Kling 4 vs Sora 2: comparing with a model that has shut down

What happened to Sora 2

Sora 2 launched on September 30, 2025 with an invite-only app built around short social videos. On March 24, 2026, OpenAI announced it would discontinue Sora in both the consumer app and the API.

According to OpenAI's help center, the web and app experiences ended on April 26, 2026, and the API, including the Sora 2 and Sora 2 Pro models, was scheduled to shut down on September 24, 2026. Data from Sora use is deleted once the export period ends.

So any Sora 2 workflow you had has already stopped, and videos or character assets you did not export may no longer be recoverable.

Two different ideas of consistency

Sora 2's answer to consistent people was the cameo. According to launch coverage, you recorded a short video and audio sample of yourself in the app, OpenAI checked it to confirm it was you, and it became a reusable likeness. You chose who could use your cameo, could revoke access at any time, and could delete videos that used it. Sora 2 also generated synchronized speech and sound.

Kling's answer is the reference set. You upload images, and in some modes short clips, of a person, character or product, and the model keeps that subject in a new scene. There is no built-in likeness ownership system: you are responsible for having the right to use the face you upload.

The cameo approach was stronger on consent and identity for real people. The reference approach is more flexible: it works the same way for products, mascots, illustrated characters and people, and it does not require a live recording.

Where Kling 4 stands

Kling 4.0 officially launched in Kling's app on October 15, 2026, after early access from September 28, according to Kuaishou. fal.ai's Kling 4.0 guide lists clips of 3 to 30 seconds, 720p to 4K, stereo audio, up to 10 keyframes and up to 15 references. Its developer API is marked as coming soon, and we found no published API price.

If you need an API now, Kling 3.0 Omni is available through providers such as fal.ai, with clips of up to 15 seconds, optional native audio and reference elements, at $0.084 to $0.14 per second depending on tier and audio.

Moving a Sora workflow to reference-based models

1. Recover what you can. If you exported Sora videos, pick frames that show your character clearly: front, three-quarter and profile, in even light. These become your reference images. RefVid takes up to five per character, each at least 300 pixels.

2. Rebuild consent for real people. Without Sora's cameo permissions, get written permission from anyone whose face you use, and keep it on file. Some models restrict real faces in references: users report that Seedance 2.0 and 2.5 reject them.

3. Separate character from scene. Sora prompts often described the character in words. With references, the images carry the character and the prompt describes only the setting, the action and the camera.

4. Expect a different look. Every model has its own motion style and color. Re-generate a few key scenes and compare before converting a whole series.

Which to pick for consistent characters and products

Sora 2 is not an option. For consistent real or photoreal people, start with Kling: Kling 3.0 Omni through the API today, Kling 4.0 when its API opens.

For short single-subject shots, Veo 3.1 takes up to three reference images and makes 8-second reference clips, according to Google's documentation. For products, mascots and stylized characters in long takes, Seedance 2.5 offers 4- to 30-second clips at up to 720p and up to 30 reference images on fal.ai.

RefVid saves your character or product once and runs it on Kling 3.0 Omni today, switching to Kling 4.0 when its API opens. Your saved characters stay in RefVid when the model underneath changes.