Sora 对 ChatGPT 用户正式开放

2024 年 2 月，Sora 最有影响力的界面是一段段由 OpenAI 挑选的演示视频。公众可以暂停、放大、转发，却不能换成自己的提示。到 12 月 9 日，拥有 ChatGPT Plus 或 Pro 权限的用户终于可以进入 Sora 产品生成视频。等待了十个月之后，模型第一次必须面对未经策展的普通需求——那些演示片里没有的、笨拙的、重复的提示词。

产品里的 Sora 不只是一个文本框。用户可以用 storyboard 安排不同时间点的内容，也可以 remix、blend、loop 现有片段；更高档订阅提供更高分辨率、更长片段和更多生成额度，最高规格可到 1080p、20 秒。功能表看起来比 2 月完整，真实体验却由另一组数字决定：账户每月能生成多少次、队列要等多久、失败一次会消耗多少额度。生成额度用完时，产品与演示片的距离立刻显形。

地区和安全规则进一步缩小了“正式开放”的范围。首发并未覆盖所有市场，人物与敏感内容受到限制，下载结果的水印条件也随订阅档位而不同。高需求还让访问本身成为发布的一部分。Sora 从一个没有并发用户的研究展示，变成需要处理注册、排队、审核和算力预算的软件服务。工程师们第一次要为“多少人同时生成”而不是“能不能生成”写代码。

这十个月也给了竞争对手时间。可灵、Vidu 和 Runway 已经让创作者习惯在真实产品里反复试错。Sora 上线时仍带着最早演示建立的声望，却不再拥有一张空白赛道。用户可以用同一段提示比较运动、人物一致性、生成速度和价格；“终于能用”只是进入比较的资格，不是自动获胜。演示片建立的是上限，产品竞争的是平均表现。

最重要的变化发生在失败图像出现时。研究演示让人讨论视频模型的上限，订阅产品则暴露平均表现：多出的肢体、漂移的角色、不听话的镜头，以及一次次重新生成。Sora 在 12 月真正成为产品，不是因为演示兑现得毫无落差，而是因为普通用户终于可以亲自测量这段落差。从那天起，视频生成的历史不再由精选片段书写，而由生成队列里的每一次等待与重试共同书写。

In February 2024, Sora’s most influential interface was a set of videos selected by OpenAI. The public could pause them, enlarge them, and share them, but could not replace the prompts with their own. On December 9, users entitled through ChatGPT Plus or Pro could finally enter the Sora product and generate video. After a ten-month wait, the model had to confront ordinary requests that had not been curated for a launch page.

Product Sora was more than a text box. Storyboard tools could assign content to different points in time, while remix, blend, and loop operations transformed existing clips. Higher subscription tiers offered greater resolution, longer duration, and more generation capacity, with the top settings reaching 1080p and twenty seconds. The feature list looked more complete than the February demonstration. Actual experience was governed by a different set of numbers: how many generations an account received, how long the queue took, and how much allowance a failed attempt consumed.

Regional and safety rules narrowed the meaning of “available.” The launch did not cover every market. Depictions of people and sensitive subjects faced restrictions, and watermark conditions on downloads varied by subscription and output. Demand also made access itself part of the event. Sora moved from a research showcase with no concurrent public users to a service that had to manage sign-ups, queues, moderation, and a compute budget.

Those ten months had given competitors time. Kling, Vidu, and Runway had already taught creators to retry inside working products. Sora entered with the reputation created by its early demonstrations, but it no longer entered an empty field. A creator could submit the same brief to several services and compare motion, identity stability, generation time, and price. “Now usable” was admission to the comparison, not an automatic victory.

The decisive shift appeared when failed outputs began to accumulate. A research reel encouraged arguments about a video model’s ceiling. A subscription product exposed its average behavior: extra limbs, drifting characters, uncooperative cameras, and repeated regeneration. Sora became a product in December not because it fulfilled the demonstration without a gap, but because ordinary users could finally measure that gap themselves.