Sora has ultimately met its end, being taken offline.
Even when it first launched, I had a premonition that it couldn't succeed. After all, what reason does a community consisting only of creators have to survive? Moreover, too many similar things have happened in the past two years: a highly capable AI product emerges, everyone marvels for a while, and then it gradually fades from discussion. Sora merely repeated this "classic path," even though it once topped the AppStore charts.
Many in the industry are discussing why OpenAI would abruptly cut multiple product lines, including Sora. Some say it's for going public, others cite cost reasons. Of course, OpenAI's official explanation is to focus on researching world models.
But is that really the case?
We all know that since Sam Altman declared the "Year of the Agent" in December 2024, OpenAI has shifted its commercial strategy from a model-capability focus to a product-first approach. The series of models it subsequently released lacked the stunning impact of earlier ones. The market has also evolved from ChatGPT's dominance to the current tripartite standoff with Gemini and Claude. Meanwhile, OpenAI began investing significant resources into productization attempts. The launch of the Sora App was one such attempt.
Therefore, rather than Sora's shutdown being a strategic contraction for OpenAI, it's more accurate to say it was just another failed AI product.
If it were just an ordinary product failure, it wouldn't be worth much discussion. But Sora is a bit different. In the text-to-video generation field, few could rival it. Over the year since its release, its output quality and stability remained outstanding. Its model capability wasn't the issue at all. However, examining the Sora App from a product perspective reveals its biggest problem: over the past year, it did almost nothing, making no progress whatsoever in "becoming a usable tool."
Anyone who has used text-to-image/video generation knows that to produce a passable video, the first step is writing a clever prompt, and the second is the model's "gacha" (random draw). With the best luck, you might get the desired effect in one or two tries, but most attempts are random shots in the dark. Not getting a Van Gogh-style effect is already considered a blessing.
This fixed routine of "high probability failure" makes it impossible to know whether the problem lies in an insufficiently good prompt, the model being in a bad mood, or because you handled pickled herring without washing your hands. You can't know if the next attempt will be closer to your ideal result, nor can you, like using a tool, gradually steer the outcome in the desired direction.
This is a major taboo for a product. Once users cannot pinpoint where the problem lies, it's hard to form any experience. Using it ten times is essentially no different from using it once. The cliff-like drop in experience caused by immense uncertainty, having to start from the beginner's village every time you boot up—no one can endure that.
The ecosystem didn't take off, the interactive experience made no progress, 30-day user retention was almost zero, and偏偏 its operational costs were extremely high. This is the real reason Sora was taken offline.
If growth had taken off, or if OpenAI had defined the paradigm for AI applications through Sora, the daily burn rate would have been a drop in the bucket. But it neither retained users nor fostered user habits. Creators couldn't profit; each generation was a one-time consumption. All the high-quality content was funneled to short-video platforms for monetization, which is extremely awkward.
In a product with a complete ecosystem, all operational costs should be offset by economies of scale. Yet Sora became a tool to help other platforms offset their costs. It was like a one-way funnel, with all costs being net outflows. In this situation, for OpenAI, more users meant more pain, fewer users meant more distress. It could neither rely on growth to save itself nor cut losses through contraction.
In the end, the only choice was to take it offline.
Looking back at Sora now, aside from those aspects directly covered by its model capabilities, its failure exhibits the typical characteristics of all past "waves" that crashed on the shore: no ecosystem to form a positive feedback loop, an interactive experience that's like a ghost car drifting out of control, and an imbalanced cost structure.
These structural issues are key factors determining whether an AI tool can truly be productized. Cross them, and you soar to the heavens; fail to cross them, and the only option is to pack up and run.