Sora 2 is gone: what OpenAI's video model left behind

Key Takeaways

• OpenAI's Sora 2 API shut down on September 24, 2026, six months after the company announced the end of Sora and five months after the consumer app closed on April 26.

• Sora 2 launched on September 30, 2025, with sound generated alongside the picture and an invite-only iOS app built around a vertical feed and "cameos" of real people.

• Its biggest fights were over rights: OpenAI moved copyrighted characters from opt-out to opt-in within days of launch, and paused Martin Luther King Jr. videos after his estate complained.

• Disney agreed to a $1 billion investment and character license in December 2025, then walked away when OpenAI shut Sora down. OpenAI gave no official reason and has named no successor model.

Sora 2's API went dark today

As of today, September 24, 2026, OpenAI's Videos API no longer serves sora-2 or sora-2-pro. It's the last piece of Sora to switch off. OpenAI announced the discontinuation on March 24, closed the Sora app and web experience on April 26, and gave developers until now. The deprecations page lists no replacement.

So Sora 2 ran for just under a year. In that time it went from one of the most-watched launches in AI video to the example people reach for when they argue about likeness rights and what these models cost to run. Most of its legacy lives in those arguments.

What the model got right

The model was a real step up. The original Sora, previewed in February 2024, made silent clips that looked great in a showreel and fell apart the moment anything had to obey physics. Sora 2 generated dialogue, sound effects and ambience in the same pass as the picture, and objects kept their weight better: OpenAI's own launch example was a missed basketball shot rebounding off the backboard instead of warping into the net. That's the right example, since the old failure was the model bending reality to satisfy the prompt.

The API arrived at OpenAI's DevDay on October 6, 2025, billed per second of output: $0.10/second for sora-2, and $0.30 to $0.50/second for sora-2-pro depending on resolution. A 10-second Pro clip in HD came to about $5. That pricing got Sora 2 wired into third-party video apps and model aggregators, which is why the name kept appearing in tool menus long after the app itself was gone.

The feed was the actual product

Sora 2 shipped as a social app: an invite-only iOS release for the US and Canada, with Android following in November, built around a vertical feed of remixable clips. The hook was cameos. You'd record a short video and voice sample of yourself, pick who could use your likeness, and friends could drop you into their generations.

It worked, briefly. The app hit the top of the US App Store in its first week. Then the curve bent the other way: Appfigures estimated US downloads fell 32% month over month in December 2025 and another 45% in January 2026. For working video people the feed was mostly beside the point. Every export carried a moving watermark and clips were short, so Sora 2 footage turned up in pitch decks and mood reels far more often than in delivered work.

The rights fights

This is the part most people will remember. At launch, copyrighted characters were in by default and rightsholders had to opt out. The feed filled with other people's IP almost immediately, and three days later OpenAI said it would switch to opt-in. That fall, Japan's Content Overseas Distribution Association, whose members include Studio Ghibli, asked OpenAI to stop training Sora 2 on its members' work without permission.

Likeness went the same way. In mid-October OpenAI paused videos of Martin Luther King Jr. after his estate objected to disrespectful depictions, and Zelda Williams publicly asked people to stop sending her AI videos of her father. Tools that stripped the watermark were circulating within about a week of launch, which didn't help the provenance story OpenAI had leaned on.

Why Sora ended

In December 2025 Disney agreed to invest $1 billion in OpenAI and license more than 200 characters from Disney, Pixar, Marvel and Star Wars for use in Sora, the first time Disney had licensed its IP to an AI platform. SAG-AFTRA, the Writers Guild and the Animation Guild all condemned it. It didn't last much past three months. According to The Hollywood Reporter, Disney learned Sora was being shut down shortly after a working meeting between the two teams, and it dropped the deal.

OpenAI never gave an official reason. Reporting pointed to compute costs, an unproven way to make money from a free video feed, and falling usage. The company has said it isn't leaving video entirely, but it hasn't announced anything to replace Sora.

What Sora 2 leaves behind

The clearest trace is policy. The opt-in rule for copyrighted characters was worked out in public, on a live feed, after launch, and so were the limits on depicting dead public figures. And the biggest studio license in AI video ended because of one product decision on OpenAI's side.

The other trace is stale information. Sora 2 still sits in plenty of "best AI video generator" roundups and in the model menus of tools that resold API access, so a fair number of pages are describing a product nobody can use as of today. For a model that ran less than twelve months, it's going to be listed as current for a while yet.