Sora 2: Video, Audio, Social App & Shutdown After Six Months

From silent clips to synchronized scenes

OpenAI's first Sora model, previewed in February 2024, generated a minute of coherent video with no dialogue, sound effects, or soundtrack; it reached ChatGPT Plus and Pro subscribers that December, still silent. Sora 2 generates synchronized audio in the same pass as the video: speech, ambient sound, and sound effects timed to what is on screen. OpenAI also said the model handles multi-shot sequences that keep the same characters and setting consistent across cuts, and cited specific physical behaviors, momentum and buoyancy among them, that earlier video models got wrong and Sora 2 gets right. [1]

Atlas interpretation: OpenAI framed these gains as more than a creative-tool update. Its own materials describe Sora 2 as a step toward general-purpose "world simulators," a research direction the company connects to eventual robotics work rather than to video generation alone. That is a claim about where the technology is heading, made by the company building it. Nothing in the September release let outside researchers test simulation accuracy on a task a robot would actually need to perform, and the six months that followed tested the product mainly on a much narrower question: what happens when millions of people can generate a realistic video of almost anything. [1]

Cameo: capture once, consent required

Cameo was the app's headline consumer feature: a short capture of a person's face and voice that Sora could then insert into generated scenes with, OpenAI said, notable fidelity, a capability it described as generalizing to any human, animal, or object. Control stayed with the person captured. They could keep their likeness private to their own generations, share it so others could cast them, revoke access at any time, and request removal of videos that used it. [1]

The app launched invite-only on iOS in the United States and Canada. It still moved fast: about 56,000 downloads on day one (a figure Appfigures later revised upward, to roughly 110,000, in November 2025), 627,000 iOS downloads in its first week, ahead of the 606,000 iOS downloads ChatGPT logged in its own first week, and past one million downloads in under five days. Sora reached number one on the U.S. App Store on October 3, three days after launch. [3]

Atlas interpretation: That speed is the context for everything that follows. Cameo gave living users a specific, revocable consent mechanism for their own likeness before launch. Nothing equivalent existed at launch for the likenesses of dead public figures, and the copyright policy governing fictional characters ran the opposite direction from Cameo's default: characters could appear unless a rightsholder opted out, rather than only if someone opted in. Both gaps produced public disputes within about two and a half weeks of release. [1][3]

MLK, and the guardrail that came after

Within weeks, Sora users had generated videos showing Martin Luther King Jr. making monkey noises during his "I Have a Dream" speech and wrestling Malcolm X, alongside recreations of other deceased figures including Malcolm X, Michael Jackson, Elvis Presley, and Robin Williams. Bernice King, Dr. King's daughter, posted publicly asking people to stop sending her AI-generated videos of her father. [4]

On October 16, OpenAI and the King estate issued a joint statement pausing Sora's generation of Dr. King's likeness while OpenAI "strengthens guardrails for historical figures." OpenAI said that while there are free-speech interests in depicting historical figures, "public figures and their families should ultimately have control over how their likeness is used," and it extended a version of its rightsholder opt-out mechanism so an estate or authorized representative could request exclusion for a deceased person, the same shape as the system it already ran for copyrighted characters. [4][5]

Atlas interpretation: The order of operations is the point worth holding onto. OpenAI built a specific, opt-in consent flow for living people who chose to capture their own Cameo. It did not build an equivalent flow for dead public figures before launch; that arrived after offensive content had already circulated and after direct pressure from a specific family. A guardrail built in response to a documented harm is not the same as one designed ahead of it, even when the eventual mechanism looks similar. [5]

Hollywood's opt-out problem

The same opt-out logic applied at launch to copyrighted characters: a studio's characters could appear in Sora generations unless the rightsholder explicitly asked to be excluded. The Motion Picture Association said that since release, "videos that infringe our members' films, shows, and characters have proliferated on OpenAI's service and across social media." Actor Bryan Cranston flagged the issue to SAG-AFTRA, whose national executive director, Duncan Crabtree-Ireland, called it "a moment of real concern and danger for everyone in the entertainment industry." [7]

On October 3, three days after launch, Sam Altman published "Sora update #1," reversing the policy: OpenAI would give rightsholders "more granular control over generation of characters, similar to the opt-in model for likeness but with additional controls," and would explore revenue sharing with rightsholders once it had worked out a structure. Altman acknowledged the company had been caught off guard by scale, writing that "people are generating much more than we expected per user." [6]

Atlas interpretation: OpenAI reversed a launch-day policy within three days, under pressure from a trade association and a performers' union rather than as a planned rollout step. That is the same pattern the MLK dispute would repeat two weeks later: a permissive default, a documented harm that reached the press, then a narrower policy built in response. Read together, the two episodes suggest the September 30 launch shipped with less policy work done on other people's likenesses and characters than on the ones users volunteered themselves through Cameo. [6][7]

What the realism made possible

NewsGuard tested Sora 2 against 20 false claims already circulating online between September 24 and October 10, 2025, and reported that the model produced a realistic video advancing the false claim in 16 cases, an 80 percent rate. The fabricated videos included a Moldovan election official destroying pro-Russian ballots and a toddler being detained by U.S. immigration officers; five of the 20 claims tested traced back to Russian disinformation operations. NewsGuard also reported it stripped Sora's visible watermark from a downloaded video using a free online tool in about four minutes. [8]

Atlas interpretation: This is the pattern connecting the technical upgrade to the backlash. Synchronized audio and steadier physics are what made Sora 2 popular enough to pass a million downloads in under five days; they are also what made a fabricated video harder to spot at a glance than the original Sora's silent, often visibly uncanny clips. A watermark that can be removed in minutes, and that OpenAI only embedded as durable metadata on downloads in the first place, was not a working answer for video that had already left the app by the time anyone checked its provenance. [8][1]

Atlas interpretation: The product did not last long enough to settle any of this on its own terms. On March 24, 2026, OpenAI announced it was shutting Sora down; reporting at the time put its compute cost near $1 million a day against downloads that had fallen from a November 2025 peak above 3.3 million to about 1.1 million by February 2026, with in-app purchases totaling only about $2.1 million. The shutdown took a downstream product of the launch's own backlash with it: The Walt Disney Company's three-year licensing agreement, signed under the opt-in framework OpenAI adopted after this launch and carrying a $1 billion OpenAI investment in Disney alongside it, reportedly reached Disney less than an hour before the shutdown was made public. [9]