What happens when you ask for a video
A video model does not draw frames one after another. It handles the whole clip at once: there is a compressed, "latent" space in which picture and time are present together, and the model denoises inside that space — the same way image models do, just with one more dimension. That is why you cannot "continue" a finished clip the way you would continue a sentence: the model has to plan the entire stretch in one go.
From that follows the single most important limitation. Clip length is not a setting, it is a ceiling that falls out of how the model is built. Ask for longer than it was trained on and you do not get something slower, you get nothing. That is why every longer AI video today is assembled from short clips, and why building a minute out of eight-second clips has a post of its own.
What it can do today
The field moves fast, but the orders of magnitude are stable. These figures are from September 2026 and come out of vendor documentation.
- Clip length: the widely available models give you four to eight seconds per request. The two at the front of the field manage thirty seconds — the biggest jump of the past year.
- Resolution: 720p everywhere, 1080p in most places, 4K at a single provider, and there only on the longest clip.
- Frame rate: 24 fps is the default. That is no accident: it reads as cinema, and there are fewer frames to invent.
- Sound: on the leading models it comes by default, not as a separate step. A year ago that was not the case, and it is what made short video usable at all.
The length you actually get
Here is a trap that is easy to walk into. Several providers have an "extend" feature that continues a finished clip, which makes it look as though eight seconds were not really a limit. But extension has a ceiling of its own, and it is surprisingly low: at one large provider thirty-seven seconds in total, and only at 720p. In other words, a chained clip does not reach a minute.
The other thing is that extension is not free in quality terms. The vendors write down the side effects themselves: at one of them, text in the prompt that does not belong "may cause a cut or a dissolve", at another the extended part plays at a different volume from the original and the output can come out a few tenths of a second shorter than requested. These are not bugs, they are what the method brings with it.
What still does not work
Five things the whole field is weak at — not one model or another, all of them.
- Physics. Objects pass through each other, liquid behaves wrongly, mass has no weight to it. A benchmark built for exactly this states it in peer-reviewed form: visual realism does not amount to physical understanding.
- Continuity across cuts. If the same character appears in two clips, the second one will show someone else. There is a separate post about this, because a method for it does exist.
- Text in the picture. The same trouble as with stills, only more obvious in motion: the caption changes from frame to frame.
- Hands and faces close up. Fine from a distance, still not fine up close.
- Predictability. Run the same prompt twice and you get two different videos. There is a seed, but one vendor says this about it word for word: "does not guarantee determinism, it only slightly improves it."
The language nobody mentions
This section matters more the further your audience sits from English. One of the largest providers writes a single word into the "supported prompt languages" field of its documentation: English. And among the limitations it states that the other languages were not evaluated. Writing the prompt in English costs you nothing if English is what you work in — but it does tell you how narrow the tested ground is.
With native sound the ground gets narrower still. The language lists of the big models run to five and eleven entries respectively, and those lists cover speech, not ambience. The practical consequence is simple: for any market outside those lists, do not make the video model talk. Generate the scene without dialogue, and add the voice afterwards from a separate audio provider — which is what AI voiceover goes through in detail, languages included.
What it costs
Prices are per second, and the spread across the field is tenfold.
- The cheapest, fast models sit at around five cents a second.
- The middle tier runs ten to twenty cents.
- The most expensive, 4K output is sixty cents a second — which puts an eight-second clip at almost five dollars.
- Generation time, according to vendor figures, runs from eleven seconds to six minutes per clip.
Do the arithmetic: the raw generation cost of a one-minute, mid-quality video lands around ten dollars — and that covers only the clips that worked. The real cost has to be read together with the attempts you threw away.
Why this post goes stale fast
One concrete case shows the pace. In December 2025 a provider announced its new model on the strength of sitting first on an independent leaderboard. Nine months later that same model is not in the top ten, and the rival product it beat in the announcement does not exist any more — it was retired.
So memorising version numbers in this area is wasted effort. What is stable: the mechanism, the character of the limits, and the workflow. What is not: the names, the numbers and the ranking. For which tool is good for what right now, choosing an AI video tool gives you a snapshot, refreshed every quarter.
Common misconceptions
- "You can make a film with it now." You can make a scene. A film comes out of that once somebody edits it together, and that job is bigger than the generating.
- "Extending solves the length problem." It does not: it has a ceiling too, and quality degrades along the way.
- "With a good prompt I get what I pictured." In video the spread is far wider than in stills. Three or four attempts is the normal rate, not a sign of failure.
- "My own language works too, the chatbot understands it." The video model's text encoder is a different one, and the vendors explicitly evaluated English only.
What to take away
In 2026 AI video is a factory for short scenes, not a film studio. Work with that in your head — eight to thirty second pieces, an English prompt, a separate audio track wherever speech matters, an edit at the end — and you get usable material. Expect a finished video out of a single request and you will be disappointed. The very first decision is whether you start from text or from a picture, and text-to-video or image-to-video is the post that weighs the two against each other.