A minute of video, and a list of what it still gets wrong
OpenAI previewed Sora on February 15, 2024, describing it as a model that can generate videos up to a minute long from a text prompt while keeping the same characters, camera motion, and background detail across the clip. It can also animate a still image or extend an existing video. [1]
OpenAI's own post listed the failure modes alongside the demo reel: Sora could confuse left and right, miss cause and effect such as a bite mark failing to appear after it happens, and produce objects that pop into a crowded scene or move in physically implausible ways over a long shot. [1]
Atlas interpretation: Publishing the weaknesses next to the highlight reel is unusual for a product announcement, and it is also the only independent check available at the time. Every clip in the post was OpenAI's own selection, generated on prompts OpenAI chose, so the failure list is the more informative half of the page precisely because it was harder for the company to spin. [1]
Access went to red teamers and a short list of filmmakers
OpenAI said it was granting access to a small group of red teamers, described as domain experts in misinformation, hateful content, and bias, to test the model adversarially, and separately to a group of visual artists, designers, and filmmakers to get feedback on making the tool useful for creative work. Everyone else could only watch the samples OpenAI chose to publish. [1]
OpenAI also described safeguards planned before any wider release: text classifiers to reject prompts that violate its usage policies, including extreme violence, sexual content, hateful imagery, celebrity likeness, and others' intellectual property, image classifiers to review individual frames before a video ships, and a detection classifier built to tell whether a video came from Sora. It said produced videos would carry C2PA metadata identifying them as AI-generated. [1]
Atlas interpretation: None of that tooling shipped alongside a product, because there was no product yet. A capability demonstration paired with a safety plan for a release date OpenAI did not name is a different kind of announcement than a launch, and reporting on Sora treated it as one: coverage of the demo outran coverage of anyone actually using it, for the better part of a year. [3]
Sora Turbo reached subscribers in December, still short of the demo
OpenAI shipped a public version on December 9, 2024, as a standalone product at sora.com, built on a model it called Sora Turbo and described as significantly faster than the version previewed in February. It generates video up to 1080p resolution and up to 20 seconds long, in several aspect ratios, with a storyboard tool for specifying content frame by frame. [2]
Access went to ChatGPT Plus and Pro subscribers: Plus accounts got up to 50 videos a month at 480p, Pro accounts roughly ten times the usage at higher resolutions and longer durations. OpenAI held the product back from the UK, Switzerland, and the European Economic Area, and restricted it to users 18 and older. It kept the same watermark and C2PA metadata commitments described in February, and said the model still often produced unrealistic physics and struggled with complex actions over a long duration. [2]
Atlas interpretation: Ten months between a capability demo and a product any subscriber could pay for, arriving with a shorter maximum clip length than the one-minute samples shown in February and the same physics complaints still listed as known limitations, set the pattern other vendors would repeat: the announcement is the model at its best, on prompts the vendor picked, and the shipped product is what survives contact with a much larger and less cooperative set of users. [2]
Sources
- Sora: Creating video from text
OpenAI · Feb 15, 2024
- Sora is here
OpenAI · Dec 9, 2024
- Sora (text-to-video model)
Wikimedia Foundation · Sep 8, 2026