In 2025, the market performance of Sora 2, a video generation product under OpenAI, can be described as a “cliff-like decline” in the latest AI news. Hailed as a “nuclear weapon” expected to disrupt the existing short-video landscape, Sora 2 achieved impressive results in its early launch phase, leveraging OpenAI’s brand influence and a full-platform promotion strategy. It recorded over 1 million installations on iOS in the first week and 470,000 downloads across seven markets (with the U.S. accounting for 63% of the total) on Android on its launch day, even topping the App Store’s free app chart at one point. However, just two months later, market feedback took a sharp turn for the worse. According to data from SensorTower, Sora 2’s 30-day user retention rate was only 1%, and its 60-day retention rate was close to 0%, even below the industry average.
Failure to Evolve from a “Toy” to a “Tool”
The core reason for Sora 2’s dismal retention rate lies in its failure to break free from its “toy attribute” and meet users’ practical needs. Firstly, both the quality and efficiency of video generation are extremely low. In practical tests, the hit rate of usable videos was only 5%-10%, meaning users often had to generate videos repeatedly to get a satisfactory result.
Additionally, a single rendering process takes several minutes, and when combined with the high failure rate, the time cost for users becomes extremely high. This issue stems from the fact that OpenAI’s image generation capability is still stuck in the GPT-4o era, which directly limits the quality ceiling of video generation—problems such as distorted facial expressions of characters and chaotic physical logic in the footage are common occurrences.
Secondly, the product has obvious functional shortcomings. Sora 2 lacks sophisticated editing features; users cannot modify videos directly and can only adjust prompts repeatedly to regenerate content. Its community function is even “disastrous”: there are no comment or favorite functions, videos cannot be paused, high-quality works fail to gain exposure while low-quality content occupies the recommendation slots instead, and the “search for similar terms” function offers a terrible experience—searching for “magic witch” actually brings up irrelevant content like tanks, the moon, and some historical figures. Furthermore, its “Cameos” feature has become embroiled in legal disputes over alleged trademark infringement, further hindering the promotion of this function.
Commercial Dilemma: The Computing Power Black Hole and a Difficult Choice
Beyond product experience issues, Sora 2 also faces an intractable commercial dilemma. OpenAI has to pay $15 million daily for Sora 2’s computing power costs (nearly $5.5 billion annually), but the current business model is “completely unsustainable”. Initially, the platform provided users with 30 free video generation quotas per day, and later launched a paid service where users could purchase 10 additional quotas for $4. However, a retention rate of less than 1% means that the vast majority of users will not pay, and the payment amount from heavy users is far from covering the costs.
OpenAI is caught in a dilemma: maintaining free quotas will lead to uncontrolled costs, reducing free quotas will accelerate user churn, and increasing the paid price may lower market acceptance. At the same time, intellectual property risks exacerbate the predicament—users may generate infringing content, so the platform has to restrict the types of content that can be created. However, such restrictions damage the user experience, leading to user dissatisfaction due to repeated policy adjustments.
Sora 2’s setback is not just the failure of a single product; it also sounds an alarm for the entire AI video generation industry: technological leadership does not equal product success, and download volume does not mean user retention. Currently, the entire track is still in the phase of burning money for exploration, and there is still a long way to go before achieving a healthy profit model.
