Sora
OpenAI's text-to-video model, unveiled in February 2024, which generates video from text by denoising spacetime patches with a diffusion Transformer.
Also OpenAI Sora · Sora 2
OpenAI, 2024
You type 'a woolly mammoth walking through snow' and get a short, realistic film clip of it, invented from scratch.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Uses / built on
How it works
In February 2024, OpenAI showed Sora: up to a minute of video generated from a sentence, with a level of consistency earlier models had not reached.
OpenAI’s technical report describes the approach, without full details:
- A video compression network turns raw video into a compact latent representation across space and time.
- The latent video is cut into spacetime patches, the video equivalent of the image patches in a Vision Transformer.
- A diffusion Transformer learns to turn noisy patches into clean ones, conditioned on text.
Because patches work for any size, the same model can generate different resolutions, durations and aspect ratios. OpenAI also reported that scaling up training compute markedly improved quality, the same lesson as scaling laws for language models.
The successor, Sora 2, arrived in September 2025 with audio and a social video app. Product availability has changed since then, so check OpenAI’s site for the current status. See text-to-video for how the field works in general.
Everyday use cases
- Case 01
How you've probably met it
Through viral demo clips in 2024 and AI-generated videos in social feeds, including from the Sora 2 app released in September 2025.
Strengths, weaknesses and when to use it
Strengths
- Showed a large jump in coherence and length of generated video.
Weaknesses
- Physics errors and objects that appear or vanish; deepfake and copyright concerns.
- Expensive to run; its availability as a product has changed several times.
When to use itAs the reference example of diffusion Transformers for video.