Why OpenAI Shut Down Sora: The Real Reasons

OpenAI shut down Sora because the product cost far more to run than it earned and was losing its audience. The Wall Street Journal reported the app cost about $1 million a day to operate, while active users fell from a peak of around one million to fewer than 500,000. The consumer app closed on April 26, 2026, and the Sora API was switched off on September 24, 2026 — which makes this week the real end of OpenAI's video business.

So the answer to "why did OpenAI shut down Sora" is not that the technology failed. Sora 2 produced some of the most convincing AI video anyone had seen. It failed as a business: every clip burned GPU time that OpenAI could sell at a much higher margin as code and enterprise tokens.

This article lays out the timeline, the money, what happened to the $1 billion Disney deal, what developers on the API need to do now, and what the shutdown says about which AI products are actually viable.

Key Takeaways

• OpenAI announced Sora's shutdown on March 24, 2026; the app closed April 26 and the API closed September 24, 2026.
• The Wall Street Journal reported operating costs of about $1 million per day, with users falling from roughly 1 million to under 500,000.
• Disney's $1 billion investment and licensing deal, announced December 11, 2025, never closed — no money changed hands.
• Downloads fell 32% in December 2025 and another 45% in January 2026, before the shutdown was announced.
• OpenAI says it is redirecting the work toward world-simulation research for robotics, and the compute toward coding and enterprise products.

When did OpenAI shut down Sora?

OpenAI shut down Sora in two stages. It announced the decision on March 24, 2026, closed the Sora web and mobile apps on April 26, 2026, and discontinued the Sora API on September 24, 2026. After the final deadline, user videos and images are permanently deleted.

The full arc, from first demo to last API call, is about two and a half years:

The gap between the two shutdown dates was deliberate. Consumer apps can be switched off quickly. API customers have Sora wired into production pipelines, so OpenAI gave developers five extra months to migrate.

According to The Decoder, OpenAI told users to download their content from the Sora library before the cutoff dates and said it had not decided whether to offer a final export window afterward. The shutdown removed sora.chatgpt.com, which handled both image and video generation.

If you still have assets you never exported, check your email for any notice from OpenAI about an export window. Beyond that, assume the content is gone.

Why did OpenAI shut down Sora?

OpenAI shut down Sora because its costs were high, its usage was falling and its compute was worth more elsewhere. The company did not give a single official reason, but reporting tied the decision to compute shortages, cost pressure and a shift toward core enterprise products such as coding tools.

There are three parts to that, and they reinforce each other.

1. The unit economics never worked

Video generation is among the most expensive things you can ask a model to do. A text response is a few hundred tokens. A video clip is many frames, each effectively an image, generated with temporal consistency across all of them.

The Wall Street Journal put Sora's operating cost at about $1 million per day, a figure repeated by 80 Level and others. Some widely shared analyses cite much larger numbers — up to $15 million per day in inference against roughly $2.1 million in total lifetime revenue. Those higher estimates come from independent commentators rather than OpenAI or a primary financial source, so treat them as unverified.

You do not need the larger number for the conclusion to hold. At the conservative $1 million a day, Sora cost around $365 million a year to run. A consumer app cannot cover that on the subscription revenue of a user base that was already shrinking.

2. People stopped coming back

Sora 2 had a spectacular launch. The app was installed 100,000 times on its first day. Then the curve bent the wrong way:

• Downloads fell 32% month over month in December 2025.
• They fell another 45% in January 2026, to roughly 1.2 million installs.
• Active users dropped from a peak of around 1 million to fewer than 500,000 by the time of the announcement.

That is the pattern of a novelty, not a habit. People tried it, made a few clips, shared them and left. Compare that with the trajectory we covered in ChatGPT reaching 1 billion weekly users: one product became daily infrastructure, the other was a toy people picked up once.

3. The compute had a better use

This is the reason that actually decided it. OpenAI is constrained by GPUs, not by ideas. Every GPU-hour spent rendering a clip for a free or low-tier user is a GPU-hour not spent serving Codex or enterprise API traffic, where customers pay per token and keep paying.

The Decoder reported that OpenAI is redirecting resources toward coding tools and enterprise customers, and toward a "super app" that rolls ChatGPT and other tools into one package. When your constraint is compute and your highest-margin product is sold out, the low-margin product goes. We looked at how tight that constraint is in the AI data center power crunch.

What happened to the Disney deal?

The Disney deal collapsed without ever being completed. Announced on December 11, 2025, it was a three-year agreement under which Disney would invest $1 billion in OpenAI and license more than 200 characters from Disney, Marvel, Pixar and Star Wars for use in Sora. The transaction never closed and no money changed hands.

Characters named in the deal included Mickey Mouse, Ariel, Iron Man, Darth Vader and Stitch. It was meant to be the template for how studios and AI labs would work together: licensed IP, inside a controlled generator, with revenue flowing back to the rights holder.

Instead, Disney reportedly learned about the shutdown about an hour before the public announcement, and has since been exploring partnerships with other AI companies, according to reporting summarized by Startup Fortune.

The deal had been controversial from the start. SAG-AFTRA, the Writers Guild of America and The Animation Guild sent an open letter describing it as a "theft of our work." And it came after a run of copyright and likeness fights:

• Sora 2 generated videos featuring copyrighted characters shortly after launch.
• In October 2025, Japan's Content Overseas Distribution Association, whose members include Studio Ghibli, demanded OpenAI stop using their content.
• Families of deceased celebrities, including Robin Williams and George Carlin, objected to generated videos of their relatives. OpenAI restricted videos of Martin Luther King Jr.

The legal and reputational exposure is the part of the ledger that does not appear in a compute bill. A product that loses money and generates lawsuits is an easy cut. OpenAI is already carrying significant legal load elsewhere, as we covered in the Apple vs OpenAI trade secrets lawsuit.

What should developers using the Sora API do now?

If your product called the Sora API, those requests now fail, and you need to move to another video model. The main alternatives are Google Veo 3.1, Runway Gen-4.5, Kling 3.0 and Seedance 2.0. Which one fits depends on what you used Sora for, because no single replacement matches it on every dimension.

Based on current comparisons of the field:

The migration lesson is more durable than any particular vendor choice. Put video generation behind your own interface so the provider is a configuration value, not a dependency scattered through your code:

import os
from dataclasses import dataclass
from typing import Protocol

@dataclass
class VideoRequest:
    prompt: str
    seconds: int = 8
    aspect_ratio: str = "16:9"

class VideoProvider(Protocol):
    def generate(self, request: VideoRequest) -> str:
        """Start a job and return a provider job ID."""
        ...

PROVIDERS: dict[str, VideoProvider] = {}

def register(name: str, provider: VideoProvider) -> None:
    PROVIDERS[name] = provider

def generate(request: VideoRequest) -> str:
    # Switching vendors is now an environment change, not a refactor.
    name = os.environ.get("VIDEO_PROVIDER", "veo")
    if name not in PROVIDERS:
        raise RuntimeError(f"No video provider registered as '{name}'")
    return PROVIDERS[name].generate(request)

Each vendor gets a small adapter class that implements generate. When the next provider is discontinued or reprices, you change one environment variable and one adapter.

This is the unique risk in generative media that text does not share as sharply. Text models are close to interchangeable for most tasks. Video models differ in clip length, resolution, audio support and how prompts are interpreted, so a swap changes your output, not just your bill. Teams that hard-coded Sora's parameters had five months to rewrite. Teams with an adapter layer changed a setting.

Is OpenAI done with video?

No. OpenAI has said Sora continues as a research effort rather than a product. A spokesperson said the team will "focus on world simulation research to advance robotics that will help people solve real-world, physical tasks." The consumer video generator is gone; the underlying research into models that predict how the physical world behaves is not.

That reframing is consistent with where the rest of the industry is heading. A model that can generate a plausible video of a cup being pushed off a table has learned something about physics. That same capability is what a robot needs to plan a movement before making it.

The bet is that "world models" are worth more as the brain of a machine than as a clip generator for social feeds. It lines up with the work we covered in Gemini Robotics 2, where a general model is adapted to control a body.

Whether that pays off is an open question. But it explains why OpenAI kept the researchers and dropped the app.

Frequently asked questions

Why did OpenAI shut down Sora? OpenAI shut down Sora because it was expensive to run and losing users, and the compute was more valuable for coding and enterprise products. The Wall Street Journal reported operating costs of about $1 million a day, while active users fell from around 1 million to under 500,000.

When did the Sora API shut down? The Sora API was discontinued on September 24, 2026. The Sora web and mobile apps had already closed on April 26, 2026, following OpenAI's announcement on March 24, 2026.

How much money did Sora lose? OpenAI has not published a figure. The Wall Street Journal reported roughly $1 million per day in operating costs, which is about $365 million a year. Larger estimates of up to $15 million a day circulate online but do not come from OpenAI or a primary financial source.

What happened to the Disney and OpenAI deal? The deal never closed. Disney announced a $1 billion investment and a three-year license covering more than 200 characters on December 11, 2025, but no money changed hands before OpenAI discontinued Sora.

Can I still download my Sora videos? Probably not. OpenAI told users to export their content before the April 26 and September 24, 2026 deadlines and said data would be permanently deleted afterward. It had not committed to a final export window.

What is the best alternative to Sora? There is no single best replacement. Google Veo 3.1 is the closest for cinematic realism, Runway Gen-4.5 for camera control and editing, Kling 3.0 for high-motion scenes and Seedance 2.0 for combined audio and multi-shot video.

The verdict

Sora did not die because AI video is bad. It died because OpenAI had a fixed supply of compute, and a video app that cost around $1 million a day while shedding half its users was the worst use of it. The Disney deal, the copyright fights and the likeness complaints made the decision easier, but the GPU math made it inevitable.

The takeaway for anyone building on AI APIs is blunt: a product being impressive tells you nothing about whether its vendor will keep it alive. Margin does. Build adapter layers around anything generative, and assume the most expensive-to-serve feature on a vendor's menu is the first to go.

If you want to see where that compute went instead, read our LLM API pricing guide — the same month Sora's API went dark, OpenAI cut its text model prices in half.

The most impressive demo of 2024 lasted two and a half years. The boring, profitable tokens outlived it.