OpenAI moved Sora from a research preview to a paid product

The shot

On 9 December 2024, OpenAI moved its text-to-video model out of research preview and released it as a standalone product at Sora.com, open to ChatGPT Plus and Pro subscribers. OpenAI's own launch announcement names the new version Sora Turbo and states it is "significantly faster than the model we previewed in February," the earlier research-only demonstration shown ten months before. This is a distinct, dated step: a lab-controlled preview becoming a subscription feature any qualifying account could open the same day, not a description of that February demo itself.

What the documents show

The announcement states subscribers could generate video "up to 1080p resolution, up to 20 sec long," in widescreen, vertical, or square formats, with a storyboard tool for setting inputs frame by frame. The accompanying system card, published the same day, adds detail the announcement omits: external red-teamers in nine countries tested over 15,000 generations before release. Both documents are OpenAI's own account of its release; the featured capabilities are a promotional selection, but the resolution, duration and testing figures are stated as facts about the shipped product. The system card is explicit that "the version of Sora we are deploying has many limitations," naming unrealistic physics and difficulty with complex actions over long durations.

The workflow

A subscriber could bring an existing image or clip to extend, remix, or blend, or generate new footage from text alone. Plus accounts were capped at 50 generations a month at 480p, or fewer at 720p; Pro accounts received "10x more usage, higher resolutions, and longer durations," per the system card. The storyboard interface let a user assign instructions per frame, closer to a shot list distributing action across a sequence than to one long prompt. Every output carried C2PA metadata and, per the system card, "animated visible Sora watermarks by default." Likeness uploads were limited to a pilot group at launch, gated by age verification.

What the tool does not change

The system card frames this release as one still routed through human review: red-teaming, moderation classifiers, and blocklists sat between a prompt and a shown result, and OpenAI states that risk areas including artist-style prompts and likeness uploads were handled with conservative, human-set policy rather than left to the model's judgment. Consent for an uploaded likeness, and what a finished generation is used for, remained assignments OpenAI gives to policy and review, not to Sora Turbo itself.

• Does a planned use fall inside the resolution and duration ceilings this document states, or does it need a later version's own documentation?
• What does the cited system card say about likeness uploads, and does that match the specific use a production has in mind?
• Is a claimed capability sourced to this release, or has it since been revised by a newer model's own materials?

Read narrowly, the 9 December 2024 announcement documents a distribution change more than a technical leap: the same lab-stated resolution and duration ceilings, now reachable by subscription rather than invitation. The system card, not the highlight reel, is the more useful record of what the tool could and could not do at that release.