Sora AI Video Generator: How to Create Stunning Videos with Text

Sora is being switched off.

OpenAI announced the discontinuation in March 2026. The app and web experience closed on 26 April 2026, and the API shuts down on 24 September 2026 — which, if you are reading this near publication, is roughly a month away.

If you have anything in production calling that API, that is the useful sentence in this article and the rest is context. Migration is now a deadline rather than a plan.

What Happened

OpenAI did not give a detailed public reason. What has been reported points at an unglamorous combination.

Running costs were substantial — figures around a million dollars a day circulated in coverage. Compute was constrained, and video generation is dramatically more expensive per output than text or images. Meanwhile usage had fallen from a peak around a million users to noticeably fewer, and the company was visibly reorienting toward enterprise products.

So: an expensive service with declining usage at a company with better uses for its compute. That is a straightforward commercial decision rather than a technical failure, and it is worth being precise about the distinction.

What It Was

Worth recording accurately, because the guides still circulating describe a product you cannot use.

Sora generated video from text descriptions, and at launch it was genuinely startling — coherent motion, consistent subjects across a clip, and a grasp of physical plausibility that earlier systems lacked. The second generation added audio and social features and integrated more tightly with OpenAI's other products.

It was, for a period, the most discussed text-to-video model in existence. That did not save it, which is the part worth carrying forward.

If You Built On It

Check what actually calls it. Production integrations, scheduled jobs, internal tools someone built and forgot. The deadline is fixed and there is no extension to negotiate.

Export anything you want to keep. Generated output stored only in the service goes with the service.

Migrate to an alternative — options below — and this time put an abstraction layer between your code and the provider. The lesson here is not that OpenAI is unreliable; it is that any hosted model can be withdrawn, and a codebase that calls one vendor's API directly makes that an emergency instead of a config change.

Where The Work Goes

Runway is the most established general-purpose alternative, with a mature editing environment around the generation rather than generation alone. For anyone doing actual production work, that surrounding tooling matters more than a marginal quality difference.

Pika is the lighter, faster option, well suited to short-form and social output.

The major labs' own video models — Google's in particular — are worth evaluating if you are already inside one of those ecosystems, since integration usually beats standalone capability in practice.

Open-weight video models are the option that did not meaningfully exist when Sora launched and does now. They are behind the best hosted models on quality and they cannot be discontinued out from under you, which after this episode is a consideration with a concrete price attached.

AI video generation tools covers the current landscape properly, and AI video editing tools covers the post-production side, which is where most of the actual work happens regardless of what generated the clip.

The Part Worth Generalising

The instructive thing here is not that a product closed. It is which product closed.

Sora was not a failed experiment. It was technically celebrated, heavily covered, and backed by the best-resourced company in the field. It still could not sustain its own running costs at the usage it had.

That should adjust how anyone plans around hosted AI services. Capability is not the same as durability, and the gap between them is filled by economics nobody outside the company can see. A model that is impressive today can be withdrawn on a quarter's notice for reasons entirely unrelated to how well it works.

Two practical consequences. Abstract your provider so switching is a configuration change rather than a rewrite. And weight open weights higher than a pure capability comparison suggests — the argument for a model you can run yourself was always about privacy and cost, and continuity has now been added to the list, with a worked example.

The wider version of this argument is in DeepSeek, which is the clearest case of open weights being the point rather than the compromise. And AI trends 2026 covers how much of this year's consensus turned out to be wrong, of which this is now one more entry.

If You Arrived Looking For A Tutorial

There is not one to give, and any guide you find explaining how to use Sora is describing something that no longer exists.

If you want to generate video, start with the alternatives above. If you specifically wanted Sora's output quality, the honest answer is that the closest current options are near it rather than at it, and the gap is narrowing quickly enough that it will not matter for long.