What happened
Cloudflare has published a model page for klingai/kling-v4, identifying it as Kling AI’s Kling 4.0 model and classifying it as a third-party offering in the Cloudflare AI catalog. The page describes two supported generation modes: text-to-video and multi-keyframe image-to-video. Developers can submit a prompt alone, or combine a prompt with images that establish points in the intended sequence.
The documented output ceiling is video of up to 30 seconds at either 720p or 1080p. Prompts may be as long as 26,000 English characters, and a request can use up to 10 keyframes. For keyframe-driven generation, the parameter scheme distinguishes a first frame, optional middle frames, and a last frame. The prompt can refer to those images as @image_1, @image_2, and so on, in the order first frame, middle frames, then last frame. This makes the supplied images part of the request’s directional specification rather than merely an optional visual reference.
Cloudflare’s examples show the distinction. One text-only request specifies aspect ratio, duration, prompt, and resolution. A second example uses one first-frame image to turn a pencil sneaker sketch into a product-style video. Another uses a first frame, a middle frame, and a last frame, with the prompt describing a continuous shot across those positions. The examples also show an ultra-wide 21:9 request, described as an aspect ratio new in Kling 4.0.
The model can be called from a Worker through env.AI.run, using the model identifier klingai/kling-v4, or through Cloudflare’s REST endpoint for AI runs. The documented responses show a completed state and a video URL in the result. That is a direct application integration surface: a program can send structured inputs and receive a location for the generated video, rather than requiring a person to operate a separate creation interface.
The page lists per-second pricing. Its pricing table gives $0.157 per second at 720p and $0.235 per second at 1080p; it also lists a $0.235 default per-second price. Duration and resolution are therefore not only creative choices but explicit cost inputs. A five-second 1080p request, for example, is priced from the stated per-second rate rather than as a flat subscription action. The source does not provide comparative quality, latency, throughput, or availability figures.
What the archive adds
This is a new event on a known Cloudflare platform thread. A recent Field Note documented Clef in Cloudflare Workers AI, another model entry with its own context and token pricing. Kling 4.0 extends that catalog in a different modality: video generation measured by seconds and resolution rather than input and output tokens.
Trend signal
This supports Generative media becomes application infrastructure. The relevant evidence is operational, not aesthetic: Kling 4.0 is exposed as a model identifier with structured request parameters, programmatic invocation, result URLs, and metered pricing. Those are the properties that let an application compose video generation into a product flow instead of treating it only as a standalone creative session.
Connections to other events
An earlier archive item described Kling O1 as combining generation and editing across text, images, motion, and video, with short clips and character consistency as part of its positioning. Kling 4.0’s Cloudflare documentation does not establish the same feature set or quality claims. What is new here is the Cloudflare-side delivery mechanism: text and multi-keyframe video generation are documented as API calls within the platform’s model catalog.
Why it matters
For an AI product team, the integration boundary matters as much as the model name. A video feature needs a reliable way to create requests, attach visual constraints, receive output, meter cost, and fit the work into the rest of the application. Cloudflare’s documented interface supplies those pieces for this model: prompt length, keyframe count, duration, resolution, invocation methods, and result shape are all specified.
That can simplify experimentation for teams already running workloads on Cloudflare. A product can test a generated teaser, product reveal, animated illustration, or other short asset without first designing a separate client integration around a different interface. The keyframe support is particularly relevant where a team needs to bound a shot’s beginning, intermediate state, or ending rather than accept an entirely unconstrained text-only result. But the interface specification is not evidence that a production feature will meet a team’s quality, latency, or rights requirements.
Limits and uncertainty
The documentation establishes the request interface and listed pricing, but it does not establish output quality, generation time, capacity, regional availability, rate limits, or reliability under production load. It does not state how well the model follows long prompts, preserves subjects across a full 30-second clip, or handles difficult transitions between many keyframes. The examples are demonstrations, not benchmarks. The source also does not describe licensing implications for a particular commercial use case beyond linking to terms and license information. Teams should inspect those terms and run their own tests before making product commitments.
Skill to build now
Build a small evaluation harness around one bounded video task this week. Create a fixed set of prompts and, for image-to-video, a fixed first-frame or three-keyframe sequence. Run variants across 720p and 1080p and across a few durations; log the returned video URL, stated duration, resolution, request inputs, and estimated cost from the published per-second rates. Review whether the keyframes actually preserve the intended scene transitions. This turns a catalog addition into evidence about a specific product workflow.
Reader takeaway
Kling 4.0 is now documented as a Cloudflare-callable video model: text prompts or up to 10 ordered keyframes can produce clips up to 30 seconds at 720p or 1080p, with per-second pricing. The immediate opportunity is to test it as an application component, while treating quality, latency, and production suitability as open evaluation questions.
Open question
Will multi-keyframe control remain dependable enough across longer clips to make video generation a repeatable product workflow rather than a best-effort creative feature?