# What is Sora, and how can you make AI videos with it now?

Date: 2024-12-10 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-sora-openai-video-generator/

**Answer:** Sora is OpenAI's AI video generator, and since December 9, 2024 a faster version called Sora Turbo is open to ChatGPT Plus and Pro subscribers at sora.com. It turns a written description, an image or a video clip into new footage up to 20 seconds long and up to 1080p, with visible watermarks and limits on videos of real people.

Yesterday, December 9, 2024, OpenAI opened Sora, its video generation model, to the public. In February the company showed sample clips and a technical report but kept the model in a research preview. Now a new, much faster version called **Sora Turbo** is available as its own website, sora.com, to people who pay for ChatGPT Plus or Pro.

You type a description, such as "a paper boat drifting down a rainy city gutter, close-up, evening light," and Sora produces a short video. You can also start from your own image or clip and ask Sora to extend it, change it or mix it with another.

## What OpenAI released

The launch came on the third day of OpenAI's "12 Days of OpenAI," a run of daily announcements, during a livestream on Monday afternoon. Here is what subscribers get:

| | ChatGPT Plus | ChatGPT Pro |
|---|---|---|
| Monthly price | Sora included at no extra cost | Sora included in the $200 plan |
| Credits per month | 1,000 | 10,000 |
| Fast ("priority") videos | Up to 50, at 720p and 5 seconds | Up to 500, at 1080p and 20 seconds |
| Extra | — | Unlimited slower "relaxed" videos, without watermarks |

Credits are Sora's internal currency. Longer and sharper videos cost more of them, and the monthly allowance resets without carrying over. Sora is not part of ChatGPT Team, Enterprise or Edu, and it's closed to anyone under 18.

The site includes several editing tools. A **storyboard** lets you lay out what should happen at each moment of a clip. **Remix** changes an existing video, **Blend** merges two videos into one, and **Loop** and **Re-cut** help you tidy a scene. Featured and Recent feeds show what others are making.

## How Sora turns words into video

Sora is a **diffusion model**. Training starts with real videos that have been covered in random visual noise, like heavy static on an old TV. The model learns to remove that noise step by step and recover the clean picture. Once trained, it can start from pure static and, guided by your text, "clean" its way to a video that never existed.

OpenAI's technical report from February describes three ideas that make this work at video scale:

1. **Compression.** A separate network first squeezes each video into a much smaller representation, keeping the essence of what's on screen and how it moves. Sora works in this compact space, and a decoder turns the result back into normal frames.
2. **Spacetime patches.** The compressed video is cut into small blocks that each cover a patch of the image over a short slice of time. These patches play the same role that tokens, the word pieces, play in a text model like ChatGPT.
3. **A transformer.** The same kind of neural network behind language models processes those patches, which lets OpenAI scale Sora up with more data and computing power.

There is also a language trick. OpenAI trained a captioning model to write detailed descriptions of its training videos, the method it used for DALL·E 3, and it uses GPT to expand a short prompt into a richer description before Sora starts. That helps Sora follow what you asked for.

## Its limits and safeguards

OpenAI is frank about the weaknesses. Sora often produces physics that look wrong, and it struggles with complicated actions that unfold over many seconds. Will Peebles, a research lead on the project, said in the livestream that this early version will make mistakes. Demand was also a problem on day one: TechCrunch reported major capacity issues by Monday evening.

To make generated videos easier to identify, every Sora clip carries **C2PA metadata**, a technical label embedded in the file that records where it came from, plus a visible watermark by default. OpenAI says it blocks sexual deepfakes and child abuse material, and for now only a small test group can upload videos of real people.

Two practical notes. Sora is not available in the UK, Switzerland or the European Economic Area yet; Sam Altman said the company has no timing to share for Europe. And TechCrunch points out that OpenAI trains on new Sora videos by default. You can turn this off in the Sora settings, under Data controls.

## Why it matters

Until now, most people's experience of AI video was watching other people's demos. With Sora inside a subscription millions already pay for, making a short clip becomes as accessible as generating an image was with DALL·E. For teachers, small businesses and hobbyists, that means quick visual sketches: a mood clip for a presentation, a product idea, a scene for a story.

It also brings old questions into everyday use. Some early testers leaked access before launch to protest how OpenAI treated artists, according to TechCrunch. Watermarks and metadata help, but they don't settle questions about consent, copyright and misinformation. OpenAI says it is releasing Sora now partly so society can work out norms while the technology is still limited.

> **Try it:** If you have ChatGPT Plus, generate the same 5-second scene twice at 480p: once with a short prompt ("a dog on a beach") and once with a detailed one that names the camera angle, light and action. Compare which details Sora kept, and look closely at the physics: paws, waves, sand.

## How do you write a good first prompt for Sora?

Think like a film director describing one shot. A useful order is: the **subject**, the **action**, the **setting**, then the **camera** and **light**. Keep it to one clear scene, since Sora's clips are short and complex sequences are where it struggles most.

```prompt
A golden retriever runs along a quiet beach at sunrise, splashing through shallow waves. Low camera angle, tracking alongside the dog. Soft orange light, light mist over the water. Realistic, calm mood.
```

Start at 480p and 5 seconds to save credits while you experiment, then raise the quality once a prompt works. Use the storyboard when you need two moments in order, such as "the door opens" then "a cat walks in." And when a result looks almost right, use Remix to adjust one thing at a time instead of starting over.

## Sources

- Sora is here: https://openai.com/index/sora-is-here/
- Video generation models as world simulators: https://openai.com/index/video-generation-models-as-world-simulators/
- OpenAI's Sora video generator is launching for ChatGPT Pro and Plus subscribers — but not in the EU: https://techcrunch.com/2024/12/09/openais-sora-video-generator-might-not-be-available-in-the-eu-at-launch/