Kling 4.0 vs MiniMax H3: The Scored Model Is the Older One

There is a version of this comparison that writes itself, and it is wrong. MiniMax H3 shipped in July 2026 and tops Artificial Analysis's image-to-video board. Kling 4.0 shipped on 15 October 2026, replaces a model that has been in production since February, and is the newest video model K​uaishou has. The obvious reading is that the challenger has arrived and the incumbent's crown is at risk. The actual state of the record is stranger: the older model is the one with independent scores, published prices, downloadable weights and a route on our platform, while the newer one has a longer spec sheet, no score, no rate card and an API still in early access. Which of those you should care about depends entirely on whether you are buying a capability or a vendor relationship. 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.

Both models are real and both are strong. MiniMax H3 is a 4-to-15-second model at 768p native, with 2K available through an API-only regeneration module, native 32kHz stereo generated in the same pass, and a unified context that reads text, images, video and audio references together. Kling 4.0 is a 3-to-30-second model at up to 4K, with up to ten keyframes, fifteen combined references and two-channel stereo. The gap in raw capability is real, and it is smaller than the gap in evidence.

The scores that exist, and the score that does not

Artificial Analysis runs three video boards with separate scales and separate vote pools, and M​iniMax's H3 appears on all three. Read 22 October 2026:

• Image-to-video (AA-Video-I2V v1.1, audio kept) — H3 (768p) is 1st at Elo 1,140 ±16 over 4,209 votes, released July 2026, recorded at $5.40 per minute. H3 Max sits 6th at 1,125 ±17.

• Text-to-video (AA-Video-T2V v2.1) — H3 (768p) is 5th at 1,158 ±16 over 5,781 votes, $4.80 per minute. Wan 3.0 leads at 1,173 and the H3 Max variant is 6th at 1,156, so the family occupies two of the top six slots.

• Video editing (AA-Video-Editing v1.1) — H3 (768p) is 4th at 1,125 ±11 over 7,457 votes, $9.60 per minute.

Kling 4.0 is on none of them. The K​ling rows the boards carry are Kling 3.0's from February — 19th on image-to-video at the 1,000 anchor, 17th on text-to-video, 9th on editing — and they describe the previous generation, not this one. Eight days after general availability is early for a board entry, so this is an absence of measurement rather than a bad measurement. It still means that for the next few weeks, "which is better" has a documented answer on one side and a press release on the other.

One caution about H3's numbers, because it cuts both ways. The image-to-video board's intervals are wide enough that first and sixth place overlap — H3 at 1,140 ±16, H3 Max at 1,125 ±17 — so "number one" is a rank, not a demonstrated gap. The honest statement is that H3 is in a leading cluster whose members have not been separated by the vote count yet.

What the extra fifteen seconds is actually for

H3's ceiling is 15 seconds, in integer steps from four. Kling 4.0's is 30, in one pass, with extensions to two minutes listed as coming soon. That is not a small difference in kind, because it crosses the line between a shot and a scene.

• Duration — MiniMax H3: 4 to 15 seconds. Kling 4.0: 3 to 30 seconds, extension to 2 minutes coming soon.

• Resolution — H3: 768p native, with 2K through an API-only regeneration module that self-hosters cannot run. Kling 4.0: 720p, 1080p, 4K, plus 21:9.

• Frame-level control — H3: first-frame, last-frame and first-and-last-frame modes. Kling 4.0: up to 10 keyframes anywhere in the timeline.

• Reference inputs — H3: up to 9 images, 3 video clips and 3 audio clips read as one context. Kling 4.0: up to 15 combined assets — 10 images, 5 videos totalling 30 seconds, 7 elements including at most 3 video-based — plus voice references.

• Audio — H3: 32kHz stereo generated with the video, at 24fps. Kling 4.0: two-channel stereo with lip sync and dialogue across nine languages.

• Prompt length — H3 publishes no token ceiling in its request documentation. Kling 4.0 accepts up to 8,000 tokens.

Where H3's design shows is consistency under a fixed camera: a 15-second take with one performance in it, driven by reference images and a first and last frame, is exactly the shape its controls assume. Where Kling 4.0's design shows is a longer sequence with beats — and the ten-keyframe timeline is the part of that upgrade that most changes how a shot gets authored.

The licence paragraph that decides it for some readers

This is the only matchup in this set where one of the two models can leave the vendor's servers, and the terms are worth reading before anything else if self-hosting is on your roadmap.

MiniMax H3's base checkpoint is downloadable under M​iniMax's community licence — not Apache, not MIT. That licence excludes the United States, the United Kingdom, the European Union and South Korea, where use requires an application to M​iniMax, and commercial use above a revenue threshold requires prior written authorisation. What is open is the base model: the Context-IR prompt-expansion module and the 2K regeneration path stay on the API side and were never part of the release. If you are inside one of those four territories, the weights are not a workaround for the API — they are a separate conversation with M​iniMax.

Kling 4.0 publishes no weights at all, and K​uaishou's launch material does not suggest any will follow. That is a different kind of closed: M​iniMax shipped something with terms you might be able to meet, and K​uaishou shipped nothing to meet. Neither position is unreasonable for a company selling a video model; they are just not the same position, and a team planning an on-premise pipeline should notice which side of the line each model sits on before it costs them a quarter.

What each costs, and the one number that is missing

M​iniMax bills per second of generated output: $0.08 per second at 768p and $0.13 per second at 2K, with reference audio free and the first five reference images free. That renders as $4.80 per minute at 768p, which is how Artificial Analysis records it on the text-to-video and editing boards, and as $5.40 to $7.80 per minute once the resolution configuration the board measured is taken into account. The draft-then-upscale trick does not beat the meter: 768p generation plus the 2K regeneration costs $0.08 plus $0.05, exactly the $0.13 direct 2K rate.

K​uaishou has published no rate for Kling 4.0 at all. The API developer console carries an early-access banner and no per-second pricing for either the full model or Kling 4.0 Flash. Every figure circulating for it today is someone's estimate. That asymmetry is the most decision-relevant fact in this comparison after the licence: one of these models has a meter you can read before you commit, and the other does not.

What we can price is the previous generation. kling/kling-v3 and kling/kling-v3-omni are listed at a flat $0.084 per request on our catalogue, kling/kling-3-turbo at $0.112, and MiniMax H3 is carried too, at the provider's own per-second rate passed through with nothing added. That matters for a side-by-side test: both a Kuaishou model and MiniMax H3 answer on the same OpenAI-compatible endpoint, so comparing them on your own prompts is a model-ID change rather than two vendor accounts, two SDKs and two quota dashboards. Failover sits behind the same call, which is what makes a week-long bake-off practical instead of a procurement exercise.

The availability question, where the open model is also the easier one

It is rare for the open-weight model to also be the more available one, and this is that case. MiniMax H3 is callable now, on our platform and on M​iniMax's own, at a published per-second rate, with three independent board entries behind it. Kling 4.0 is callable on K​uaishou's creation page and, for developers, through an early-access API with no published rate and no model identifier we can route.

Two things follow, and they point in opposite directions. If you need to ship something this month, H3 is the model with a price, a score and an integration story. If you need a 30-second take with ten directed beats and stereo dialogue, H3 cannot do it at any price, and waiting for Kling 4.0's rate card is the cost of that capability. The good news is that these are not exclusive: routing a short-clip pipeline through H3 while you trial Kling 4.0 on the vendor's platform is a normal way to run a transition, and it keeps the production path on a model whose behaviour is documented.

One caveat that routing cannot fix: if you are in the US, UK, EU or South Korea and your plan was to self-host H3's weights, calling it through an API is not a workaround for the licence. It is a different product with different terms, and the licence question stays yours to answer.

Picking one

Take MiniMax H3 when the clip is short, the budget is fixed and the evidence matters. It is a top-five model on the text-to-video board and first on image-to-video, across three boards and thousands of votes, at $0.08 per second for 768p. It generates its own stereo audio. It accepts nine images, three video clips and three audio clips in one context. And if you are outside the four excluded territories, it is the only one of these two you can download.

Take Kling 4.0 when the sequence is long and the number of moving parts is high. Thirty seconds, ten keyframes, fifteen references and two-channel dialogue is a capability H3 does not have, and no amount of benchmark position substitutes for it. Go in knowing that you are buying an unmeasured model at an unpublished price, that 10-bit HDR and the two-minute extension are both still listed as coming, and that the API is in early access rather than generally available.

What would change the shape of this comparison is a Kling 4.0 entry on one of the boards and a published rate card. Either one alone would move it from a capability-versus-evidence argument to an ordinary price-and-quality decision, which is where a buyer wants to be. Until then, the two sentences that survive scrutiny are: H3 is the better-documented model by a wide margin, and Kling 4.0 is the only one of the two that will hold a thirty-second take.