Sora 2's distinguishing strength is physical simulation — it models how things fall, splash, tear and settle. Prompts that describe cause and effect get results no other model matches; prompts that just list objects waste that entirely.
A ceramic mug slips from the edge of a marble counter, tumbles once in the air, and shatters on the tile floor, sending fragments skidding outward while coffee splashes up the cabinet door
A broken mug on a kitchen floor, coffee spilled, cinematic lighting, 8K
The test: does your prompt contain a because? Something happens, and something else follows from it. That's the capability you're paying for.
Sora simulates differently depending on what things are made of. Being specific changes the output substantially:
Sora 2 generates up to 10 seconds — long enough for a beginning, middle and end. Sequence words help it allocate the time:
At first the surface is perfectly still. Then a single drop lands, sending rings outward. As the ripples reach the edge, the reflection of the window breaks apart.
Liquids, destruction, weather, cloth, granular materials, and anything where the believability of the motion is the point. For talking characters, Veo's native lip-sync wins; for cheap volume, Kling wins; for pure visual polish on a still-ish shot, Seedance wins. Compare in Veo vs Sora and Sora vs Kling, and check the real cost in the cost calculator.
Ready-made prompts in this style: 12 Sora examples. Other models: Seedance guide · Veo guide · Kling guide.
Physical simulation — liquids, breaking, falling, cloth, weather and anything involving cause and effect. Prompts that describe a chain of consequences get results other models can't match; prompts describing a static scene waste that capability.
No dedicated field. Fold the intent into positive phrasing instead — "clean composition, no on-screen text" — the way you would with Veo. Kling, Runway and Wan still accept explicit negative lists.
Up to 10 seconds, which is long enough for a three-beat structure. Use sequence words — "at first", "then", "as" — so the model distributes the action across the full duration instead of front-loading it.