Kling 4.0 vs Runway Gen-4.5: Read the Footnote Before the Score

The most quotable fact in this matchup is hidden in a footnote. On Artificial Analysis's text-to-video board, read 22 October 2026, Runway Gen-4.5 sits 24th at Elo 913 — and the board attaches a note to that row explaining why the number is not what it looks like: Runway's API accepts prompts of up to 1,000 characters, so Gen-4.5 was generated on the 61% of the board's prompts that fit inside that limit, and its results reflect those prompts only. A model scored on three-fifths of the test, against models scored on all of it, is not a model that lost the test. It is a model that took a shorter version of it. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Kling 4.0, generally available since 15 October 2026, is not on that board at all — it is eight days old and unscored. So the two headline numbers in this comparison are a qualified score and no score, which is a poor basis for a purchase and a good reason to read the rest of the page: what each model is inside, what it costs, and what the footnote tells you about how Runway builds for a different customer than K​uaishou does.

What the 61% note actually says

Artificial Analysis generates every model's outputs from the same prompt set, then has evaluators vote blind on pairs. When a model cannot accept a prompt — too long, wrong modality, unsupported aspect — that pairing is not run. Runway's 1,000-character ceiling means 39% of the text-to-video prompt set never reached Gen-4.5, and the note is the board's own disclosure that its position reflects a subset.

Two conclusions follow, and only one of them is comfortable. The first is that 913 is a floor of unknown depth: the prompts Gen-4.5 could not accept are, by construction, the longer and more detailed ones, and those are usually where a model's instruction-following is tested hardest. The second is that the ceiling itself is a product decision, not an accident. A 1,000-character cap is roughly 150 to 200 words — enough for a shot description, not enough for the multi-paragraph creative direction Kling 4.0 accepts at up to 8,000 tokens. Runway has built an interface where you direct with images, keyframes and presets; K​uaishou has built one where you direct with prose.

That distinction is the whole comparison in miniature. If your prompts are short, Gen-4.5's board position is probably pessimistic. If your prompts are long, Gen-4.5 is not the model you were going to use, and the score was never relevant.

Two products that happen to share a category

Runway Gen-4.5 shipped on 1 December 2025 and held the top of the text-to-video board at 1,247 Elo at launch, displacing Veo 3 and Sora 2 Pro. Its API opened on 10 February 2026. What it is sold as is a studio: an editor, generation, iteration and assembly in one place, a credits plan, ProRes and PNG sequence export, an enterprise roster. The model is one component of that. Ten months after launch it has slid down a cumulative Elo table, which is what happens to any model that stops collecting votes while the field around it keeps moving — not evidence that it got worse.

Kling 4.0 is sold as a model, with a creation page wrapped around it. Thirty seconds per pass, ten keyframes, fifteen references, two-channel audio, up to 4K, an MCP and CLI surface for agents, and a redesigned canvas that the release note spends several paragraphs on. There is no editor in the Runway sense and no subscription ladder published alongside it; the commercial surface is credits on K​uaishou's own platform plus an early-access API.

• Duration — Gen-4.5: 2 to 10 seconds at 720p, with ceilings reported elsewhere ranging from 10 to 60 seconds, a spread wide enough to mean nobody has pinned it. Kling 4.0: 3 to 30 seconds in one pass.

• Resolution — Gen-4.5: 720p generation, with 4K as a plan-level upscale rather than a native mode. Kling 4.0: 720p, 1080p, 4K native, plus 21:9.

• Audio — Gen-4.5: contested. Launch coverage described native audio as a headline feature; at least one model-reference page states the model has no native audio and that sound in the demos was added in post. Kling 4.0: two-channel stereo in the same pass, with lip sync and nine languages, which the release note specifies in detail.

• Control surface — Gen-4.5: image-to-video, keyframes and video-to-video inside an editor. Kling 4.0: 10 keyframes on a timeline, 15 references, and in-model editing across up to 5 input clips.

• Delivery formats — Gen-4.5: ProRes and PNG sequence export, documented and priced. Kling 4.0: 10-bit HDR at 1080p and 4K listed as coming soon; no ProRes or sequence export published.

• Where it runs — Gen-4.5: Runway's platform and API. Kling 4.0: K​uaishou's creation page, with an API in early access and no published model identifier.

The money, in the units each vendor chose

Runway bills 12 credits per generated second, with a credit worth about a cent on the developer tier — roughly $0.12 per second — under a subscription floor of about $12 to $76 a month. Artificial Analysis records $7.20 per minute for the 720p configuration, which is the same arithmetic in a different unit. Billed in seconds of output at a fixed resolution, the invoice is knowable before you generate, and the rate does not move with the output settings because the settings barely vary: 720p, two to ten seconds.

K​uaishou publishes nothing for Kling 4.0. The API console shows an early-access banner and no per-second rate for either tier. What is published is the previous generation: kling/kling-v3 and kling/kling-v3-omni at a flat $0.084 per request, kling/kling-3-turbo at $0.112, on our catalogue, with provider list price passed through and no markup added — which is the mechanism that matters here, because it means a K​uaishou launch discount or a later rate cut appears on our line the same day rather than at renewal. When Kling 4.0 gets a rate, that property applies to it too. Today it has no rate to pass through.

The arithmetic worth doing before either invoice is cost per accepted clip, not cost per second. Gen-4.5's ten-second ceiling means a 30-second deliverable is three generations stitched, with the joins and the re-rolls that implies; Kling 4.0's 30-second pass means one, at a price nobody has published. A cheaper second is not cheaper if it takes four of them.

Where each one can actually be called

We do not route Runway Gen-4.5 and we do not route Kling 4.0. Runway is reachable through its own platform and API; Kling 4.0 through K​uaishou's creation page and an early-access developer API. Neither claim in this paragraph is an availability claim on our side, and this piece should not be read as one.

What is routeable today is the previous K​ling generation and the rest of the video catalogue, on the same OpenAI-compatible endpoint as the text models. For a team already holding a Runway subscription because of the editor, the practical arrangement is to leave the editing workflow where it is and put the API generation behind a router — the two decisions are separable, and most pipelines treat them as one. Automatic failover is the part that changes day-to-day operations: a serving path degrading upstream becomes a route change rather than a failed batch, and the client code does not move.

The other thing a router buys in a comparison like this is a fair test. Running the same prompts against Kling 4.0 on K​uaishou's platform and Gen-4.5 on Runway's means two accounts, two billing systems and two sets of terms; running them against models that share a key means the difference you measure is the model, not the integration.

Who the footnote is really addressed to

There is a reading of the 61% note that is unkind to Runway and a reading that is unkind to the board, and the truthful one is unkind to neither. Runway optimised its API for a workflow where the prompt is a caption and the direction lives in images, keyframes and an editor. That is a defensible product decision for the customer it sells to: agencies and studios whose briefs arrive as boards, not as paragraphs. The board optimised its prompt set for testing text-to-video as a language task. The mismatch is structural, and 913 is what it looks like when a scoreboard grades a model on a task adjacent to the one it was built for.

Kling 4.0 is built for the board's task and then some. Eight thousand tokens of prompt input, a director-style prompt expansion the release note describes, ten keyframes, an editing mode that takes five input clips. It is, on paper, the most text-directed video model in this comparison. It is also entirely unscored, so the paper is all there is.

That is why the useful question is not which model wins. It is which workflow you are in. If the deliverable leaves the building as a ProRes file that a finishing house will conform, Gen-4.5's export path and its editor are the parts you will touch every day, and its board rank will never enter the conversation. If the deliverable is a 25-second sequence authored from a written brief, with dialogue and a product that has to stay identical across shots, Kling 4.0 is the only one of these two built for that, and the absence of a score is a reason to test it yourself rather than a reason to skip it.

The test that settles it

Run one brief through both, this week, and score the outputs against the brief rather than against each other. Take something with a beat structure — an establishing shot, a character entrance, a product reveal, a closing line — because that is where the two designs diverge most sharply: Gen-4.5 will ask you to build it in cuts, keyframes and an edit, and Kling 4.0 will ask you to write it as a shot plan with keyframes on a timeline. Count the generations it took, the re-rolls, and the minutes you spent directing rather than generating.

Then look at what came out. If the Runway version needed six clips and the K​ling version needed one, the price difference between them is not the number on either rate card. If the Runway version was faster to finish because the editor did the assembly, that is a real advantage and no Elo table records it. And if Kling 4.0's 10-bit HDR is still marked coming soon when you run the test — it is, as of 22 October — then the colour comparison is not yet available and the honest version of that row is "unshipped on one side".

The two sentences that survive the footnotes are these: Runway Gen-4.5 is a scored model measured on 61% of a test it was not designed for, and Kling 4.0 is an unscored model designed for exactly that test. Neither is a verdict, and together they are enough to tell you which one to put in front of your next brief.