Kling 4.0 Flash Leaves Early Access: a $0.094 Second and a 720p Ceiling

Kling 4.0 Flash stopped being a subscriber perk on 4 November 2026. Kuaishou's Kling AI opened the model to every user on the Kling AI website and through its public API, and in the same move published the two things the five-week early-access phase never had: a request path you can call, and a per-second price. That price is the headline. At 720p with no input video, a second of Kling 4.0 Flash output costs $0.094 — about 40% below the $0.157 that the full Kling 4.0 charges for the same second at the same resolution. Everything else in the announcement is smaller than that number, and one line in the rate card is actively misleading.

The short version for anyone who just needs to know whether they can use it: yes. There is no waitlist and no annual plan requirement any more. The model generates 3 to 20 seconds at 720p from a text prompt, from a single first-frame image, or through Omni reference, and it is reachable at /text-to-video/kling-4.0-flash, /image-to-video/kling-4.0-flash and /omni-video/kling-4.0-flash. The longer version involves a billing rule that turns the cheapest-looking line on the rate card into the most expensive way to use the model, and a resolution ceiling that decides what the model is actually for.

What 4 November actually added

Kling 4.0 Flash was not new on 4 November. It has existed since 28 September, when Kuaishou announced the 4.0 generation and handed Flash to Ultra Yearly subscribers — the "Black Gold" annual tier — inside the Kling app. At that point there was no API, no endpoint, no model identifier, and no price. Coverage written in that window, including ours, had to describe Flash as a speed-and-value tier with "keyframe and reference limits not stated," because they genuinely were not stated anywhere a reader could check.

The change is that the documentation caught up with the model. Four things became concrete at once:

• Access — the waitlist ended. Any user on any plan can now run the model on the Kling website.

• Distribution — the API is public, and so are three non-API surfaces: MCP, a command-line client, and a set of AI plugins. There is also a Kling plugin for Adobe Photoshop, Premiere Pro and After Effects.

• Price — a published rate card with two billing rows, both at 720p only.

• Limits — 3 to 20 seconds, 720p, 8-bit SDR, a 26,000-character English prompt budget, up to 10 reference images and 5 reference videos, and up to 15 inputs on an Omni task.

The Adobe integrations are the part that changes where the model sits rather than what it can do. A model that lives in a browser tab is something you leave your edit for. A model that lives in a Premiere panel is something you use in the middle of a cut. That is a workflow claim, not a quality claim, and Kling has not published the plugin's round-trip behaviour — whether it returns a clip into the timeline as a linked media item or as a flattened render. Treat the integration as a convenience until you have used it on a real project.

The rate card, read line by line

Kling prices in "units" on its own page and converts at $0.157 per unit. For Kling 4.0 Flash there are exactly two rows, and no third row for a higher resolution because no higher resolution exists:

• No input video, 720p — 0.6 units, or $0.094 per second, billed on the output duration.

• With input video, 720p — 0.4 units, or $0.063 per second, billed on twice the output duration when the input is no longer than the output.

• 1080p — no row. The model does not generate it.

• 4K — no row, and none announced for Flash.

• Native audio — no row, and no audio parameter in the changelog entry.

For comparison, the full Kling 4.0 that went fully available on the API on 28 October carries $0.157 per second at 720p and $0.235 at 1080p, with 4K marked as forthcoming in both the model documentation and the promotional material. Kling 3.0, the older model, is still the only tier in the family with a 4K price at all — $0.42 per second — and the only video tier that sells a with-audio and a without-audio variant as separate rows.

Why the "with input video" line is not a discount

The $0.063 line looks like a 33% cut against the $0.094 headline. It is not. Kling's billing rule for every 4.0 tier is that when the input video is the same length or shorter than the output, you pay the unit price multiplied by the output duration times two. So a 10-second output from a 10-second input is charged as 20 seconds:

• 10-second output, no input — $0.094 × 10 = $0.94.

• 10-second output from a 10-second input — $0.063 × 10 × 2 = $1.26.

Supplying the input video costs $0.32 more, which is 34% above the simple text-to-video call, not a third below it. The arithmetic only gets worse in the second branch, where an input longer than the output is billed on the input and output durations added together: a 20-second input driving a 10-second output is $0.063 × 30 = $1.89, double the plain call.

There is a defensible reading of the rule — you are being charged for the conditioning work as well as the generation, and a reference video is more work than a prompt — but the practical instruction is simple. If you are moving from text-to-video to video-to-video on Flash, budget roughly 1.3× what the headline rate suggests, not 0.67×. The same trap exists on Kling 4.0 at both resolutions, where the with-input rate of $0.094 and $0.141 works out to $0.188 and $0.282 per output second — 20% above the plain rates rather than 40% below them.

The ceiling is the product decision

Two limits matter more than the price. The first is 720p. Every platform that will pay for this footage — a social feed, a vertical ad slot, a 720p preview in an edit — is fine with 720p. Every platform that wants a master file is not, and Flash cannot produce one at any price. The second is 20 seconds, against 30 on the full model, which rules out the single-take scenes that the 4.0 generation was largely sold on.

Then there is control. Flash's image-to-video path accepts a single first-frame image and nothing else. It has no last-frame pinning and no keyframe track, so you cannot specify where the shot ends. The full Kling 4.0 accepts up to 10 keyframe images and, on the Omni path, can reference voices directly. If your shot has a defined end state — a product landing in a specific position, a character arriving at a mark — Flash can only be retried until it happens, and retries are what the cheap rate is supposed to protect you from.

Where the integrations change the workflow

MCP and the CLI are the interesting distribution surfaces, because they put a video model inside an agent loop rather than behind a web form. A coding agent that can call /text-to-video/kling-4.0-flash and read back a task id is a different product from a user pasting prompts into a box, and Kling shipping both on the same day as the API suggests the model is meant to be a component. The Adobe plugin is the consumer-facing version of the same idea.

What Kling has not published is a rate limit, a concurrency ceiling, or a queue policy for the public API. Those numbers decide whether the cheap per-second price survives contact with a production pipeline, and they are usually the first thing a team asks after the price. Until they exist, the responsible assumption is that Flash is a burst tool: fine for a hundred drafts in an afternoon, unproven for a hundred an hour every day.

Calling it, and what OrcaRouter does with it

The three request paths are the whole public surface: /text-to-video/kling-4.0-flash for prompt-to-clip, /image-to-video/kling-4.0-flash for a single first frame, and /omni-video/kling-4.0-flash for the reference-and-edit path. Omni tasks accept up to 15 inputs and can mix images, video and character elements. Duration and aspect ratio can be set by parameter or by prompt, including 21:9 and an adaptive mode.

OrcaRouter does not carry Kling 4.0 Flash, and it does not carry Kling 4.0 either. What it carries from the Kling family is the older generation — kling/kling-v3, kling/kling-v3-omni and kling/kling-video-o1 — served through a single /v1/video/generations endpoint at $0.08 per request, with an OpenAI-compatible client so the call shape is the one you already have. If you are building a pipeline around Flash, the part OrcaRouter actually covers is the text side: writing the prompts, grading the outputs, moderating them, and orchestrating which model gets which job, across 200+ models on one key with 0% markup on provider list prices and automatic failover between providers. The video call itself goes to Kling.

What to watch

Three things would change this read. An independent benchmark, because there still is not one: neither 4.0 tier appears on a public video leaderboard with a score, and every quality claim about Flash remains Kling's own. A 1080p or 4K row, which would make the model a delivery tier rather than a draft tier. And published throughput limits, which decide whether $0.094 a second is a price or an advertisement.

Until then, the honest summary is that Kling 4.0 Flash is the cheapest way to generate a lot of Kling-shaped footage, and it is priced like a component rather than a product. The 40% discount against Kling 4.0 is real. The 33% discount on the input-video line is not. And the 720p ceiling is not a discount at all — it is the thing you are paying with.