Kling 4.0 vs MiniMax H3 Max: A Post-Train Overtaken by Its Own Base

In late August, MiniMax H3 Max debuted at number one on Artificial Analysis's image-to-video board, twenty Elo ahead of the base MiniMax H3 it was post-trained from. On the board as I read it on 23 October 2026, H3 Max is sixth at Elo 1,125 and the base H3 is first at 1,140. The tuned variant did not get worse in any way anyone has demonstrated; the board changed around it, more votes arrived, and a newer prompt set reshuffled the order. That is what an arena does, and it is the first thing to understand before comparing either of them to Kling 4.0.

What "Max" is, and why the name now misleads

MiniMax H3 Max is not a new generation from MiniMax and it is not a bigger model. It is a post-trained derivative of MiniMax H3 — the Hailuo 3 generation — built by an external inference platform that tuned the base model for stronger prompt adherence and better aesthetics, then co-optimised it with its own serving stack for throughput. Those are the builder's claims. The only independent evidence for them was, and remains, its arena placement.

The board on 23 October, with the drift left in

Artificial Analysis splits its video rankings by whether the clip carries generated audio. The with-audio image-to-video board is the one this class of model competes on, and it is currently on version v1.1 rather than the v1.0 that most published comparisons quote.

• First — MiniMax H3 (768p), Elo 1,140 ±16 over 4,209 samples, released July 2026, listed at $4.80 per minute of 1080p output at default settings.

• Second — FLUX 3 Max (720p), 1,138 ±15 over 5,054 samples, released October 2026, $2.55 per minute.

• Third — Dreamina Seedance 2.5, 1,138 ±17, $34.12 per minute.

• Sixth — MiniMax H3 Max, 1,125 ±17 over 3,476 samples, released August 2026, $5.76 per minute.

• The overlap — from rank one to rank six the spread is 15 Elo against intervals of ±15 to ±17. This is a leading cluster, not a ranking you can defend at the individual-model level. Anyone quoting "first" or "sixth" without the date and the interval is overreading the board.

• Kling 4.0 — absent. The board added three models in the last 30 days; Kuaishou's newest is not one of them.

• The prior generation — Kling 3.0 1080p Pro anchors the board at 1,000; Kling 3.0 Omni 1080p Pro sits eighteenth at 1,004.

The drift is not a footnote, it is the finding. In two months the same model moved from 1,204 at debut to 1,195 in the October board to 1,125 today — and the board itself changed version between the first and last of those readings. If you are choosing a video model for a six-month project, a number that moves 79 points in eight weeks is a snapshot, not a specification.

What Kling 4.0 offers that no Elo captures

Kling 4.0 went live for everyone on 15 October 2026. Kuaishou publishes no benchmark of its own for it, and no independent score exists. What does exist is a specification that is clearly aimed at a different job from H3 Max's:

• Duration — 3 to 30 seconds in a single pass, with extension toward 2 minutes listed as coming soon. H3 Max does 5 to 15 seconds. For a 30-second shot, only one of these can generate it without a seam.

• Resolution — 720p, 1080p and 4K, with 10-bit HDR at the top two tiers marked coming soon on the vendor's own release note and feature page. H3 Max tops out at 768p. This is the largest difference between them, and it appears on no leaderboard.

• Control — up to 10 keyframe images and a 15-item reference budget across 10 images, 5 videos capped at 30 seconds combined, voice references and 7 elements. H3 Max's control surface is the base model's, plus first-and-last-frame handling.

• Audio — two-channel stereo with lip-sync across nine languages and named accents. H3 Max inherits H3's native audio, which is the reason the with-audio board exists as a separate ranking.

• Openness — MiniMax H3's weights are open under a community licence with territorial carve-outs and a revenue threshold for commercial use; H3 Max's weights have been promised but not shipped. Kling 4.0 is closed outright.

Read that list as a description of two different products rather than a scorecard. One is a short-clip, low-resolution, well-measured model that is cheap and callable. The other is a long-clip, high-resolution model with a control surface built for authored sequences, and no independent evidence yet that its picture quality matches its spec sheet.

Cost per minute, on today's numbers

• Kling 4.0 — $0.157 per second at 720p and $0.235 at 1080p on Kuaishou's preview sheet, which is $9.42 and $14.10 per minute. There is no 4K rate; the column reads "-".

• MiniMax H3 Max — $5.76 per minute of 1080p output at default settings, per Artificial Analysis.

• MiniMax H3, the base model — $4.80 per minute on the same basis, and $0.08 per second of generated video at 768p on OrcaRouter, which is $4.80 a minute for the model currently ranked first.

Take a hundred ten-second clips, a plausible monthly volume for a social or e-commerce team. That is 1,000 seconds, or 16.7 minutes of output.

• On H3 Max at $5.76 per minute — about $96.

• On Kling 4.0 at 720p, $0.157 per second — $157.

• On Kling 4.0 at 1080p, $0.235 per second — $235.

The gap is roughly 1.6 to 2.4 times, not the four-to-one that earlier rate snapshots implied. And the cheap option is the one with a score. That is the trade in one line: Kling 4.0 charges a premium for 30 seconds and 4K, and the premium is smaller than the spec gap suggests — but you are paying it on faith, because nobody outside Kuaishou has measured the output.

One caveat on the Kling figures. They come from a page that still marks the model "Coming Soon" and states that the prices "are for preview purposes only and cannot be used to submit requests before launch." Kuaishou announced on the evening of 21 October that the API had gone live, opening gradually to enterprise clients and developers. Both things are on the record; the practical reading is that the rate is an intention until a billing statement proves otherwise.

What you can call, and what that is worth

MiniMax H3 Max is served through the platform that built it, on that platform's own API. It is not an OrcaRouter route. The base MiniMax H3 is: $0.08 per second at 768p, billed per second of generated output plus metered reference inputs, on the same OpenAI-compatible endpoint as more than 200 models, at provider list price with 0% markup. That is also the model currently first on the board, which makes the routable option the better-ranked one — an unusual and temporary alignment.

Kling 4.0 is not routable either. There is no identifier for it in our catalogue, and saying otherwise would be a fiction. What is routable is Kling 3.0 at $0.084 per request, the generation it replaces.

What would make this a real contest

A Kling 4.0 entry on the with-audio image-to-video board. Kuaishou has the model in public hands and the arena has a slot; the absence is now eight days old and it is the single fact that keeps this comparison from being decidable.

A published 4K rate. The preview sheet leaves that column empty, and 4K is the one capability that justifies a premium over a 768p model. A model that costs more per minute needs to deliver something the cheaper one cannot — and at 768p it does not.

A shipped H3 Max weight release. If the post-train's weights land under anything like the base model's community licence, the 768p ceiling stops mattering for teams with their own GPUs, and the price argument changes shape.

The defensible position on 23 October is narrow. If you need 30 seconds, 4K, ten keyframes or a text-to-video path with a real reference budget, Kling 4.0 is the only one of the two that offers them, and you should budget from $9.42 a minute and expect it to move. If you need a scored, cheap, callable model today, MiniMax H3 Max is no longer the best answer — the base MiniMax H3 outranks it and undercuts it on the same board. That is the least glamorous conclusion available, and it is the one the evidence supports.