The technology behind Sora 2
How the model is built and what it was trained to do, drawn from official docs and independent reviews.
Sora 2 is OpenAI's video and audio generation model, released on September 30, 2025 through sora.com, a standalone iOS app, and later the API. OpenAI's system card describes it as a step beyond the original Sora, with more accurate physics, sharper realism, synchronized audio, enhanced steerability, and a wider stylistic range. OpenAI calls it what may be the GPT-3.5 moment for video, framing the 2024 Sora as the GPT-1 moment.
The headline improvement is physical plausibility. OpenAI's launch example is a missed basketball shot: earlier models would teleport the ball into the hoop, while Sora 2 lets it rebound off the backboard, and the company notes that the model's mistakes now tend to look like mistakes of the agent it is simulating rather than breaks in world physics. Sora 2 also generates dialogue and sound effects natively, and its cameo feature can insert a verified real person into generated scenes with their appearance and voice after a one-time consent recording.
OpenAI has not published architecture details for Sora 2, so specifics like parameter count or exact design remain undisclosed. The original Sora was described in OpenAI's 2024 technical report as a diffusion transformer operating on spacetime patches of compressed video latents. The system card states that Sora 2 was trained on diverse datasets, including publicly available internet data, third-party licensed data, and data from users and human trainers, with filtering and safety classifiers applied in the pipeline.
Every Sora asset carries C2PA provenance metadata, videos downloaded from sora.com or the app carry a visible moving watermark, and OpenAI runs internal detection tools to identify Sora-made media. At launch OpenAI deliberately limited the product: no video-to-video generation, no text-to-video of public figures, blocked generations of real people outside the consent-based cameo system, and restrictions on uploading images of photorealistic people. The company acknowledges these safeguards do not fully solve provenance and says mitigations will keep changing as risks emerge.