What is Sora AI?
Sora was OpenAI's text-to-video AI: you typed a description and it generated a short video clip. OpenAI discontinued it — the app and website closed April 26, 2026, and the API shuts down September 24, 2026. It ran roughly $1 million a day while users fell from about 1 million to under 500,000.
Why — the first-principles explanation
Sora turned text into video using a diffusion transformer. Diffusion models learn by watching noise destroy a real video, then learning to reverse that destruction. Show the model millions of clips, progressively scramble each one into static, and train it to undo each step. Run that process from pure random noise while steering it with your text prompt, and the noise resolves into a video that matches your words. Sora did this in a compressed "latent" space and worked on 3D patches — chunks that span both space and time — which is why it could keep a subject roughly consistent as it moved across frames rather than flickering into a new object every frame.
That mechanism explains both what Sora was good at and why it died. Video is expensive in a way text never is. A sentence is a few hundred tokens; a few seconds of video is a huge grid of pixels across dozens of frames, and the model has to denoise the entire clip through many steps. Every clip you watched cost real GPU time. Text chat costs fractions of a cent per reply — video costs orders of magnitude more, and unlike a chat reply, a bad video gets thrown away and regenerated.
The economics never closed. TechCrunch, citing a Wall Street Journal investigation, reported Sora burned roughly $1 million every day while its worldwide user count peaked near 1 million and then fell below 500,000. A social video app has to make money from ads or subscriptions, and neither covers a dollar-a-clip cost base. On March 24, 2026, about six months after the Sora 2 app launched, OpenAI announced it was shutting the whole thing down and redirecting the compute to enterprise products. A planned $1 billion Disney character-licensing deal collapsed alongside it.
So Sora matters now mostly as a lesson: the hard limit on consumer AI video isn't whether the model can make something impressive. It's whether anyone can afford to let millions of people press "generate" for free.
An example that makes it click
Think of a photo left out in the rain until it's a blurry mess. Now imagine someone who has seen ten million ruined photos and their originals, so well that handed any blurry smear they can guess what it looked like before. Hand them pure random static and say "a golden retriever on a beach," and they clean it up one pass at a time — 30 or 50 passes — until a retriever on a beach is sitting there. That's diffusion.
Now make it a video. Instead of one photo, you're un-blurring 300 photos at once, and they all have to agree the dog is the same dog walking the same direction. That's why one 10-second clip could cost more computing power than a thousand chat messages — and why an app doing that for a million people, free, bled about $1 million a day until OpenAI turned it off.
Key facts
- OpenAI publicly previewed Sora on February 15, 2024, and released it to ChatGPT Plus and Pro users in the US and Canada on December 9, 2024.
- Sora 2 launched September 30, 2025 with an iOS app; the Android app followed roughly two months later.
- OpenAI announced Sora's discontinuation on March 24, 2026; the web and app experiences ended April 26, 2026, and the API is scheduled to end September 24, 2026.
- Sora used a denoising latent diffusion model with a transformer as its denoiser, generating video by denoising 3D patches in latent space.
- Sora's maximum video length was one minute, and every generated video carried a visible moving watermark.
- Per a Wall Street Journal investigation reported by TechCrunch, Sora cost roughly $1 million per day to run while users fell from a peak near 1 million to fewer than 500,000.
▶ The 60-second explainer (script)
Sora was OpenAI's text-to-video AI. You typed a sentence, and it made a short video. And it's gone — OpenAI shut down the app and website on April 26th, 2026. The API follows on September 24th. Here's how it worked. Sora was a diffusion model. It learned by watching real videos get scrambled into static, over and over, until it could run that backwards. Give it pure noise plus your prompt, and it un-scrambles the noise, step by step, into a video matching your words. It worked on chunks that covered both space and time, which is why the dog in your clip stayed the same dog as it walked. So why kill something that impressive? Money. Video is brutally expensive. A chat reply costs a fraction of a cent. A few seconds of video means denoising a huge stack of frames, dozens of times. The Wall Street Journal reported Sora burned about one million dollars a day, while users dropped from a million to under half a million. A billion-dollar Disney deal collapsed too. So the lesson isn't that AI video failed. It's that letting millions of people generate it for free never added up.
What authoritative sources say
People also ask
Can I still use Sora?
No. The Sora app and website were discontinued on April 26, 2026. The API remains available only until September 24, 2026, and OpenAI is not issuing new access for it.
What does the name Sora mean?
Sora is the Japanese word for "sky." OpenAI chose it to suggest limitless creative potential.
How long could Sora videos be?
Up to one minute. Every video also carried a visible, moving watermark to make AI-generated footage identifiable.
Is Sora the same as ChatGPT?
No. Both are OpenAI products, but ChatGPT is a text and image assistant while Sora was a separate video generator with its own app. ChatGPT still operates normally.
What can I use instead?
Google's Veo and Runway remain active text-to-video products as of July 2026. Check each vendor's current pricing directly, since this market changes fast.