OpenAI's Sora was a family of generative video models and a consumer creation product. It progressed from a research preview in 2026 to Sora Turbo and then Sora 2, which combined video generation with synchronized audio. The current 2026 status is different from the original launch story: OpenAI discontinued the Sora web and app experiences on April 26, 2026, and says the Sora API will be discontinued on September 24, 2026.
Sora should therefore be described as a discontinued product and an important stage in OpenAI's video research—not as a tool that a new user can sign up for today.
What Sora was designed to do
Sora generated new video from natural-language instructions and, in some versions, from image or video inputs. Its purpose was broader than applying a filter to existing footage. The model attempted to synthesize scenes, subjects, movement, camera behavior and visual style over time.
The original research preview attracted attention because it could produce coherent short scenes with detailed environments and cinematic framing. OpenAI described the work as research toward models that could represent aspects of the physical world. That ambition did not mean Sora had a dependable physical simulator inside it. Generated footage could still contain impossible motion, changing identities, inconsistent geometry or cause-and-effect errors.
A short timeline
- February 2026: OpenAI presented the original Sora research preview and examples of text-to-video generation.
- December 9, 2026: Sora Turbo launched as a product through sora.com for eligible users. OpenAI described tools for creating, extending, remixing and blending video, including a storyboard interface.
- September 30, 2026: OpenAI introduced Sora 2, adding synchronized dialogue, sound effects and broader audio generation while improving control and physical consistency.
- April 26, 2026: OpenAI discontinued the Sora web and app experiences.
- September 24, 2026: OpenAI's current help notice lists this as the scheduled discontinuation date for the Sora API.
Historical product pages may still describe subscriptions, invitation access or creation limits that applied when those pages were published. Their presence online does not mean those offers remain available.
What changed with Sora 2
Sora 2 was presented as a video-and-audio generation model rather than video alone. OpenAI reported better prompt adherence, a wider range of styles, sharper realism and more plausible responses to physical events. Demonstrations included speech, ambient sound and effects synchronized with generated action.
The model could follow detailed creative direction about subject, setting, camera, lighting and sound. It was intended for ideation, previsualization, short-form storytelling and experimentation. Those strengths made it useful as a creative system, but not as a reliable recorder of facts. A realistic-looking output was still synthetic and could misrepresent people, places or events.
How the model worked at a high level
OpenAI's earlier Sora research described a diffusion-transformer approach. A diffusion system learns to turn a noisy representation into structured visual output, while a transformer architecture helps model relationships across space and time. Video and image data were represented in smaller units often described as patches, allowing the model to learn from material with different resolutions, durations and aspect ratios.
This is a useful conceptual explanation, not a complete technical specification. Training data, post-training methods, safety classifiers and production infrastructure all affect the final system. It is also wrong to assume that visual fluency equals an exact understanding of physics. Sora 2 improved consistency but the Sora 2 system card still framed deployment as an iterative process with emerging risks and limitations.
Safety, likeness and provenance
Powerful video generation creates risks that do not arise in ordinary editing software. A model can be used to fabricate events, imitate a person's appearance or voice, create harassment, or produce misleading political and financial material. OpenAI highlighted nonconsensual likeness and deceptive generations as significant Sora 2 risks.
At launch, OpenAI said Sora videos included visible and invisible provenance signals. Outputs carried a visible watermark, and Sora videos embedded C2PA metadata, an industry-standard mechanism for recording content origin. OpenAI also described internal image and audio search systems intended to help trace generated material.
Provenance controls are useful but not absolute proof. Metadata can be stripped by later processing, a watermark can be cropped, and viewers may encounter copies outside the original platform. The safer rule is to verify surprising footage through independent reporting and primary records rather than trusting realism alone.
OpenAI also used restrictions around photorealistic people, consent-based likeness features and stricter safeguards for content involving minors. According to its responsible-launch explanation, the company combined policy enforcement, moderation, red teaming and provenance signals. These measures reduced some risks but did not make every generated clip accurate, consensual or suitable for publication.
What discontinuation means in 2026
The official Sora discontinuation notice says the web and app experiences ended on April 26, 2026. OpenAI recommends that former users export their Sora content as soon as possible and provides a sunset page for that purpose. The notice says associated user data will be permanently deleted after any final export window has passed.
The same notice schedules the Sora API for discontinuation on September 24, 2026. Developers should treat that date as a migration deadline and verify the current notice directly, because service schedules and export windows can change. This article does not recommend building a new dependency on an API that is already scheduled for retirement.
OpenAI's public notice provides the discontinuation dates and data-export information but does not offer a detailed strategic explanation for ending the product. Claims about an unannounced replacement or the company's internal reasons would therefore be speculation.
How Sora differs from a how-to guide
A historical prompt tutorial can explain techniques that once helped users structure a shot, describe motion or iterate on a storyboard. This overview answers a different question: what Sora was, how its capabilities evolved, what risks accompanied it and whether it is currently available. Because the consumer product is discontinued, instructions that begin with creating a new Sora account are obsolete.
Why Sora still matters
Sora demonstrated how quickly generative video moved from silent short clips toward controllable audiovisual scenes. It also made provenance, consent and synthetic-media literacy central product questions rather than abstract policy topics. Those lessons remain relevant to any successor video system, regardless of brand.
For creators, the durable workflow is to treat generated media as source material that requires review, rights clearance and disclosure. For viewers, the durable habit is to check origin and corroboration. For organizations, procurement should address data retention, model retirement, export rights and continuity before a tool becomes embedded in production.
Bottom line
Sora began as OpenAI's ambitious text-to-video research, became a public creation product, and evolved into Sora 2 with synchronized audio and improved control. It also exposed difficult questions about deepfakes, likeness consent, provenance and the gap between visual realism and factual truth.
As of July 2026, the Sora web and app product is no longer available, and the API is scheduled to end in September 2026. The accurate current description is therefore retrospective: Sora was an influential OpenAI video-generation system whose capabilities and safety work helped define a stage of generative media, but it should not be presented as an active consumer service.