AI Video Generation: Create Videos with Sora, Runway, and Pika Labs

The most important thing that happened in AI video this year was not a model release. It was a shutdown.

OpenAI discontinued Sora — the app closed in April 2026, the API closes on 24 September. This was arguably the most technically celebrated video model of its generation, from the best-resourced company in the field, and it could not sustain its own running costs.

That tells you more about where this category actually stands than any capability demo, and it should shape how you approach it.

Why That Matters For Your Decision

Video generation is expensive in a way text and images are not. Every second of output costs meaningfully more compute than a page of text, and that cost does not fall as fast as usage grows.

Which means the economics here are genuinely unsettled, and any tool you build a workflow around might not exist in eighteen months. That is not pessimism about the technology — the technology is improving quickly — it is a statement about the business model underneath it.

Two practical consequences, both of which most coverage skips. Do not build a production pipeline that depends on one provider's API without an abstraction layer. And weight open-weight models higher than a capability comparison alone would suggest, because a model you can run yourself cannot be withdrawn. That argument used to be about privacy and cost; this year it acquired a continuity dimension with a worked example. The Sora shutdown covers what happened in detail.

What Is Actually Available

Runway is the most established option for people doing real work. The important thing about it is not generation quality — it is the editing environment around the generation. Once you are assembling multiple clips, fixing timing, and iterating, that surrounding tooling matters considerably more than a marginal quality edge on a single output.

Pika is lighter and faster, and suits short-form and social output where turnaround beats polish.

The major labs' video models, Google's in particular, are worth evaluating if you already work inside one of those ecosystems. Integration with tools you use tends to beat standalone capability in practice.

Open-weight video models exist now in a way they did not when this category started. They trail the best hosted models on quality and they cannot be discontinued, repriced, or changed under you.

I am deliberately not ranking these on output quality. That ordering changes every few months, and it is the least durable basis for a decision.

Where Generated Video Is Genuinely Usable

The honest boundary, which most coverage blurs.

It works for B-roll and establishing shots, abstract and atmospheric sequences, backgrounds behind other content, concept and pitch work where the point is conveying an idea rather than shipping a final asset, and short social clips where production values are not the differentiator.

It does not work anywhere a viewer studies a face or a pair of hands, anywhere continuity must hold across shots — the same character, same location, same lighting — anywhere precise control over what happens is required, or anywhere a specific real place, product or person must appear accurately.

The failure mode is specific and worth knowing: output is frequently impressive for two seconds and then something goes subtly wrong. A hand acquires a finger, a face shifts, an object passes through another. Viewers notice, even when they cannot articulate what they noticed.

The Workflow Nobody Mentions

Generation is a minority of the work.

The realistic loop is: generate many candidates, discard most, take the usable seconds from what survives, and assemble them in a conventional editor. You are not producing finished video — you are producing raw material with a low hit rate.

Budget accordingly, in both time and cost. The failed generations are billed the same as the good ones, and there will be considerably more of them than any demo suggests.

This is also why the editing environment matters more than the generator, and why AI video editing tools is arguably the more useful read if you are actually producing something.

Prompting For Video

Different from image prompting in one way that matters: you are describing motion and time, not just a frame.

Say what moves and how. Camera behaviour — static, slow push, handheld — changes the output substantially and is frequently the difference between usable and not. Keep the described action simple; complex multi-step action is where coherence collapses. And specify duration, because generating longer than you need is how the artefacts arrive.

Expect to run the same prompt repeatedly. Variation between generations is high, and the good one is often the fourth attempt at something identical.

What To Actually Do

If you are experimenting, use a free tier on one of the established tools and calibrate on what the failure modes look like in practice. That teaches you where the boundary sits faster than any article.

If you are producing commercially, use it for the categories listed above and shoot or license everything else. The moment you find yourself fighting the tool to get a specific result, you have crossed the boundary and the conventional route is cheaper.

If you are building it into a product, abstract the provider. That is the lesson this year handed everyone for free.

For the wider creative stack, AI image generation covers still work where the tools are considerably more mature, and AI for YouTubers and content creators covers the production workflow around all of it.