OpenAI saw two senior departures on April 17. Bill Peebles, a key leader behind Sora, and Kevin Weil, the company’s former chief product officer and founder of OpenAI for Science, both said they were leaving. At the same time, OpenAI for Science stopped operating as a standalone division.
Sora shutdown was followed by Peebles' exit
The source says Sora was shut down last month, not because it lost in product competition, but because computing costs became too heavy to sustain. The report estimates the video generation model was losing about $1 million per day in compute expenses. Peebles had helped take Sora from a two-person effort to a formal launch, and chose to leave after the release of Sora 2.
In his post on X, Peebles said building Sora from zero was the honor and adventure of a lifetime. He described one of the best parts of working on video models as being able to directly observe scaling behavior. The message was restrained in tone, though it made clear how central the project had been to his time at OpenAI.
Weil leaves one day after GPT-Rosalind launch
Kevin Weil’s departure came just one day after the April 16 debut of GPT-Rosalind, a model developed by OpenAI for Science under his leadership. Named after Rosalind Franklin, the scientist associated with the discovery of the DNA double helix, the model was positioned for life science reasoning. Its listed partners included Amgen, Moderna, Allen Institute, and Thermo Fisher Scientific.
On X, Weil said that April 17 was his last day at OpenAI because OpenAI for Science was being broken up and integrated into other research teams. He said the past two years had taken him from chief product officer into the research side of the company, where he went on to start OpenAI for Science.
OpenAI narrows focus to enterprise AI
The source cites Bloomberg’s reading of the reshuffle: OpenAI is continuing to cut so-called side quests and concentrate resources on enterprise AI and a superapp strategy. Sora sat in consumer video generation, where costs were high and monetization was slow. OpenAI for Science was aimed more at academic and scientific work, which was also outside the company’s current commercial center.
Placed side by side, the sequence is striking. GPT-Rosalind was launched, then its division was folded into other teams. Sora had already been shut down, and its core lead then walked away. The two moves point to the same internal direction: fewer non-core efforts, tighter use of compute, and a stronger concentration on businesses closer to OpenAI’s main revenue focus.