What was OpenAI Sora? The video AI, explained

In plain words

Sora was a text-to-video model made by OpenAI, the company behind ChatGPT. You wrote a short description, such as "a paper boat sailing down a rainy street", and it produced a video clip of that scene from nothing.

OpenAI first showed Sora in February 2024 and opened it to paying ChatGPT users that December. In September 2025 it launched Sora 2, which added matching sound and speech, along with a Sora phone app where people shared and remixed each other's clips.

In March 2026 OpenAI announced it was ending Sora. The app closed on 26 April 2026 and the service for developers on 24 September 2026. OpenAI said it would keep the work going as research.

Everyday example

A flip book drawn from a sentence

Remember flip books, where each page shows a drawing slightly changed and flicking them makes it move? Imagine handing an artist one sentence and getting back a whole flip book: the boat, the rain, the ripples, each page consistent with the last. Sora made the pages and kept them matching, in seconds.

Where the example breaks: A flip-book artist knows that water flows and boats float. Sora only learned how video tends to look, so objects could merge, vanish or move in impossible ways.

Latestas of October 2026

Status: Closed: app and website shut on 26 April 2026; developer API shut on 24 September 2026
Your videos: OpenAI's help page asked users to export their Sora content; anything not exported is deleted after the final export window

These details change often. The rest of this page explains what it is, which stays true.

How it actually works

OpenAI described Sora's design in a February 2024 technical report.

Squeeze the video

Training videos were first compressed into a smaller latent space, keeping what matters and dropping pixel-level detail.

Cut it into spacetime patches

The compressed video was split into small pieces covering a patch of the image across a few moments of time. These patches played the role that tokens play for text.

A diffusion transformer

OpenAI called Sora a diffusion transformer: a transformer that learns to remove noise from those patches step by step, guided by the prompt. See diffusion model.

Decode back to video

Once the noise was gone, a decoder turned the patches back into ordinary video frames. Sora 2 also produced matching audio, such as dialogue and sound effects.

Made by

OpenAI

First shown

15 February 2024; available to ChatGPT Plus and Pro users from December 2024

Sora 2 and app

Announced 30 September 2025

Shut down

App closed 26 April 2026; developer API closed 24 September 2026

Other companies offer text-to-video tools, including Runway and Google, whose models are covered under Google DeepMind.

Questions people ask

Is Sora still available?

OpenAI closed it in 2026. It announced the end of Sora on 24 March 2026, the Sora app and website closed on 26 April 2026, and access for developers ended on 24 September 2026. OpenAI said the work continues as a research project.

Why did OpenAI shut down Sora?

OpenAI gave few details in its announcement. Reporting at the time said the company wanted to move computing power toward coding tools and business customers, with Sora continuing only as research.

Is Sora the same as ChatGPT?

No. Both came from OpenAI, but ChatGPT is a chat app built on GPT language models, while Sora was a separate model and app for making video.

How did Sora make a video from text?

It started from random noise and removed it step by step, guided by your prompt, in the way a diffusion model makes an image. The difference is that it worked on short chunks of video across time, so frames stayed consistent with each other.