Wan 3.0 Tops the Rebuilt Video Leaderboard, and Kling 3.0 Becomes Its Anchor

Artificial Analysis rebuilt its text-to-video board on September 30. The new scale, AA-Video-T2V v2.0, judges every model at 1080p and draws on more than 68,000 human preference votes across roughly 1,000 prompts, grouped into ten use categories and ten capability axes. Wan 3.0 sits first with audio at an Elo of 1,156, and it takes the top spot in ten of the twenty category rankings.

The number that tells you more about the state of the market is the one at the bottom. Kling 3.0 at 1080p is the fixed anchor at 1,000. Nobody is ranked below it because the scale is pinned there. Kuaishou's previous flagship used to be the domestic standard for AI video; on this board it is the reference point that everything else is measured against.

A pack, not a podium

The top four models sit within 18 points of each other, on error bars of roughly nine to ten. Wan 3.0 leads at 1,156. Utopai X, which is post-trained on MiniMax H3 and has no public API, is second at 1,148. Dreamina Seedance 2.5 follows at 1,142, and MiniMax H3 at 768p is fourth at 1,138. fal's H3 Max rounds out the group at 1,134.

Read those numbers as a pack rather than a ladder. A model scoring inside an error bar of the leader is not measurably worse than the leader. Three of the top five are H3 or built on it, which is the more interesting fact: the open-weight base MiniMax released has become the substrate other teams post-train on, the same way people built on Stable Diffusion three years ago.

There is a distribution story underneath the ranking that the board makes visible for the first time. A year ago, the top of a video leaderboard was a list of closed models from a handful of well-funded labs. Today the top five include an open-weight base and two models post-trained on it by smaller teams. Utopai X has no public API, so the model sitting second on a public leaderboard stays out of reach for anyone outside the company. That is a new kind of ranking artifact, and it says something about how quickly a good post-training run can lift an open base into contention.

On the companion no-audio board, launched the same day, Wan 3.0 leads at 1,129 with H3 and H3 Max taking second and third. Removing audio narrows the field, which suggests the Chinese models' advantage runs through their audio handling rather than their visuals alone. It also flips the interpretation of the leaderboard: a studio that only needs silent footage is looking at a different competitive picture than one that needs lip-synced dialogue.

What the price column says

The leaderboard also carries a per-minute price, and that column reads differently from the Elo column. Wan 3.0 costs $12 per minute of video. Seedance 2.5 is the most expensive model in the top ten at $34.12 per minute. MiniMax H3, which anyone can download, runs at $4.80 per minute.

So the open-weight model in third place costs a quarter of what the leader charges and a seventh of what the second-place commercial model charges. For a team shipping a product, that gap matters more than nineteen Elo points. The question shifts from "which model is best" to "which model is good enough that the per-minute bill does not kill the business case." H3 is the answer for a lot of teams that would have paid for Seedance a year ago.

Seedance 2.5's $34.12 per minute figure is worth a second look, because it is the top ten's most expensive entry and it ranks third. The pricing implies a customer who values quality enough to pay seven times the open-weight rate, and those customers exist in advertising and film. What the board cannot show is whether that premium survives contact with a budget. If a production needs two hundred shots, the difference between $34 and $4.80 per minute is the difference between a project and a spreadsheet problem.

Wan 3.0's lead comes with a caveat: it is in invite-gated commercial beta. H3 is the one you can actually buy today at a published price. A leaderboard position for a model you cannot access is a signal about where the field is heading, not a purchasing decision.

The video-editing board adds a second dimension. Wan 3.0 holds first place there at an Elo of 1,188, taking the category from MiniMax H3, which had held it since its August debut. Editing is where the practical workflows live, because most commercial video work starts from footage that already exists. A model that leads on generation but trails on editing is less useful to a production team than a model that is strong at both. Wan 3.0 currently leads on both, which is the strongest position any model has held on this board.

Kling 4.0 shipped as a spec sheet

The other thing this board clarifies is how much of the recent Kling 4.0 launch exists outside a press release. Kuaishou's account announced the model on September 28: up to 4K, 10-bit HDR, 15 multimodal references, 10 keyframes, native 30-second generation, stereo audio. The release date was "coming this October."

What actually shipped is Kling 4.0 Flash, live for Ultra Yearly subscribers. The feature page carries no price, no date, no resolution text, and no mention of October. The newest Kling entry on any Artificial Analysis board is still 3.0. A creator post that disclosed Kling paid for it says launch output is 1080p and cheaper than Seedance, which conflicts with the "up to 4K" headline.

None of that makes the announced capabilities fake. It means the gap between an announcement and a usable model is where most of the work lives, and this launch has not closed it yet. A spec sheet that lists 4K, HDR, fifteen references and ten keyframes describes a capability set that, if it ships, would be the strongest in the open market. It also describes a set of features that every competitor will need to answer within a quarter.

The pattern is familiar from the last two years of model launches. Announcements set expectations, and the interval before general availability is where the actual product gets built. For a buyer, the practical advice is the same as always: do not plan a production pipeline around a feature page. Plan it around what you can call today.

What to watch

The next signal is whether Wan 3.0 leaves invite-gated beta and publishes a price. If it does, the gap between the leader and the open-weight pack becomes a real budget decision rather than a theoretical one.

The second signal is Kling 4.0 itself. Its last flagship set the anchor on this scale. The next one is being priced into expectations before anyone outside Kuaishou has run it. When it does arrive, it arrives against a board where three of the top five slots are already occupied by models that cost less.

The third is what happens to the anchor itself. Kling 3.0 sits at 1,000 by definition. If Kling 4.0 ships and lands above it, the scale's fixed point moves and every Elo on the board becomes a comparison against a new baseline. That is the moment this ranking stops describing September and starts describing where the year is going.