Veo
Google's Video Model, With Sound Built In
OpenAI's text-to-video model closed on 26 April 2026, less than seven months after its social app hit number one. The API follows on 24 September 2026. What it was, why it ended, and where to go instead.
OpenAI discontinued the Sora web and mobile experiences on 26 April 2026. The Sora API is scheduled to follow on 24 September 2026. Visiting sora.com today returns a single line: Sora is no longer available. Video generation was removed from ChatGPT as well — there is no surviving consumer route to the model.
If you made anything on Sora, exports are handled at sora.chatgpt.com/sunset. OpenAI's own guidance is blunt: after the final export window closes, data associated with your use of Sora is permanently deleted. Credits you bought can be redirected to Codex, which tells you something about where the company's attention went.
We keep shut-down tools in this catalogue on purpose. What a product promised, and how long it lasted, is part of the record — and Sora's record is worth reading closely, because almost nothing about its failure was technical.
Sora arrived in February 2024 as a research preview and immediately reset expectations. Minute-long clips with coherent camera moves and stable subjects, generated from a sentence, at a moment when competing systems struggled past four seconds. It was the demo that convinced a lot of people that generative video was arriving faster than anyone had planned for.
The public product landed in December 2024 for ChatGPT subscribers: a web app with a storyboard timeline, remixing, blending two clips and looping. Useful, constrained, and clearly a first version — the gap between the launch demos and what a paying user could actually get out of it was the first sign that the hard part was not the model.
The second act came in late 2025. Sora 2 added synchronised audio — dialogue, effects, ambience generated with the picture rather than dubbed onto it — and shipped inside a standalone social app rather than a creative tool. The app's signature feature, Cameos, let you upload your own likeness and voice so friends could drop you into their generated clips. It reached number one on the App Store within days.
OpenAI announced the closure in March 2026 without giving a public reason. What follows is the reporting and the visible evidence, not a company statement.
Downloads peaked around 3.3 million in November 2025 and had fallen to roughly 1.1 million by February 2026. The app slid from the top of the App Store to somewhere in the hundreds. A feed made entirely of generated video turned out to be something people wanted to look at once and not return to — novelty is not retention, and the numbers said so within a single quarter.
Video inference is expensive in a way text generation is not. Reporting put Sora's running costs in the region of a million dollars a day against lifetime in-app purchase revenue estimated at about two million dollars in total. Treat those figures as journalism rather than filings, but the shape is not in dispute: a free social feed of the most compute-hungry output in AI was never going to pay for itself.
Cameos and the model's willingness to render recognisable characters produced a steady stream of trouble — impersonation, deceased public figures, studio-owned properties appearing in clips their owners had not agreed to. OpenAI tightened controls repeatedly and moved rightsholders from opting out to opting in. Each fix narrowed what the app was fun for, which is a hard place for a consumer product to sit.
The shutdown landed while OpenAI was visibly consolidating around coding tools, agents and enterprise revenue. Sora's credits being made redeemable against Codex is the clearest possible signal of that reordering. Compute spent on a social video feed is compute not spent on the products that were growing.
The video generation market did not contract when Sora left it — it is more crowded now than when Sora launched. Nothing here reproduces the Cameos social experience, because nobody else is trying to; what they replace is the actual work of generating video.
For production work with editing tools around the model, Runway remains the most complete environment and the one most used by people who ship. Google's Veo is the closest match to Sora's original strength — long, coherent shots with audio — and reaches most people through Google's own subscriptions. Kling and Hailuo built serious followings on motion quality and on being available when others were not.
Luma and Pika sit at the accessible end, quick to try and cheap to iterate on. If you need something self-hosted or licence-flexible, the open-weight video models that appeared through 2025 are now good enough for real work and immune to exactly the failure mode described on this page. That last point deserves weight: a vendor closing a product is a risk you can price in advance.
Whatever you pick, export your finished work as you go. Sora's users had roughly a month between the announcement and the app going dark.
Sora was not shut down because the model was bad. By any technical measure it was among the best video systems ever released to the public, and it was killed anyway. That is the useful lesson, and it applies well beyond OpenAI.
Generative video is the most expensive consumer AI product to run and the hardest to monetise, because the people who value it most — studios, agencies, marketers — want control, rights clearance and repeatability, none of which a social feed provides. Sora chased the audience that costs the most and pays the least, and the arithmetic caught up inside two quarters.
For anyone building on top of an AI service, the second lesson is plainer. A product backed by the best-funded company in the field, sitting at number one on the App Store, was gone in seven months. Keep your source material, keep your exports, and prefer tools whose disappearance would cost you a subscription rather than an archive.
No. The web and mobile apps closed on 26 April 2026 and video generation was removed from ChatGPT. The API remains until 24 September 2026 and then ends too.
Only through the API, and only until it shuts down in September 2026. Any site claiming to offer consumer access to Sora after that is not offering what it says it is.
Use the export tool at sora.chatgpt.com/sunset and wait for the email. Do it now rather than later — OpenAI has said data will be permanently deleted once the export window closes.
OpenAI directed subscribers to its standard ChatGPT refund process. Credits already purchased can be spent on Codex instead, which for many users is the practical outcome.
No official reason was given. The visible factors were collapsing usage after a spectacular launch, running costs far above what the app earned, persistent problems with likenesses and copyrighted characters, and a company-wide shift toward coding and enterprise products.
For serious production work, Runway. For long coherent shots with generated audio, Google's Veo. For fast and cheap iteration, Luma or Pika. None of them replicate Cameos, and no mainstream tool is likely to try.
Sora was the product that made generative video feel inevitable, and then demonstrated that being technically extraordinary is not a business model. It ran for under two years as a public product and under seven months as the app most people meant when they said its name.
If you used it, export your work before the window closes. If you are choosing a video tool now, choose one whose economics you can see, and keep your own copies of everything you make. Sora's users learned that lesson on a deadline.