Sora is OpenAI's gambit to build a 'world simulator'

OpenAI just pulled back the curtain on Sora, a new diffusion model that generates up to a full minute of high-fidelity video from a text prompt. It’s not just a longer clip generator. The company frames the project as a deliberate step toward ‘general purpose simulators of the physical world,’ a phrase that elevates the ambition well beyond anything we’ve seen from Runway, Pika, or Meta’s efforts. The technical report is light on implementation details—no model architecture specifics or training recipes here—but heavy on a philosophical shift in how to approach visual data. The core idea is to treat video patches the way large language models treat text tokens, borrowing a page directly from the LLM scaling playbook. By compressing raw video into a lower-dimensional latent space and then slicing it into spacetime patches, Sora can train on a chaotic mix of content: widescreen, vertical, high-res, low-res, 4 seconds or 60. It’s a single model that handles all of it natively, without the standard pre-processing step of cropping everything to a uniform square. That matters more than it sounds. The team found that training on native aspect ratios directly improves framing and composition, avoiding the all-too-common generative artifact where subjects get awkwardly decapitated by the edge of the frame. The report includes a side-by-side comparison showing exactly this, with the square-crop baseline leaving a dog partially out of view while Sora’s version keeps the subject centered. Scaling behavior tracks with what we’ve come to expect from transformers: as training compute increases, sample quality gets markedly better across fixed seeds and inputs. The model is a diffusion transformer, an architecture that’s already proven its worth in image generation, now demonstrating it can scale effectively for video too. What’s conspicuously absent is any mention of the training data—its source, volume, or whether it includes copyrighted material scraped from the web. For a report that opens with the word ‘simulator,’ that omission feels significant. The decoder can map generated latents back to pixel space at variable resolutions up to 1920x1080, and the same model handles both video and static images since they’re just single-frame clips in this framework. There’s no public release announced, no API timeline, just a clear signal that OpenAI considers video generation a solved scaling problem and is already thinking about what comes after photorealism—simulating the physics, causality, and object permanence of the real world. Whether Sora actually grasps those things or merely mimics them convincingly is the billion-dollar question the report leaves deliberately unanswered.