What is Sora 2, and how does the new Sora app work?

Yesterday, Tuesday, September 30, 2025, OpenAI released Sora 2, the successor to the Sora video generator it first showed in February 2024. You describe a scene in words and Sora 2 produces a short video, now with synchronized audio: dialogue, background sound and effects that match what happens on screen.

OpenAI did not stop at the model. It also launched a new iPhone app called Sora, a social app with a scrolling feed, where people make videos, remix each other's work and, with permission, put themselves and their friends into scenes.

What OpenAI released

OpenAI describes the first Sora as the "GPT-1 moment for video," the point where generated video began to work at all, and suggests Sora 2 may be video's "GPT-3.5 moment," the stage where a technology becomes good enough for wide use. The concrete improvements:

Area	What changes in Sora 2
Physics	Objects behave more plausibly; a missed basketball shot bounces off the backboard instead of jumping into the hoop
Control	Follows detailed instructions across several shots and keeps the scene consistent between them
Styles	Realistic, cinematic and anime looks
Sound	Speech, sound effects and background soundscapes generated together with the picture
Real people	Can insert a person, animal or object it has observed, with their look and voice

OpenAI is frank that the model still makes plenty of mistakes. Its interesting observation is that many errors now look like errors by the people in the scene, a gymnast who stumbles, rather than reality bending to satisfy the prompt.

How a video model works

A video model learns from huge amounts of video. Instead of predicting the next word, as a chatbot does, it learns to produce sequences of frames that look and move like real footage. OpenAI frames this as building a world simulator: to make a convincing clip of a paddleboard backflip, the model has to capture something about weight, balance and water.

Older models were, in OpenAI's word, "overoptimistic." Asked for a successful shot, they would bend reality to deliver it. Being able to show failure, the ball that misses, is a sign the model has learned more about how the world behaves.

Prompts work best when they read like a short shot description. For example:

Prompt

A rainy evening at a small bakery. Through the window, a baker in a
flour-dusted apron slides a tray of croissants onto the counter and
says, "Last batch of the day!" Soft rain on the glass, a bell rings as
the door opens. Warm light, handheld camera, 10 seconds.

That prompt names the subject, the action, a line of dialogue, the sounds, the lighting and the camera, which gives the model clear targets for both picture and audio.

The Sora app and cameos

The Sora app is where OpenAI wants most people to meet the model. You can create videos, remix others' creations and browse a feed. Its signature feature is cameos: after a one-time recording of video and audio in the app, which verifies your identity and captures your likeness, you can drop yourself into any Sora scene. You can also let friends use your cameo.

TechCrunch compares the feed to TikTok and Instagram Reels. OpenAI says its feed is designed differently: by default it favors people you follow or interact with, and videos likely to inspire your own creations, and it is not tuned to maximize time spent scrolling. You can steer the recommendations by describing what you want in plain language, and the app will periodically ask how you are feeling about it.

Safeguards and open questions

Using someone's likeness raises obvious risks, and OpenAI lists several protections:

You alone choose who may use your cameo; you can withdraw that permission later and remove any video that shows you, even drafts made by others.
Teens get daily limits on how many feed videos they see and stricter cameo permissions; human moderators are being added to handle bullying.
Parents can use ChatGPT's parental controls to override scroll limits, turn off personalization and manage direct messages.

TechCrunch points out what these controls cannot fully prevent: a friend you trust with your cameo could still make a misleading video of you. It also notes that recommendations may draw on your location from your internet address and on your ChatGPT history, which can be switched off.

On money, OpenAI says its only current plan is to let people pay for extra videos if demand outruns its computing capacity.

Try it: Write a ten-second scene for Sora 2 that includes one line of dialogue and two specific sounds. Then rewrite it with a deliberate failure, such as a juggler dropping a ball, and see which version looks more natural.

How do you make your first video with Sora 2?

Download the Sora app from Apple's App Store. Access is by invitation, rolling out first in the United States and Canada; you can sign up in the app to be notified when your account is let in. Once invited, you can also use Sora 2 on the web at sora.com. Sora 2 starts out free with generous limits that depend on available computing power, and ChatGPT Pro subscribers can try the higher-quality, experimental Sora 2 Pro on sora.com.

For your first clip, keep it short and specific: one setting, one action, one or two sounds. Watch the result, then change a single detail at a time, the camera, the lighting, the line of dialogue, so you learn what each part of the prompt does.

If you want to appear in your videos, record your cameo in the app and decide right away who may use it. Start with only yourself, and widen it to close friends once you are comfortable with how it works. Everything you made with the first Sora stays in your sora.com library.