Sora generated video by using an input to guide a model that progressively transformed a noisy video representation into a clip. OpenAI’s 2024 materials describe compressed video latents, spacetime patches, a transformer, and diffusion. They explain the broad approach, but not every implementation detail. (OpenAI’s technical report; Sora System Card)
For production planning, the key update is that Sora is no longer available: OpenAI says the Sora web and app experiences ended on April 26, 2026, and the Sora API was discontinued on September 24, 2026. As of November 9, 2026, Sora is not a viable API to select for a new production integration. (OpenAI Help Center)
How Sora turned an input into video
OpenAI’s technical report and 2024 System Card describe the original Sora. The distinction below matters: compressing and patching videos describes the model’s visual representation; it does not mean every text-only request first needs an existing video.
Representation: OpenAI’s report says videos are compressed into a lower-dimensional latent space and then divided into spacetime patches for the model to process. It says Sora was trained on and generated videos within this compressed space. A decoder maps generated latent representations back to pixels. These are descriptions of the model’s representation, not a required sequence of user actions. (OpenAI’s technical report)
Text-to-video generation, at a high level:
- The prompt guides generation. OpenAI’s System Card says the original Sora could generate a video from text alone. It also describes image and video inputs, but the reviewed materials do not specify their exact conditioning pipeline. (Sora System Card)
- Generation starts from noise. OpenAI describes Sora as starting with a base video that looks like static noise, then progressively removing noise over multiple steps, guided by the input. (Sora System Card)
- The representation becomes a video. The technical report says Sora generates within compressed latent space and describes a decoder that maps generated latents back into pixel space. (OpenAI’s technical report)
- Frames are considered together. OpenAI says Sora had “foresight of many frames at a time,” which helped keep a subject consistent when it temporarily went out of view. That is a design benefit, not a guarantee that every detail remains consistent. (Sora System Card)
OpenAI explicitly says its technical report omits model and implementation details. So this is a high-level explanation, not a complete account of every internal step or of how image- and video-conditioned generation works. (OpenAI’s technical report)
What Sora could do—and what its limitations meant
At the December 2024 launch, OpenAI described Sora Turbo as supporting text prompts and supplied assets, with tools to extend, remix, and blend material. The launch announcement listed outputs up to 1080p and 20 seconds. These are historical product details, not current service limits: Sora has since been discontinued. (OpenAI’s Sora launch announcement)
OpenAI also warned at launch that Sora often generated unrealistic physics and struggled with complex actions over long durations. For example, an action in a generated clip might not follow the physical outcome a viewer expects. That warning was qualitative, not an error rate or reliability benchmark. (OpenAI’s Sora launch announcement)
Can developers use Sora through a production API?
No—not as of November 9, 2026. OpenAI’s discontinuation notice says the Sora API was discontinued on September 24, 2026. OpenAI’s developer deprecations page also identifies Sora 2 video-generation models and the Videos API as removed on that date. (OpenAI Help Center; OpenAI API deprecations)
A historical API listing or tutorial does not change that status. For a new video-generation system, choose a currently supported API and verify its model availability, input and output formats, limits, access terms, and pricing in that provider’s current official documentation. Do not assume a ChatGPT plan provides API access.
Bottom line
The original Sora’s documented approach used compressed video representations and spacetime patches; for text-to-video, OpenAI described progressively removing noise from a base video representation and decoding generated latents into pixels. The documentation does not disclose every implementation detail, and it does not establish the exact conditioning pipeline for image or video inputs. For production selection, the decisive point is that Sora’s app and API have both been discontinued, so Sora is not an available production API choice as of November 9, 2026.
Sources
Video generation models as world simulators — OpenAI
What to know about the Sora discontinuation — OpenAI Help Center