OpenAI's first Sora preview reached artists, not the public

The shot

On 15 February 2024 OpenAI published a page introducing Sora as a text-to-video model, stating it could generate clips 'up to a minute long' while holding visual quality and prompt adherence. The announcement page says Sora was 'becoming available to red teamers' to probe for harms, and that OpenAI was separately granting access to visual artists, designers and filmmakers for feedback. A follow-up page dated 25 March 2024, Sora: first impressions, states the company had worked with these creators 'since we introduced Sora to the world last month,' and profiles participants including the Toronto studio shy kids and director Paul Trillo. Neither document describes a general launch; a broader paid tier arrived only in December that year.

What the documents show

OpenAI's own materials frame February 2024 strictly as a research preview on two access tracks: red-teaming for safety evaluation, and a curated artist program for creative feedback. The announcement page states the company planned to add a detection classifier and C2PA metadata only 'in the future if we deploy the model in an OpenAI product,' phrasing that confirms no deployment had happened at the time of posting. The first-impressions page corroborates the access model from outside the company: it quotes invited artists describing hands-on experimentation rather than self-service signup, consistent with a closed cohort rather than a tool available to filmmakers generally.

The workflow

For a production weighing Sora at this stage, the entry point was an invitation, not a subscription or an API key, routed through the safety-testing channel or the creative partner program. The company's own weakness examples double as a workflow warning: a bitten cookie shows no bite mark afterward, and a chair the model fails to treat as a rigid object while it is being excavated. Any resulting footage needed frame-by-frame review against ordinary continuity checks before it could inform a shot list, the discipline a supervising editor applies to any pre-visualisation asset regardless of how it was produced.

What the tool does not change

The published weaknesses - spontaneous entities appearing mid-shot, left-right confusion, implausible object behaviour - are named in the vendor's own words, not inferred from outside criticism. Nothing in either document substitutes a director's judgment of whether a shot serves a scene, and neither page claims the featured clips were production-verified footage rather than selected demonstrations. That distinction is editorial but load-bearing: a showcase reel is evidence of what a model can be made to produce under curation, not what it reliably produces on a single request.

• Which access track, safety testing or the artist program, would a given request have qualified for at this stage of the release.
• Does a demonstration reel disclose how many attempts or how much selection produced the clips shown to the public.
• What has the vendor stated it still plans to build, and does that mean the feature was not yet functioning at announcement.

Read against its own text, OpenAI's February 2024 announcement is a bounded research disclosure, with an admitted list of weaknesses and a safety program under construction, not a broadcast that a production tool was open for general use.