Update, October 2026. OpenAI has shut down Sora: the app stopped working on 26 April 2026 and the API was switched off on 24 September 2026. You can no longer generate video in Sora 2 Pro, and KADR no longer has the model either. What happened and what to use instead: Sora was shut down.
What Sora 2 Pro was
Sora was OpenAI’s video model, from the people who made ChatGPT. Version 2 Pro was the flagship and was positioned accordingly: maximum quality, maximum price.
Three things Sora was valued for:
Physics of motion. Objects in frame behaved realistically: water poured like water, fabric swayed like fabric, things fell the right way. This is where simpler models stumble.
Coherence in complex scenes. When the frame held many interacting objects, weaker models got confused and everything fell apart. Sora kept a complex scene whole.
Native sound. Like Veo, it generated video together with sound.
The key fact: it was one of the most expensive models
While the model was running, Sora 2 Pro cost 462 tokens for a 6-second clip at 1080p, about 730 rubles on KADR. For comparison, Veo 3.1 Lite costs 53 tokens for the same clip, so Sora was almost nine times more expensive.
A minute of finished video from such segments came to thousands of rubles, and with regenerations the bill went into tens of thousands. What a minute costs on the models available today is in the article on how much a minute of AI video costs.
What it was worth paying for
Sora justified its price where its strengths were really needed:
- A complex scene with physics: water, fire, destruction, fabric, liquids, many moving objects.
- A scene where a mistake is costly: the final shot of an important project, where you need the maximum on the first try.
- Realistic interaction between objects that fell apart on simpler models.
These tasks have not gone anywhere. Today the closest replacement for them is Veo 3.1: detail, lighting, believable physics and native sound.
When you did not need Sora
The same logic applies to any flagship today:
- Drafts and tests. Never tune a prompt on the most expensive model.
- Simple static scenes. A landscape, an object on a background, a calm portrait: models several times cheaper will cope.
- Backgrounds and bulk content. If you need a lot, fast, a flagship will ruin you.
How to work economically with any flagship
One rule from the Sora days still holds: the flagship only for the final, and only when it is worth it.
- Debug the prompt and composition on a cheap model: Veo 3.1 Lite or Seedance at low resolution.
- Make sure the scene works: camera, motion, logic.
- Move the finished prompt to the flagship (Veo 3.1, Seedance 2.5 or Kling v3) and render the final take.
That way you pay the flagship price once for a result, not twenty times for tests.
Sora versus Veo and Kling, and what replaces it
Without tests, based on the nature of the models:
- Sora 2 Pro was the strongest at physics and coherence in complex scenes, but also one of the most expensive. It is no longer available.
- Veo 3.1 is about detail, lighting, physics and sound, 1080p only. The main replacement for Sora’s photorealistic shots.
- Seedance 2.5 is about long story clips: up to 30 seconds in one pass and a stable character.
- Kling v3 is about camera movement and people, cheaper, and works with real faces.
Roughly: if the scene needs physics and photorealism with sound, Veo. If you need a long story clip, Seedance 2.5. If camera movement and people, Kling. If budget, Seedance 1.5 Pro or Veo Lite.
In short
Sora 2 Pro was OpenAI’s flagship: the strongest physics, coherence in complex scenes, native sound and one of the highest prices in the catalogue, about 730 rubles for 6 seconds while it was running. In 2026 OpenAI shut Sora down, and the model is no longer available anywhere. Its tasks are now covered by Veo 3.1 (physics and photorealism with sound), Seedance 2.5 (long story clips) and Kling v3 (people and camera).
