OpenAI previews Sora, its text-to-video model

Background

A text-to-video model takes a written description and produces moving pictures instead of text or a single still image. That is a harder task than generating a photograph, because the model has to keep the same object looking consistent from one frame to the next and has to make motion read as continuous rather than as a series of unrelated pictures. A prompt describing a scene, a subject and a camera movement has to become a clip that a viewer can watch without noticing obvious breaks.

The preview

On 15 February 2024, OpenAI previewed Sora, its text-to-video model, and showed sample high-definition video clips generated with it^[1]. OpenAI said the model could generate videos up to one minute long^[1][2]. The preview was not a release: the model was not made available at the time, so the published sample clips were the main evidence people outside the company had of what it could produce^[1][2].

From preview to release

Access came later, and it arrived in stages. In November 2024, some testers leaked API access on Hugging Face in protest, and OpenAI revoked that access^[1]. The public release followed on 9 December 2024, for ChatGPT Plus and Pro users in the United States and Canada^[1][2].

The release was divided by subscription tier, and the tiers differed mainly in how many videos a subscriber could make and how long and how sharp each clip could be. Plus subscribers could generate up to 50 videos per month, at up to 720p and up to five seconds long^[2]. Pro subscribers could generate 500 videos per month, with clips up to 20 seconds long at 1080p^[2].

What the limits mean

Generating video is costly: every extra second of footage and every step up in resolution means more computation for a single request, so limits on length and resolution are one of the practical ways a video model gets shaped before it reaches users. The limits also make the preview and the release two different things. The one-minute figure described what the model could do in a demonstration^[1][2]; the five-second and twenty-second clips describe what a particular subscription allowed at launch^[2].

Why it matters

Sora made a text-to-video model reachable through an existing subscription product rather than through a separate, invitation-only research channel: the release put it in the hands of ChatGPT Plus and Pro users in two countries on the same day^[1][2].

The path from the February preview to the December release is also part of the story. Between the two, access was something people argued over and, in one case, took: testers who had been given API access leaked it on Hugging Face in protest in November 2024, and the company withdrew it^[1]. For a model that is demonstrated publicly long before it is sold, the questions of who gets to try it, under what terms, and for how long are settled in the months between the demo and the launch, not at the moment the first clip is shown.