What happened
Replicate’s page for kwaivgi/kling-v4.0 describes Kling 4.0 as Kuaishou’s latest video model. It says the model can generate clips from 3 to 30 seconds at either 720p or 1080p, and can also edit existing video. Rather than asking a user to choose among separate text-to-video, image-to-video and Omni models, the page says Kling 4.0 selects a mode based on the inputs supplied. The page shows 9.7K runs, but does not say when those runs occurred or provide a comparison baseline.
For text-to-video, users specify a prompt, aspect ratio, resolution and duration. Supported aspect ratios are 16:9, 9:16, 1:1 and 21:9. For keyframe animation, a request can include a start image, an end image and any number of middle images; the output takes its aspect ratio from the images. A separate reference_images input accepts up to 10 images, which prompts can refer to as @image_1 and so on. The page suggests using these references for characters, products, places or styles, but does not quantify how consistently the model preserves them.
Video references can guide camera movement, motion or style. Editing uses a base_video and a prompt describing a change; the page says the output keeps the original video’s length and aspect ratio. There can be up to five videos in total across the base and reference inputs. Each input must be 3–30 seconds, and combined input-video length cannot exceed 30 seconds. Supported video files are .mp4 and .mov, up to 300 MB, with dimensions from 700 to 4553 pixels per side and an aspect ratio between 2:5 and 5:2. Images can be up to 30 MB and must also fit that aspect-ratio range. Prompts can contain up to 26,000 English characters or 11,000 non-English characters.
The listed prices are $0.157 per second at 720p and $0.235 at 1080p for text, keyframe or reference-image inputs. With any video input, the listed rates are $0.094 per second at 720p and $0.141 at 1080p. Billing uses the output duration plus whichever is longer: input-video duration or output duration. So editing a five-second input into a five-second output bills 10 seconds. The page links to privacy, paid-service and service-level-agreement documents, but the supplied material does not detail their terms. It does not state an absolute publication or launch date.
What the archive adds
Our earlier Kling 4.0 note covered availability through Cloudflare AI Gateway and Workers AI, including published pricing and API examples. This is a new access-and-pricing detail on a known story: Replicate’s page documents its own model endpoint, inputs and billing examples. It does not establish that Replicate access began on a particular date.
Trend signal
This is not a strong fit for an agent-system pattern: the page describes a generative-video model endpoint, not agent interoperability or an agent runtime. It does, however, provide a concrete instance of generative media being offered as a composable service rather than only as a standalone artifact generator. That is consistent with Generative media becomes application infrastructure, though this page alone does not show integrations or production adoption.
Why now
The model page shows that the listing was created one week and six days ago and updated one week and three days ago. It provides no absolute date or explanation for the timing, so the available evidence does not support a claim about a launch announcement or why this access route appeared then.
Why it matters
For a team assessing video generation, the page makes several practical decisions more explicit: whether the task needs text, images, keyframes or source footage; what input constraints to validate; and how resolution and clip length affect the bill. The video-input rates are lower per second than the non-video rates, but the billing rule means an edit is not priced simply by output length. A five-second edit with a five-second result is billed as 10 seconds, so a prototype budget should include both sides of that calculation.
The page also consolidates generation and editing capabilities under one model identifier, which may reduce the need to route between separate model endpoints for those tasks. That is a description of the interface, not evidence that quality, consistency or workflow reliability is better than alternatives. Teams still need to test representative footage and prompts, check whether the result meets their quality bar, and review the linked service terms before sending sensitive material. The published numbers are useful for an initial estimate, not a full production-cost or vendor comparison.
Limits and uncertainty
The supplied page does not give a dated launch announcement, guarantee availability in particular regions, or describe a service-level commitment in the excerpt. Its 9.7K runs figure has no timeframe or context. It provides no independent quality evaluation, comparison with another model, or quantified evidence for consistency across clips. The listed input limits and prices describe this Replicate listing; they should not be assumed to apply to other providers or access routes. The source also does not establish how the model handles rights, consent or other governance questions for user-supplied media; linked privacy and service documents are not reproduced here.
Skill to build now
Before testing the endpoint, create a small cost-and-validation worksheet for one representative task. Record the input type, input-video duration if applicable, intended output duration and resolution. Calculate billed seconds as output duration plus the longer of input and output video duration, then apply the listed rate. Use the same prompt and reference assets across a few runs, and have a reviewer score whether the result preserves the required subject, motion and edits. This will turn the page’s price and limits into a local estimate, without treating a model listing as proof of production suitability.
Reader takeaway
Replicate’s Kling 4.0 page gives builders a concrete way to estimate video-generation and editing costs, along with input constraints to check before testing. The useful new detail is the access route and its billing rules, not proof of a newly launched model: existing coverage already documents Kling 4.0 elsewhere, and this page supplies no absolute launch date.