A version of this story appeared in CNN Business’ Nightcap newsletter. To get it in your inbox, sign up for free here.

Six months and millions of dollars down the drain, OpenAI is pulling the plug on what it once called “the most powerful imagination engine ever built.”

There seems to have been two key missteps in the Sora saga:

ICYMI: OpenAI said Tuesday that it would wind down Sora, the TikTok-like text-to-video app that, to OpenAI’s credit, did make some super real-looking stuff. Its release was a turning point for AI-generated video, elevating it from the realm of goofy slop to sophisticated deepfakes (for better or worse).

The Sora rollback is part of a broader shift within OpenAI, once the undisputed frontrunner in the AI race that now faces serious competition from rivals such as Anthropic and Google.

Earlier this month, OpenAI’s head of applications told staff that the company couldn’t afford to be “distracted by side quests,” the Wall Street Journal reported. The company is doubling down on its core products, including an updated ChatGPT focused on office work and a coding tool called Codex.

Sora downloads soared after the invitation-only rollout in September, with more than 1 million daily active users after just over a month, according to data from Similarweb. But the novelty wore off fast. Usage peaked in early November, and then tumbled. Downloads are down 70% from November, and daily active users have fallen 34%.

That was stumble No. 1: OpenAI built a really slick machine and expected everyone to love using it as much as its engineers did. But AI videos, even the really sophisticated kind, take much of the fun out of scrolling.

The joy humans derive from, say, a video of a husky that sounds like it has an Italian accent, or a cat moving to the beat of Nelly’s 2005 banger “Grillz,” comes partly from knowing that someone else experienced a funny thing in real life and managed to capture it as it happened. When an AI version of the same thing creeps into your social feed, it feels like cheating. The person who posts that video didn’t witness anything impressive or unusual, they just typed some words into a box and uploaded the result.

But beyond philosophical questions about authenticity,it took almost no time for the internet to get around Sora’s content restrictions. People used the app to generate fake videos of women being strangled or splattered with mysterious white goo, people committing crimes, andpublic figures wearing Nazi-like uniforms. Less than a month after Sora’s launch, OpenAI had to put a pause on videos of some historical figures after users created “disrespectful depictions” of Martin Luther King, Jr. (The app banned the use of living public figures’ likenesses, but allowed depictions of people who’d died.)

We’ve seen this kind of stumble with consumersfrom OpenAI before.Its launch of ChatGPT-5 in August was a disaster, as the new model had a flatter, more terse personality, as well as an alarming inability to answer basic questions. Users recoiled immediately, forcing OpenAI to backtrack and restore the old models.

Sora stumble No. 2 was much more prosaic, with the cold, pragmatic glare of a balance sheet: OpenAI seemingly underestimated how much it would cost to run the thing. The first sign that the cash burn on Sora was becoming a problem came in late October, when the head of Sora, Bill Peeples, posted on X that “the economics” were “currently completely unsustainable.”

In November, Forbes estimated the app cost OpenAI — which is still burning through cash faster than it can bring it in — about $15 million a day.

The company didn’t respond to questions from CNN about that estimate or how much Sora’s financial burden played into the decision to end the app. In a statement, OpenAI said the Sora team would continue to focus on “world simulation research” to advance OpenAI’s robotics efforts.

Bottom line: OpenAI has a math problem: It reportedly made about $13 billion in revenue last year, and it aims to triple that in 2026 while burning through tens of billions on computing power. That conundrum is forcing the company to relent on revenue drivers it once avoided — like showing ads in ChatGPT results, which CEO Sam Altman once derided as a “last resort” — and throw in the towel onlosing bets like Sora.