Sora
Why it matters
Deep Dive
Sora is built on a diffusion transformer: raw video is compressed into a lower-dimensional latent space, chopped into spacetime patches, and denoised by a transformer that treats those patches much the way a language model treats tokens. When OpenAI showed the first clips in February 2024 — a woman walking down a neon-lit Tokyo street, golden retrievers playing in snow — the jump in length and coherence over earlier video generation systems like Runway and Pika was obvious, with single generations running up to a minute. Public access followed in December 2024 as Sora Turbo, a faster, cheaper variant bundled with ChatGPT subscriptions, and in 2025 OpenAI shipped a standalone app built around a cameo feature for inserting real people's licensed likenesses, alongside Sora 2 with synchronized dialogue and sound effects.
Patches Instead of Pixels
Under the hood, Sora adapts the diffusion model recipe that powered image generators to the time dimension. A compression network first squeezes raw video into a latent representation, and an encoder cuts that latent volume into spacetime patches — small cubes of space and time that play the role of tokens. The transformer then learns to reverse a noising process over these patches, starting from static and ending with a coherent clip, conditioned on the text prompt.
Two training tricks matter as much as the architecture. First, Sora trains on videos at their native durations, resolutions, and aspect ratios rather than cropping everything to squares, which lets one model serve vertical phone clips and widescreen shots alike. Second, OpenAI reused the recaptioning idea from DALL-E 3: a captioning model writes dense, descriptive text for training videos, so the generator learns from far richer supervision than the short alt-text most video datasets ship with. Neither trick is exotic, but at scale they are a large part of why prompt adherence looks so much better than in earlier systems.
From Research Preview to App
The February 2024 unveiling was a research preview only — a curated set of demos plus access for red teamers and a small group of artists — and it stayed that way for most of the year. The December 2024 launch of Sora Turbo on sora.com opened text-to-video, image-to-video, and remix tools to ChatGPT Plus and Pro subscribers, with clips capped well below the minute-long demos. OpenAI has never published parameter counts or training data specifics, so most outside understanding of the system comes from its technical report and hands-on use.
The 2025 app marked a shift from tool to platform. Its cameo feature lets a user record a short verification video and then license their likeness to friends, who can drop that person into generated scenes with their permission. Sora 2, released in late 2025, added synchronized audio — dialogue, sound effects, and ambience generated alongside the video — closing one of the biggest gaps with Google's Veo 3. The app launched invite-only and climbed download charts quickly, which also meant its content moderation decisions were stress-tested in public within days.
Where It Shines and Where It Breaks
Sora's strengths are cinematic: smooth camera moves, plausible lighting, multiple shot compositions within one generation, and a wide stylistic range from photorealism to animation. Image-to-video is often the more practical mode — feeding a still frame as an anchor gives far more control than a text prompt alone, and teams building product mockups or storyboards tend to start there. Extending an existing clip forward or backward in time also works surprisingly well for building longer sequences.
The failure modes are equally consistent. Physics still breaks in predictable places: objects pop into existence, hands and text garble under motion, and cause-and-effect sequences (pouring liquid, shattering glass) often look right frame-by-frame but wrong as events. Long or complex prompts get silently truncated in effect, with later clauses ignored, which makes prompt engineering for video a matter of iterating short, concrete scene descriptions rather than writing screenplays. Generation is also slow and compute-hungry compared with image models, so iteration cycles feel closer to rendering than to chatting.
It Is Not a World Simulator
OpenAI's technical report framed scaled video models as a step toward general-purpose simulators of the physical world, and the claim is worth pushing back on. Sora learns the statistics of how pixels evolve, not an explicit causal model of objects, forces, or agents. That is why it can produce a convincing ocean wave yet fail at a person biting a cookie and the cookie showing a bite mark — surface plausibility is not the same as state tracking. Treating it as a learned renderer rather than an emergent physics engine sets expectations correctly and explains both the magic and the breakage. The same caveat applies to claims that video models are close to a true world model in the reinforcement-learning sense: the representations are impressive, but nobody has shown they support reliable prediction and planning.
Consent, Provenance, and the Deepfake Question
A model this good at fabricating footage inherits the entire deepfake problem, and OpenAI's answers have been layered rather than absolute. Generated videos carry a visible moving watermark plus embedded provenance metadata, cameo requires recorded consent from the person being depicted, and policy filters block public figures and certain content categories — though each of these has been tested by determined users. Watermarks can be cropped or degraded, and provenance metadata survives only while platforms cooperate, which is why watermarking is best treated as one signal among many rather than a guarantee.
The rights side is equally unsettled. Rightsholders have pressed OpenAI over copyrighted characters appearing in generations, pushing the company toward opt-out and revenue-sharing arrangements, and the cameo model has become a test case for how likeness licensing might work at scale. For anyone building on generative video, the practical read is that copyright in AI outputs and training data remains a moving target, and provenance tooling should be designed in from the start rather than bolted on after a viral incident.