With Sora 2, we are jumping straight to what we think may be the GPT‑3.5 moment for video. Sora 2 can do things that are exceptionally difficult—and in some instances outright impossible—for prior video generation models: Olympic gymnastics routines, backflips on a paddleboard that accurately model the dynamics of buoyancy and rigidity, and triple axels while a cat holds on for dear life.

Prior video models are overoptimistic—they will morph objects and deform reality to successfully execute upon a text prompt. For example, if a basketball player misses a shot, the ball may spontaneously teleport to the hoop. In Sora 2, if a basketball player misses a shot, it will rebound off the backboard. Interestingly, “mistakes” the model makes frequently appear to be mistakes of the internal agent that Sora 2 is implicitly modeling; though still imperfect, it is better about obeying the laws of physics compared to prior systems. This is an extremely important capability for any useful world simulator—you must be able to model failure, not just success.

The model is also a big leap forward in controllability, able to follow intricate instructions spanning multiple shots while accurately persisting world state. It excels at realistic, cinematic, and anime styles.

Prompt: Vikings Go To War — North Sea Launch (10.0s, Winter cool daylight / early medieval)...