OpenAI’s Sora in 2026: From Sora 2 Hype to Shutdown

Posted on by IE Contributor

Few AI products have had a run as short and as dramatic as Sora. In the space of about 18 months, OpenAI’s text-to-video model went from a jaw-dropping research preview to a viral social app, a billion-dollar Hollywood partnership, and then a full shutdown. As of late October 2026, there is no Sora you can use: the consumer app closed in April and the developer API followed in September. For anyone tracking where generative AI is heading, the story of Sora is less about one model and more about what it takes to turn a dazzling demo into a sustainable product.

AI-generated still from OpenAI's Sora Tokyo Walk demo
An AI-generated still from OpenAI’s “Tokyo Walk” Sora demo. AI-generated image, public domain, via Wikimedia Commons.

What Sora was, and why Sora 2 mattered

Sora is OpenAI’s family of models that turn written prompts (and, later, images) into video. OpenAI has described the original Sora, first shown in February 2024, as the “GPT-1 moment for video”: the first time video generation started to look like it genuinely worked, with behaviours such as object permanence emerging from scale, according to the company’s Sora 2 announcement. The first public version reached users at the end of 2024, as The Verge notes.

Sora 2, released on September 30, 2025, was pitched as a much bigger leap. OpenAI called it potentially “the GPT-3.5 moment for video” and highlighted three advances:

  • Better physics. Earlier models tended to “cheat” to satisfy a prompt, for example teleporting a missed basketball into the hoop. In Sora 2, OpenAI said, a missed shot rebounds off the backboard, because the model is better at representing failure as well as success.
  • More control. The model could follow multi-shot instructions while keeping the scene consistent, across realistic, cinematic and anime styles.
  • Synchronised audio. Sora 2 generated dialogue, sound effects and background soundscapes alongside the video, rather than producing silent clips.

OpenAI also offered a higher-quality Sora 2 Pro model to ChatGPT Pro subscribers and said it planned to bring Sora 2 to its API. The company was candid that the model was “far from perfect,” but framed it as evidence that scaling neural networks on video data moves AI closer to simulating the physical world (OpenAI).

The Sora app: a social network built on “characters”

The more surprising part of the Sora 2 launch was the product wrapped around it. Instead of a simple generation tool, OpenAI released an invite-only iOS app, simply called Sora, with a TikTok-style feed. Users could create clips, remix each other’s videos, and drop themselves or friends into scenes using a feature called “characters,” which captured a person’s likeness and voice after a short one-time recording that also verified their identity (OpenAI). The rollout started in the U.S. and Canada.

Because realistic video of real people raises obvious risks, OpenAI published a dedicated safety framework alongside the launch. Its “Launching Sora responsibly” post listed several protections:

  • A visible watermark on every video at launch, plus embedded C2PA provenance metadata and internal tools to trace clips back to Sora.
  • Consent controls on characters: only the owner decides who can use their likeness, can revoke access at any time, and can see every video that includes it, including other people’s drafts.
  • Blocks on depicting public figures (unless they used the characters feature themselves) and on music imitating living artists or existing works.
  • Stronger defaults for teens, including limits on mature content, scroll limits, and parental controls through ChatGPT.

Early momentum was real. According to Axios, the app rose to the top of Apple’s App Store charts and reached a million downloads faster than ChatGPT did. In December 2025, Disney announced a landmark deal to license more than 200 characters for AI video and to invest $1 billion in OpenAI.

The shutdown: what happened in 2026

The momentum did not last. Axios, citing TechCrunch, reported that Sora app downloads had fallen 45% by January 2026. On March 24, 2026, OpenAI announced it was “saying goodbye to Sora.” In a message to users, the team said the decision followed “careful internal discussion about our broader research priorities” (The Verge).

The official timeline, per OpenAI’s Help Center and API deprecations page, ran as follows:

  • March 24, 2026: shutdown announced; developers using the Videos API were notified the same day.
  • April 26, 2026: the Sora web and app experiences were discontinued.
  • September 24, 2026: the Videos API and every sora-2 and sora-2-pro model and snapshot were removed. OpenAI lists no replacement, and its Sora 2 model page now states plainly that “there is no one-to-one replacement API.”

Why pull the plug on a product that had been a headline success months earlier? The consistent answer is compute and focus. Axios reported that Sora was consuming significant computing power at a time when every frontier lab is short on it, and that OpenAI was prioritising capital, chips and enterprise products amid competition from Anthropic and Google. The Verge, citing Reuters, reported that Sora’s future had been debated internally for some time because it left other teams with less compute, and that the Disney partnership ended with no money having changed hands. The Verge also pointed to OpenAI’s broader push to concentrate on practical products such as Codex.

The research is not entirely gone. An OpenAI spokesperson told Axios that the Sora team will continue to focus on “world simulation research to advance robotics that will help people solve real-world, physical tasks.”

What it means for the AI video market

Sora’s exit carries a few lessons that reach well beyond OpenAI:

  • Video is expensive. Generating high-quality, audio-synced video is among the most compute-hungry things AI does. Even the best-funded lab decided the return did not justify the cost against coding and enterprise tools.
  • Virality is not retention. A record-setting launch did not translate into lasting daily use. A novelty feed of AI clips proved easier to start than to sustain.
  • Platform risk is real. Developers who built pipelines on the Sora 2 API got six months’ notice and no migration path. The OpenAI deprecations page itself warns that preview and specialised models can be retired faster than general-purpose ones.
  • World models remain the long game. OpenAI has consistently framed video generation as a route to AI that understands the physical world. That ambition now lives in research and robotics rather than in a consumer app.

Practical takeaways

  • Former Sora users: OpenAI says you can still request an export at sora.chatgpt.com/sunset, and that it will permanently delete Sora data once any final export window ends. Do it now if you have not. Purchased ChatGPT/Sora credits can be used for Codex instead (OpenAI Help Center).
  • Developers: any code calling v1/videos or Sora 2 models no longer works. Audit for dead integrations and design new pipelines so the video model can be swapped without a rewrite.
  • Businesses: treat generative video tools as interchangeable suppliers, not foundations. Keep your source assets, prompts and brand guidelines outside any single vendor’s platform.
  • Everyone: provenance matters more than ever. Sora’s watermarking and C2PA metadata set a useful baseline; look for similar labelling in whichever tools you use next.

Conclusion

Sora 2 was, by most accounts, a technically impressive model, and its app briefly showed what a social network built on AI video might look like. But in 2026 OpenAI concluded that the compute it consumed was better spent elsewhere, and the product disappeared within about seven months of launch. The episode is a useful reminder that in generative AI, capability is only half the equation. Cost, sustained demand and strategic focus decide which breakthroughs become lasting products, and which become case studies.

Sources