AI video tools in 2026: what actually ships, and what it costs

For three years the interesting question about AI video was whether it could produce something watchable. That question is settled. The models generate clean, coherent, well-lit footage, and the gap between the best and second-best output is now small enough that most viewers could not reliably tell you which tool made which clip.

Which means the interesting question has moved. It is no longer can this produce usable footage but what does it cost to get footage I can actually ship, and can I get it into my edit without rebuilding my whole process around it. Those two questions have very different answers depending on which tool you pick, and the answers are not visible in any showreel.

The bottleneck moved from quality to iteration

The number that matters is not the price of one generation. It is the number of generations required before you get a shot you would put in front of a client.

That ratio varies enormously — by tool, by shot type, and by how specific your brief is. A generic atmospheric establishing shot might land on the second attempt. A shot requiring a particular character to turn their head and say a particular line might take twenty, and might never land at all.

Price your work on accepted output, not attempted output. A tool that costs twice as much per generation but halves your retry count is the cheaper tool, and the pricing page will never tell you that.

This is why per-second pricing is so misleading as a comparison tool. Two products can advertise near-identical rates and differ by a factor of five in what a finished minute actually costs you.

How to measure your own retry rate

Before committing to a subscription, run the same three shots through every tool you are considering:

1. An easy shot — a landscape, an abstract texture, something with no subject that needs to stay consistent.
2. A medium shot — one subject, simple motion, no dialogue.
3. The shot you actually need — whatever the hardest recurring requirement in your real work happens to be.

Count generations to an acceptable result for each. That number, multiplied by the per-generation cost, is your real price. It is usually the only benchmark worth running, and it takes an afternoon.

Consistency is still the hard problem

Everything difficult about AI video reduces to the same underlying issue: the model has no persistent memory of what it made a moment ago. Each generation is a fresh roll.

In practice this shows up as:

• Character drift. The same person, described identically, comes back with a different face, a different build, or different clothing between shots.
• Set drift. A room’s layout, lighting or contents reorganise themselves between cuts that are supposed to be continuous.
• Text. On-screen lettering remains unreliable. Signage, screens and labels come back as convincing-looking nonsense often enough that you should plan to composite any text you need legible.
• Hands and interaction. Improved substantially, still the first thing a viewer notices when it goes wrong.

The workarounds are all versions of the same idea: give the model less to invent. Reference images, locked seeds, image-to-video rather than text-to-video, and shorter clips assembled in the edit rather than longer generated sequences. Every one of these trades creative range for predictability, and for commercial work that trade is almost always correct.

What this means for how you brief

The teams getting good results are not writing more elaborate prompts. They are building small libraries of reference material — character sheets, location plates, lighting references — and feeding those in on every generation. The prompt describes the action; the references hold everything that must not change.

This is a meaningful shift in where the skill sits. It rewards people who think like art directors rather than people who think like prompt engineers.

Rights and licensing are the quiet risk

The commercial terms attached to these tools are considerably less settled than the technology, and they are where a project can go badly wrong after delivery.

Three specific things to check on the exact plan you are paying for:

• Commercial use. Several tools grant it only above a certain tier. A free-tier test clip that finds its way into a delivered edit is a real and recurring problem.
• Training on your inputs. Some tools reserve the right to train on prompts and uploads unless you are on a business plan. If you are uploading a client’s unreleased product photography as reference, this matters a great deal.
• Indemnification. A few vendors now offer some protection if generated output draws a claim. Most do not. Know which side of that line your tool sits on before you need to know.

Keep a dated copy of the terms you agreed to. These documents change more often than pricing does, and “the terms said something different when I generated this” is worth being able to demonstrate.

Where it genuinely works today

Cutting through the discourse in both directions, the honest picture is that AI video is production-ready for a specific and growing set of jobs:

• B-roll and texture. Atmospheric footage that fills space under a voiceover. This is the clearest win and it is displacing stock libraries quickly.
• Concept and previsualisation. Showing a client an idea before committing budget. Enormously faster than a storyboard and considerably more persuasive.
• Backgrounds and set extension. Generated plates behind live-action foregrounds, where continuity requirements are lower.
• Transitions and motion design. Short abstract sequences where nothing has to stay consistent because nothing is a subject.

And it remains unreliable for narrative work with recurring characters, anything requiring accurate lip-sync, and anything where a specific real product must appear exactly as it exists.

What to do about it

If you are evaluating tools this quarter, the shortest useful version:

Run the three-shot test above against your real work rather than against a demo brief. Compare on accepted output, not per-generation price. Check the licence on the specific tier you will actually pay for, and keep a copy. Choose the tool that fits your existing export and review process, because the time you lose fighting a pipeline mismatch will exceed anything you gain from marginally better generation quality.

And keep the evaluation cheap to repeat. The differentiator six months from now will not be the one that matters today, and any decision you make should be one you can revisit without having rebuilt your entire workflow around a single vendor.