There was a moment watching a Sora-generated video where I genuinely thought: this changes everything. The motion was fluid, the scene looked almost cinematic, and for about fifteen seconds I forgot I was looking at something a machine built from a text prompt.
Then the next six tries reminded me exactly where we actually were.
That’s the Sora story in a nutshell. And now it’s over. OpenAI announced this week it’s shutting down the Sora app and dissolving its December deal with Disney along with it. A lot of people in the industry are framing this as a stumble. I think they’re reading it wrong. This is OpenAI learning a word that every company eventually has to learn: focus.
But before we get there, let’s be honest about what Sora actually was to use.
One out of seven videos came out usable. I’m not exaggerating — that was roughly the hit rate. The rest? Random audio injected from nowhere, like the platform decided your carefully crafted scene needed a soundtrack you never asked for. Text overlays that came out garbled, misspelled or completely unreadable. Audio cloning that couldn’t hold a consistent voice for more than a few seconds. Videos that cut off too soon because the platform couldn’t hold the full scope of what you asked for. Prompts that got quietly truncated mid-generation, so what came out had nothing to do with what you put in.
And then there was the copyright situation. Using other people’s images without clear licensing was a problem that never really got resolved. Worse, there were legitimate concerns that your own generated content could surface for other users. For a consumer hobbyist, that’s annoying. For an enterprise buyer — exactly the audience that was supposed to make Sora a business — that’s a dealbreaker.
So yes, Sora was frustrating. But here’s what’s also true: it was legitimately exciting. When it worked, it worked. It brought AI video generation out of the research lab and into the hands of regular people. It moved the Overton window on what was possible. That’s not nothing. That’s actually a meaningful contribution to where this industry goes next.
The problem was never the vision. The problem was the execution, the economics, and the distraction.
On the economics alone, I kept waiting for someone to do the math out loud. Video generation is among the most compute-intensive workloads in AI. The infrastructure costs are brutal. OpenAI was offering access at price points that made no sense given what it costs to actually run the thing. How were they not losing a fortune on every render? Maybe they were. An IPO has a way of forcing those conversations into the open, and suddenly every product that isn’t on a path to profitability becomes a very uncomfortable line item.
Which brings us back to focus.
OpenAI has been trying to be everything to everyone. Voice interfaces, search integration, the GPT store, agents, hardware ambitions, image generation, video generation. That’s a lot of surface area for any company, let alone one preparing to go public under intense scrutiny. Anthropic, meanwhile, made a different call early — stay in the lane of text and code, allocate compute to what actually scales into enterprise revenue, and let the shiny objects go. That bet is looking pretty smart right now.
Shutting down Sora is the first real sign that OpenAI is making the kind of hard resource allocation decisions that serious companies have to make. Not every moonshot earns a permanent seat at the table. Sometimes the right move is to acknowledge what isn’t working, cut it loose, and put those resources where they can actually win.
The AI video generation category isn’t going anywhere. Better tools already exist. More are coming. The players who get this right will be the ones who solve the quality consistency problem, nail the economics, and build real trust with enterprise buyers on IP and content rights. Sora didn’t get there. But it showed everyone else the map.
OpenAI will be fine. They just needed to learn when to say no. That’s a lesson most companies only learn the hard way.
