Sam Altman’s $1bn deal with Walt Disney to bring characters such as Mickey Mouse and Luke Skywalker into OpenAI’s artificial intelligence video platform marks an effort to boost the Sora app after it has struggled to build an engaged audience.
The content and investment deal announced on Thursday averts a costly legal showdown over intellectual property but faces serious obstacles to reviving the Sora platform, which is grappling with disappointing uptake and challenging economics.
“The demand for Disney characters in particular from our users is sort of off the charts,” Altman, chief executive, told CNBC on Thursday. He said the deal would be “a wonderful start for what our customers want to do when it comes to putting themselves in that . . . lightsabre fight in Star Wars or making a Buzz Lightyear custom birthday video for their kid”.
The companies’ vision of bringing the magic of Disney’s characters to Sora for the first time belies the fact that users are already creating bootleg animations, with several accounts making imagined movie trailers and other content in the distinctive aesthetic of Disney and Pixar films.
Unofficial accounts have used the Disney and Pixar logos, and portrayed characters resembling Toy Story’s Buzz Lightyear and the child from the film Up. However, these profiles have generally amassed only hundreds of followers, with more popular accounts gaining just over 1,000 likes.
The limited reception for Disney-style content on Sora reflects the app’s wider struggle to attract enough engaged users to justify its high running costs.
Despite leaping to the top of the download charts when it launched its standalone app in September, Sora has not translated that early interest into a habitual daily audience comparable to those of conventional social media services such as TikTok or YouTube.
The app offers an experience similar to scrolling on videos on TikTok or Instagram’s Reels, except users create the clips that they share by giving simple prompts to the app’s AI model.
Prior to this week’s deal, Sora was not “gaining traction”, said two people familiar with its performance.
About 25 per cent of Sora’s 7mn monthly users open the app daily, according to estimates from Sensor Tower, with people spending an average of 13 minutes on the app.
Conventional social media apps such as TikTok keep users engaged for as much as 90 minutes a day, according to the market intelligence firm.
“People aren’t spending time scrolling [on Sora],” said another person with knowledge of the matter.
Audiences have shown an aversion to so-called “slop”, the term used in the AI industry for low-value, AI-generated content. Sora users “dip in and dip out because [there is] only so much slop you can take”, the person said.
OpenAI cited its “feed philosophy” for Sora, which said it aims to “optimise for creativity. We’re designing ranking to favour creativity and active participation, not passive scrolling. We think this is what makes Sora joyful to use.”
Another problem is the enormous cost to OpenAI of letting users experiment for free on Sora. Cantor Fitzgerald’s analysts estimate video model providers spend an average of $1.30 to generate a 10-second video, while consultancy SemiAnalysis said the average ChatGPT query costs about half a cent.
Video generation is vastly more complicated and demands more computing power than processing text for chatbots. SemiAnalysis said the difference between Sora and ChatGPT resembled comparing the fuel efficiency of a jet to that of a pick-up truck.
Wei Zhou, head of AI utility research at SemiAnalysis, said: “The resolution of the videos generated by Sora is low because they have to optimise for cost and consider their compute resources, alongside ChatGPT and research. They’re a business, and they have to think about optimising business outcomes, and Sora is expensive to run.”
Bill Peebles, OpenAI’s head of Sora, in October said “the economics are completely unsustainable” and the company had to limit the number of free-generated videos users could make. Currently, users of the app get a limited number of daily credits to create videos for free and can purchase more.
Sora is among the OpenAI products that Altman plans to deprioritise until next year after declaring a “code red” in an internal memo last week and telling staff to focus on keeping its core ChatGPT chatbot ahead of rivals such as Google and Anthropic, which have made rapid advances.
Against the backdrop of Sora’s high costs and limited traction, the Disney deal allows OpenAI to avoid the risk of a costly legal fight over IP.
The three-year agreement makes Disney the first Hollywood studio to license its content to OpenAI. Sora’s users will be able to choose from about 200 characters from Disney, Marvel, Pixar and Star Wars — though the agreement does not include any talent likenesses or their voices. This means users can make videos using a cartoon image of Han Solo, but not Harrison Ford’s face or raspy voice.
Disney will take a $1bn stake in the $500bn start-up and will be granted warrants to buy more equity in OpenAI at a nominal price over time in return for licensing its characters.
The tie-up falls short of OpenAI’s ambitions to have Hollywood studios adopt Sora for filmmaking, but keeps the company out of Disney’s legal crosshairs.
Bob Iger, Disney’s CEO, on Thursday said: “We’ve been aggressive in protecting our IP.” The Hollywood group this week issued a cease-and-desist letter to Google accusing it of infringing on its copyrights “on a massive scale” by copying works without authorisation to train its AI services.
The letter claimed Google, which has its own video generation product Veo, was “flooding the market with infringing works”.
Google said it has had “a long-standing and mutually beneficial relationship with Disney, and will continue to engage with them”, adding that it uses public data to build its AI and has “additional innovative copyright controls”.
AI companies are facing lawsuits from media organisations, authors and news publishers. In September, Anthropic settled a lawsuit for $1.5bn, paid out to authors who claimed their content was used to train its AI model Claude. Other cases brought against Meta and Stability for using books and photos to train their models have failed to win in court.
“Litigation is proving to all sides a blunt instrument with outcomes unpredictable,” said Joel Smith, an IP partner at international law firm Simmons & Simmons. “Recovery of damages for past use is always difficult and expensive to quantify.”
Smith said content deals are often “more straightforward” and can cover “retrospective use of content, as well as future use”. (The Financial Times has a partnership and licensing agreement with OpenAI.)
Additional reporting by Hannah Murphy
