OpenAI Shut Sora and Shipped “Spud”—but the Safety “Demotion” Is Unproven
The mystery surrounding OpenAI’s March 2026 changes has partly cleared. Axios identified the “Spud” model as GPT-5.5 when it launched on April 23, while OpenAI’s current Sora page records that the video product became unavailable on April 26.
What remains unresolved is the claim that safety was demoted. Axios’s March 24 account of the restructuring placed safety work under chief research officer Mark Chen, security under president Greg Brockman and responsibility for capital and computing infrastructure more firmly with CEO Sam Altman. Those reporting lines are consequential, but they do not reveal who can block a model release or whether safety teams lost formal authority.
The March decisions were related, but not one proven AGI program
OpenAI combined a consumer-product closure, a research redirection and a management reshuffle during the same period. It also completed initial development of the model family carrying the internal “Spud” name. The timing supports the conclusion that the company was concentrating leadership attention, staff and scarce computing capacity.
It does not establish that OpenAI had achieved artificial general intelligence or reorganized in response to a system already beyond human control. “AGI Deployment” may express the company’s ambition, but an organizational label is not a capability evaluation. No disclosed evidence connects the label to a model that has met a generally accepted AGI threshold.
The strongest business interpretation is narrower. OpenAI reduced investment in a compute-intensive consumer experiment while prioritizing models designed to complete longer sequences of digital work. That strategy can be aggressive without requiring the more extraordinary explanation that an autonomous system is setting corporate priorities.
Sora ended as a product, not as a field of research
Sora’s consumer closure is no longer speculative. The product is unavailable, ending the app and web experience that allowed users to generate and share videos. That makes “Sora was killed” a reasonable shorthand for the consumer offering, but an inaccurate description of everything OpenAI was doing with video models.
The research team was set to continue work on world simulation for robotics and physical tasks. That distinction matters: retiring an application can free computing capacity and employees while preserving the underlying technical program. It also means there is no basis for claiming that Sora’s entire dataset was transferred to a single robot-training project.
The closure therefore reflects a portfolio decision rather than the disappearance of video research. OpenAI abandoned a public-facing distribution model while retaining an interest in systems that learn how objects and environments behave. Whether that research produces commercially useful robotics technology remains an open question.
“Spud” became a product release, not evidence of AGI
The codename ceased to be the central mystery once GPT-5.5 arrived. The released model was positioned for agentic coding, computer use, office tasks and early scientific research, with an emphasis on planning and completing multi-step workflows with less continuous direction from the user.
That focus fits the resource shift visible in March. A model capable of sustained work across software and business tools has a clearer route to enterprise revenue than a standalone social-video application. It also depends on the same expensive ingredients OpenAI was prioritizing: advanced chips, data-center capacity and infrastructure able to serve longer-running tasks.
Agentic performance should not be confused with general intelligence. A system may become better at tool use, planning and persistence while remaining unreliable outside tested workflows. Neither the internal codename nor the public release name demonstrates that GPT-5.5 can perform the full range of cognitive work implied by AGI.
Why the safety claim exceeds the available evidence
The management change creates a legitimate governance question because safety work now sits closer to the research organization whose models it evaluates. That arrangement could reduce independence if research leadership controls budgets, escalation routes and release decisions. It could also improve coordination by placing evaluations and mitigations directly inside model development.
An organization chart alone cannot distinguish between those outcomes. The missing information includes whether safety personnel retain direct access to senior leadership or the board, whether they can delay deployment, and what happens when research and risk teams disagree. Without those decision rights, describing the change as a definitive demotion turns a plausible concern into an unsupported fact.
Safety and security also cover different functions. Model safety concerns evaluations, dangerous capabilities, behavioral safeguards and deployment controls; corporate security protects infrastructure, model weights, personnel and sensitive information. Assigning them to different executives may alter accountability, but it does not show that either function was abolished.
The durable signal is concentration around agentic work
The verified sequence points to strategic concentration. OpenAI withdrew from a visible consumer-video product, preserved selected world-model research, released the former “Spud” project as a model for extended computer work and reassigned senior responsibilities while Altman focused on capital and infrastructure.
That concentration carries a real tension for the business. The systems with the strongest commercial case—agents that can code, operate software and complete multi-stage assignments—also demand stronger evaluation, monitoring and security because they act for longer and with less supervision. Faster coordination between research and safety could help, but reduced independence could create pressure to favor deployment.
The current record therefore supports a hard pivot toward agentic AI, not the more dramatic claims attached to it. Sora’s consumer product is gone, “Spud” surfaced as GPT-5.5 and safety responsibility changed hands. Evidence that OpenAI has achieved AGI, eliminated meaningful safety oversight or ceded corporate decisions to an AI system has not emerged.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.