Choosing an AI Video Tool

Before you pick a tool

Answer one question in advance: what is the output, and who is going to watch it? A social clip, a product demo, a training video and a film test are four different jobs, and there is no tool that is best at all four.

The second question is whether you are working from the EU. In this field that counts for unusually much: at several providers there are functions and licences that explicitly exclude the Union. There is a separate section on it at the end of the post.

Who is who in the autumn of 2026 — two layers, do not mix them

Most of the confusion comes from calling two different things an "AI video tool". There is the model, built by a research lab, and there is the interface you actually open. Most interfaces today do not run their own model but other people's — and inside a single interface, which model you pick matters more than the interface itself.

The models

• Alibaba — Wan. It leads both the text-to-video and the editing ranking, and it gives you thirty seconds in a single request. Its top version is not downloadable; the public release is a generation behind.
• Google — Gemini Omni and Veo. Two separate lines, and they do not sit in the same place: Gemini Omni is second on the text-to-video ranking, Veo is seventeenth. Native audio, up to 4K. Clip length is the tightest here.
• MiniMax — H3. Leads the image-to-video ranking. The Hailuo name it used to be known by now belongs to the mobile app; the model is H3.
• ByteDance — Seedance. The other thirty-second one, and its reference allowance is the most generous in the field.
• Kuaishou — Kling. Shot-level script input, and it carries identity and voice across from a reference video.
• Lightricks — LTX. Open-weight, with native 4K and synchronised audio. This line matters because it is usable from the EU on licence grounds too.
• Vidu, PixVerse, Sand.ai, Skywork. Smaller names, but all of them sit in the top ten of the rankings — Vidu specifically for reference-based, character-consistent generation.
• SpaceXAI (formerly xAI) — Grok Imagine. Available on the large aggregator platforms; its documentation states honestly that what they call generation from text is in fact image generation followed by video generation.
• OpenAI — Sora. Discontinued. The consumer interface in April 2026, developer access on 24 September 2026. It can be mentioned in the past tense.

The interfaces

• Higgsfield. Not a mere aggregator: it builds its own model family too — camera control, character-consistent imagery and storyboarding — and runs a good part of the large models alongside. Its positioning pulls towards advertising and brand creative.
• Magnific (formerly Freepik). A full creative platform: image, video, audio, plus the old stock library. Its model list is the widest, and its pricing is the most transparent — it prints how many credits buy how many seconds, per model.
• Krea. Many models on one surface, with the most generous free daily allowance.
• Runway. Today more of an aggregator than a model maker — but its own editing model is on the editing ranking as well. Since 8 September 2026 it has a plug-in for the established editing and compositing program: generation happens on the timeline, with no export round trip.
• LTX Studio. A storyboard-centred workflow: it builds a sequence of scenes, not a clip.
• Luma. A brand-consistent, agent-like workflow with its own model family.
• Adobe Firefly. Rights-cleared positioning and the most transparent pricing of the lot: exactly ten cents a second throughout, whichever plan you are on.
• CapCut, Canva, VEED, Kapwing, Descript, InVideo, Pika. The editor and template side. Here video generation is one feature among many rather than the product itself — in exchange, the rest of the workflow is there too.

And a warning that follows on its own: the model lists of these interfaces go stale within months. In September 2026 several interfaces were still advertising a model in their catalogue whose developer access shut down within ten days. A model list is marketing material, not a guarantee.

What the rankings say — and where they disagree

There is one arena built on blind human votes, and it keeps three separate lists: text-to-video, image-to-video and editing. The three do not have the same model at the top — which is the most useful thing a ranking can teach you.

Two things stand out here, and neither is what you would expect.

First: the leading field is Chinese. In the top ten of the text-to-video list, Alibaba appears with three models, MiniMax with two, ByteDance and Sand.ai with one each — there is exactly one Western model among them, Google's Gemini Omni line. Veo is seventeenth. Anyone picturing Google's video line at the top is in fact thinking of Gemini Omni, not of Veo.

Second: the top placing is not statistically sharp. Between the top four models there are only seventeen points of difference, and the confidence intervals largely overlap. Anyone choosing "the model that leads the ranking" is paying for a difference inside the error bars.

And on the editing list it is worth knowing that only nine models are ranked at all. The provider most famous precisely for editing is only seventh even on that short list — which means the correct answer to "which is best" is at which job.

And an instructive case about the pace: in December 2025 one provider advertised its new model on the grounds that it was first in the rankings. Today that same model is not in the top ten, and the competitor's product it beat in the advertisement no longer exists at all.

Nine jobs, nine picks

• A social clip, fast. Any aggregator — Krea, Magnific, CapCut — with native audio. Eight seconds is plenty, and this is where the choice of model matters least.
• A longer scene with fewer cuts. The thirty-second model from Alibaba or ByteDance. This is the biggest practical difference in the field.
• A recurring character, a series. ByteDance for its reference allowance, Kuaishou for carrying identity across from a reference video, or Vidu, which was built for exactly this — the details are set out in character consistency, a post of its own.
• Editing and fixing existing material. The top of the editing ranking, or Runway's editing model — and watch the EU restriction on uploaded video.
• A film workflow, inside the editing software. Runway's plug-in, from September 2026. An important limit: it exists only for the two Adobe programs, not for the other established editing and compositing software.
• A full scene sequence from a storyboard. LTX Studio was built for this, and Higgsfield's storyboard line aims at the same thing. Here it is not clip quality that decides but whether the scenes hold together.
• An ad from a product photo or a product link. The answer here is not the generative scene but the purpose-built ad maker: Creatify builds video from a product link, HeyGen works with product placement. Different category, different price logic.
• A training video, internal education. A talking head avatar, not a generative scene — and the price difference is an order of magnitude. Avatar video for business covers this in detail.
• Recutting existing video, captions, long to short. Descript edits from the transcript, Kapwing captions and cuts. Breaking down long material is the subject of short clips, a separate post.

And two situations you solve with a decision rather than a tool. If there is no budget: Krea gives a daily allowance, Runway and LTX a one-off entry credit, Kapwing a watermarked free tier — these are fine for trying, not for producing a series. If the data cannot leave: the open-weight line is what remains, together with its licences, which has its own section below.

Prices: two completely separate logics

On the developer interface it is per second, and the spread is tenfold.

• The cheapest output is five cents a second.
• The middle category sits around ten to twenty cents.
• Top-end 4K output is sixty cents with audio — forty without it, and thirty on the fast variant.
• Avatar video on the same interface is ten cents a second.

So an eight-second clip can be anything from forty cents to five dollars, for the same eight seconds. And since roughly one clip in three turns out usable, multiply the real cost by three.

On the interfaces, though, it is credits — and there it is not the interface that decides but the model. Within one subscription, out of the same credit allowance, the effective per-second price ranges from three cents to fifty-seven — purely depending on which model you click. That is a wider spread than the price difference between interfaces. So the question is not which interface you subscribe to, but which model you use inside it.

And a lesson that runs against intuition: on the top subscription tier the interface can be cheaper than raw developer access, for the same model and the same output. Anyone thinking purely in API terms overestimates the cost.

And the talking-head avatar is a price category of its own. There the per-second rate on a subscription is roughly three cents — an order of magnitude below the generative scene. If the job is a training video, the price of the generative models is not even worth thinking about: you are shopping for a different product.

The legal small print

This is the part people skip, and it is the most expensive thing to skip.

• The terms can contradict each other. At one provider the general terms make commercial use conditional on written permission, while the billing terms say it is not restricted for subscribers. The sensible reading: not on a free account, yes on a paid one — but it is worth clarifying in writing if much is at stake.
• How long the output is kept differs by plan and by provider. According to Google's developer documentation the generated video is deleted after two days from their servers; elsewhere it is at least seven. If you do not download it, it is gone.
• The free plan rarely grants commercial rights. It was like that with image generation, and it is no different here.
• EU availability differs function by function. It is not the provider that is available or not, it is the individual capabilities.

The open-weight line, which is narrower from the EU

This is the most frequently misquoted part of the field, and by 2026 it has narrowed rather than widened.

The model leading the rankings is not downloadable. Alibaba's top model has no open version; the public release is a generation behind, under an Apache licence. Anyone who believes the top model can be run at home is mistaken.

Two popular open licences explicitly exclude the Union. Tencent's Hunyuan licence defines the territory so that the EU, the United Kingdom and South Korea are carved out of it, and spells out in capitals that the licence does not apply there. MiniMax's H3 community licence is narrower still: alongside the EU, the United Kingdom and South Korea, the excluded territories include the United States as well, and above twenty million dollars of annual revenue a separate permission is required. That one hurts, because MiniMax H3 leads the image-to-video ranking — meaning the best model in that category is not usable from the EU with open weights.

There are two large video models cleanly usable from the EU today: an Apache-licensed Alibaba release from a generation back, and the LTX line, which is free up to ten million dollars of annual revenue. Read the licence before you download, not after.

How to choose

1. Write down THREE real jobs you would do in the next month.
   Not "videos" — specifically: a 15-second product clip, the
   opening of a training video, the motion for a social post.

2. Check which of them NEEDS a clip longer than 8 seconds.
   If one does, that narrows the field to two providers.

3. Check whether you work with UPLOADED video (editing, extension).
   From the EU this is the tightest bottleneck.

4. Run the same three jobs on two tools, 3 attempts each.
   Do not look for the prettiest: look at HOW MANY OUT OF NINE are usable.

5. Read the ownership and retention parts of the terms.
   Five minutes, and it causes the most damage when it is skipped.

Common misconceptions

• "The model that leads the rankings is the best choice." Two lists, two orders, and the editing order is different again. The job decides.
• "If it is open-weight, I can run it at home and use it freely." The top models are closed, and two widespread open licences explicitly exclude the EU.
• "The interface is what matters." Most interfaces run the same models. Within one subscription, the choice of model produces a twentyfold price difference — the difference between interfaces is smaller than that.
• "The more expensive model is proportionally better." With a tenfold price difference the quality difference is not tenfold — what differs most is resolution and length.
• "What was best last year is still good." In this market the field turns over in nine months, and products get discontinued.

What to take away

Do not look for the best tool, look for the one that delivers your three jobs most reliably — and make the decision on the "how many out of nine are usable" question. Give the terms five minutes, and if you work from the EU, check availability function by function. On the background to the capabilities and the limits, How AI Video Works is the place to start, on sound there is Native Audio to read, and on the repair side Editing AI Video goes through what is possible today. The same decision for image generation is what Choosing an AI Image Tool walks around, one category over.