Kling 4.0 vs Luma Ray 3.2: One Makes the Sound, the Other Makes the Grade

Put Kling 4.0 and Luma Ray 3.2 in the same sentence and most comparisons go straight to resolution, because both advertise high-end output and neither is cheap. That is the least interesting axis. The two models disagree about something more fundamental: whether a generated clip is a finished piece or a plate. Kling 4.0, generally available since 15 October 2026, generates picture and two-channel stereo sound in one pass, holds a take for up to 30 seconds and expects you to publish what comes out. Luma Ray 3.2, Luma's flagship since 9 June 2026, generates silent video and puts its engineering into what happens after the take — native 16-bit HDR, EXR export in ACES2065-1, and a reframe mode that changes the framing without re-rendering the shot. 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.

That difference decides more purchases than a resolution comparison does, and it is worth stating before the numbers: if your deliverable is a clip you post, silence is a defect nothing else on the spec sheet fixes. If your deliverable is a shot a colourist will finish, an EXR in the right colour space is worth more than any amount of generated ambience. Everything below is about which of those two you are.

The scoreboard problem, stated plainly

Neither model can be compared on an independent score today, and for different reasons. Luma Ray 3.2 has no row on any of Artificial Analysis's three video leaderboards — text-to-video, image-to-video or editing — checked 22 October 2026. The only Luma entry on the image-to-video board is the older Ray 3, from September 2025, sitting 47th at Elo 1,215 on the silent scale. Ray 3.2 is not there at all.

Kling 4.0 was released eight days before this reading, and has not been entered on any board either. The K​ling rows that exist are Kling 3.0's, from February 2026, and they say nothing about the new generation. So this is a matchup with no shared measurement on either side — which means anyone quoting an Elo gap between these two models is quoting something they built themselves. Judge the rest of this page on specifications and billing, and treat any score you find elsewhere with suspicion.

Picture, sound, and what each leaves for later

K​uaishou's release note for Kling 4.0 is unusually specific about audio. Two-channel stereo, tighter lip sync, dialogue across nine languages plus regional accents — American, British and Indian English; Beijing, Taiwanese, Northeastern and Sichuanese Mandarin; Cantonese. It also generates on-screen text in those languages, which is a different capability from speech and is the one that fails most visibly when a camera moves.

• Sound — Kling 4.0: native two-channel stereo with lip sync, in the same pass. Ray 3.2: no audio from text or image prompts; audio survives only on Modify and Reframe, where it is the source clip's audio, preserved rather than generated.

• Duration — Kling 4.0: 3 to 30 seconds in a single pass. Ray 3.2: 5 or 10 seconds for fresh generation; the widely quoted 20 seconds belongs to Modify Video, which transforms existing footage.

• Resolution — Kling 4.0: 720p, 1080p or 4K, with 21:9 available. Ray 3.2: 360p, 540p, 720p or 1080p, and no 4K generation — Luma's 4K is an upscale utility applied afterwards.

• Colour pipeline — Kling 4.0: 10-bit HDR at 1080p and 4K, listed as coming soon. Ray 3.2: native 16-bit HDR generation and EXR export in ACES2065-1, available now, priced at twice the SDR rate for HDR and three times for HDR plus EXR.

• Frame-level control — Kling 4.0: up to 10 keyframe images placed across the timeline. Ray 3.2: up to 16 keyframes per clip, including on image-to-video.

• References — Kling 4.0: up to 15 combined assets across images, video, elements and voice. Ray 3.2: text, images, keyframes and character references, with no video-reference budget published.

Read down that list twice. On four of six lines Kling 4.0 wins on paper. On the one line that a finishing house cares about most — a 16-bit EXR in the colour space the compositor actually consumes — Ray 3.2 is not just ahead, it is the only one of the two that has shipped the feature at all.

Two ways to be billed, and why they do not convert

Luma sells clips, not seconds. Its published ladder for text-to-video and image-to-video runs 20 credits for a five-second draft-resolution clip, 50 at 540p, 100 at 720p and 400 at 1080p; a ten-second clip is 60, 150, 300 and 1,200 respectively. Modify Video and Reframe are priced separately and far higher, which is consistent with them being the finishing tools rather than the generator.

What a credit costs depends on the plan you bought it in, so there is no single dollar figure that is honest to quote for Ray 3.2 — a subscription buyer and a pay-as-you-go API buyer pay different rates for the same 100 credits. The property that does survive the conversion is the shape of the invoice: it is fixed per clip, knowable before you press generate, and it does not change whether the model renders two seconds or all ten.

K​uaishou has published no rate card at all for Kling 4.0. The API is in early access, the developer console says so in a banner, and there is no per-second price for either the full model or Kling 4.0 Flash on the public pages. That is not a small omission in a comparison about cost — it means the cheaper-looking option cannot actually be costed yet, and the honest position is that Kling 4.0's economics are unknown while Ray 3.2's are at least structured.

For reference, the previous generation is priced and routeable: kling/kling-v3 sits at a flat $0.084 per request on our catalogue, with kling/kling-3-turbo at $0.112 and the 2.x and O1 generations below that. Ray 3.2 is not a route we carry. So the two models a reader can actually call today are Kling 3.0 and Ray 3.2 — and the new K​ling is the one waiting on a rate card.

Where the 20-second number comes from

Luma's marketing around Ray 3.2 leans on length, and the length figure is easy to misread. Fresh generation is 5 or 10 seconds. Twenty seconds is what Modify Video can output when the source is 24fps: at 30fps the ceiling falls to 15 seconds, at 60fps to 7. That is not a limitation Luma hides — it is printed on the pricing page — but "up to 20 seconds" travels without the frame-rate clause, and a team that buys expecting a 20-second prompt-to-video clip will be disappointed by the first call.

The same discipline applies to HDR. Ray 3.2's HDR mode is limited to five-second clips at 720p or 1080p, and is not available on ten-second output. So the model's best colour pipeline and its longest fresh generation are mutually exclusive, which is a trade-off nobody puts on a comparison slide. Kling 4.0's 10-bit HDR, when it arrives, is specified for 1080p and 4K without a duration qualifier published — but it is listed as coming soon, so it is currently a promise on both sides of this particular line.

What the decision looks like inside a studio

Ray 3.2's real product is iteration on a shot you already have. Reframe changes aspect ratio, extends the frame or swaps a background after the take while holding the original lighting; Modify Video transforms source footage rather than inventing it; EXR leaves the model in a form Nuke and Resolve will open without a conversion argument. A team doing client work with a colourist in the loop gets value from those features on every job, and no amount of stereo audio substitutes.

Kling 4.0's real product is the first assembly. Thirty seconds, ten keyframes and fifteen references mean a sequence can be authored rather than stitched, and the audio arrives attached — which matters most for social, UGC and ad work where the deliverable is vertical, captioned and published the same day. The vendor's own examples are vlogs, short drama, product ads and music video, which is a fair description of where a finished-in-one-pass model earns its keep.

There is one route consideration that cuts across both. If a team ends up using each model for what it is good at — K​ling for the publishable take, Ray for the plate that goes to grade — the integration cost is the thing that decides whether that happens in practice. One OpenAI-compatible key in front of more than 200 models makes a second model a model-ID change rather than a second vendor contract and a second quota to monitor, and automatic failover keeps a batch moving when one serving path degrades. We route neither Kling 4.0 nor Luma Ray 3.2, and this piece is not claiming otherwise; the routing argument is about the models around them, not these two.

Which one, concretely

Take Luma Ray 3.2 when the shot is the problem and the sound is somebody else's job. A camera move you cannot get right, a framing that has to change after the client sees it, a deliverable that has to arrive as HDR or EXR — those are the problems Reframe, keyframes and the colour pipeline solve, and Kling 4.0 does not solve them at any resolution. The per-clip billing is a genuine advantage in that workflow, because direction is what you iterate on and a fixed clip price does not punish the tenth attempt more than the first.

Take Kling 4.0 when the output is the deliverable and the sound is part of it. Thirty seconds in one pass, ten keyframes, two-channel stereo and fifteen references is a combination nothing else in this comparison offers, and the 21:9 and 4K options make it viable for delivery formats Ray 3.2 will not generate natively. The caveats are real: no published price, no independent score, HDR still promised rather than shipped, and an API that is in early access rather than generally available to developers.

What would settle this is a Ray 3.2 entry on one of the video boards, and a Kling 4.0 rate card. Until both exist, the two sentences that hold are these: one of these models generates the soundtrack and the other exports the colour space, and the choice between them is made by your edit, not by your benchmark tab.