study

AI video models are converging on quality. Their prices are not.

The top five models in our benchmark sit within a tenth of a point of each other. The same clip can cost several times more depending on where you make it.

By Mara Voss · 3 October 2026

Two years ago, picking an AI video model meant picking the one that did not fall apart. That is no longer the hard part. On the 40 prompts we run through every model, the top of the table has bunched up so tightly that the ranking among the leaders says less than the price tag next to them.

The top five fit inside a tenth of a point

A spread of 0.10 out of 5. With 40 clips per model, a re-run could plausibly reorder these five, which is why our leaderboard shows a 95% confidence interval next to every score. Widen the view to the top eight and the spread is still only 0.24 points. The industry reading of this is that base model quality is converging, and our data agrees with it.

Convergence does not mean the models are interchangeable. They still fail on different things: one handles hands well and stumbles on liquids, another nails on-screen text and cannot hold a camera orbit. But the overall gap that used to separate a usable model from a toy has closed at the top.

The price gap has not closed

Our price radar reads the published prices of six platforms and APIs twice a day and converts everything, credits included, to US dollars for one clip. For the same model and the same settings, the spread is wide:

The pattern is not "APIs cheap, web apps expensive". Veo 3.1 costs exactly $3.20 for that clip on every API we track: fal, Replicate and Runway's own API. The spread comes from web subscriptions, where a credit is worth a different amount on every plan. On Runway's web app, the same Veo clip costs $2.56 on the top plan and $7.68 on the entry one. The cheapest Seedance 2.5, Seedance 2.0 and Kling clips are all on Higgsfield's largest plan, and the most expensive Kling clip is a pay-per-use API. Prices move, so the live figures are on each model page: Seedance 2.5, Seedance 2.0, Kling 3.0 and Veo 3.1.

What this does to the market

When the models converge, a lead in raw quality gets short-lived, and the value moves to the layers around the model: control over the shot, reuse of approved assets, review and approval, distribution into the tools people already use, and cost per finished clip. The market is splitting into a handful of labs competing on the base model and a growing set of platforms that sell the workflow on top, often routing between several models rather than owning one.

You can see the second group in our radar data. Platforms like fal, Replicate, Runway and Higgsfield each carry several of the same frontier models, and they compete on price, packaging and convenience rather than on a model nobody else has.

How to choose when the scores are this close

• Start from your failure, not the overall score. Check the category boards for the thing your brief depends on, for example hands or lip-sync. A model ranked fifth overall can be first where it matters to you.
• Then shop the price. Once two or three models are good enough, the platform you buy them from can change the bill more than the model does.
• Watch the entry plan. On subscriptions, the cheapest plan usually has the most expensive credits, so a big plan only pays off if you use it. Our radar shows the same clip on every plan.

How we score, and what the scores do not cover, is in our methodology.

Frequently asked

Which AI video model is the best right now?

On our benchmark, Google Gemini Omni leads at 4.91 out of 5, followed by Kling 3.0 Turbo at 4.89. The top five are within 0.10 points, so for most briefs the choice comes down to price, control and the specific failure you can live with.

Why does the same model cost different amounts?

Platforms resell the same models with different markups and billing: per second through an API, or credits inside a subscription, where a credit is worth more or less depending on the plan. The same 10-second Seedance 2.0 clip costs $1.80 through Replicate's API and $8.64 on Runway's entry web plan.