112 per second (x4/3)	$0.126 per second (x1.5)
pro	$0.112	$0.56	$0.14 per second (x1.25)	$0.168 per second (x1.5)
4K	$0.42	$2.10	$0.42 per second (no surcharge)	not applicable

On pro reference-to-video, sound without a video is x1.2 ($0.1344 per second) rather than x1.25. A 15-second 4K clip with audio is $6.30; the same clip on pro is $2.10, and on std $1.68.

Video-to-video endpoints

Endpoint	Per second	Billed range	3s	5s	10s	Max
std video-edit	$0.126	3-10s input	$0.378	$0.63	$1.26	$1.26 (10s)
pro video-edit	$0.168	3-10s input	$0.504	$0.84	$1.68	$1.68 (10s)
4K video-edit	$0.462	3-16s input	$1.386	$2.31	$4.62	$7.392 (16s)
4K video-reference	$0.462	3-15s output	$1.386	$2.31	$4.62	$6.93 (15s)

images, element_list, and keep_original_sound do not change the price on any tier.

Which tier to pick

• std for iteration and volume. At $0.084 per second it is one fifth the 4K price, and the parameter surface is nearly identical, so prompts and reference sets transfer upward once they work.
• pro when std output is close but not clean enough, particularly for reference-to-video with several identity references. The step from std to pro is a third more per second; the step from pro to 4K is nearly four times.
• 4K when the delivery target is actually 4K: broadcast, large-format display, or footage that will be cropped or reframed after generation. It is also the only tier where audio does not add to the price, and the only tier with a 16-second edit window and a dedicated video-reference endpoint.

A practical workflow: prototype the edit prompt on kling-video-o3-std/video-edit at $0.63 per 5-second clip, then run the approved prompt once on the 4K endpoint at $2.31. That keeps most spend at the cheap tier while the final asset is 4K.

Related reading

• Kling O3 Pro Video Edit on WaveSpeedAI covers the pro edit endpoint in more depth.
• Kling O3 4K Reference-to-Video explains the image-reference path at 4K.
• Kling O3 4K Text-to-Video walks through multi-prompt and element lists.

Model pages: 4K video-edit and 4K video-reference. The full family and pricing are on the Kling O3 API page.