OpenAI launches Sora 2 and the Sora app

Background

A video generation model takes a written prompt and returns a sequence of moving images. Making those images look plausible is one problem; giving them sound is a separate one, and most such systems have produced silent clips, with audio added afterwards. OpenAI already offered a video model, Sora 1 Turbo, when it announced the next version ^[1].

What happened

On 30 September 2025 OpenAI unveiled Sora 2 and called it its flagship video and audio generation model. Picture and sound are generated together, so speech, noise and music line up with what happens on screen instead of being dubbed on later ^[1].

OpenAI also said the model obeyed the laws of physics more closely. Its illustration was a basketball that misses: the ball rebounds off the backboard rather than teleporting into the hoop, a kind of error video models have often made ^[1]. The claim is about continuity — an object that bounces has to travel, slow down and change direction in a way that matches what the viewer expects, and that is hard to keep consistent across hundreds of frames.

The app

The model did not arrive on its own. It launched with a new iOS app, available by invitation in the United States and Canada ^[2]. The app was not a plain gallery of clips: it was built around a social feed and around cameos, which let users insert themselves and friends into the videos they generated ^[2]. Videos made this way carried visible watermarks ^[2].

Shipping a generator as a feed changes who uses it. A website with a prompt box suits people who already know what they want to make; a feed with faces in it suits people who want to react to what others have posted. The watermark serves a different purpose: it marks a clip as synthetic wherever it travels, even after it leaves the app.

Availability and plans

Sora 2 rolled out first in the United States and Canada, initially free and with generous limits. Pro users could try a higher-quality version, Sora 2 Pro, on sora.com ^[1]. OpenAI said it planned to bring Sora 2 to its API, which would let other software call the model directly, and that Sora 1 Turbo would remain available ^[1]. Keeping the older model in place is a reminder that access through an API and access through an app are different products for different users.

Reaction

The launch was followed quickly by friction. Complaints from rights holders over copyrighted characters in generated videos led OpenAI to promise more control within days ^[2]. The episode put a question on the table that the technical announcement did not answer: when a model can put well-known characters and real people's faces into new footage, who decides what may be made, and how quickly can a maker of the model respond when the answer is disputed.

Sora 2 therefore stands as two launches at once — a model that generates sound and picture together, and a consumer app that puts generated video into a social loop, with watermarking and rights questions arriving alongside it ^[1][2].