OpenAI Disbands Robotics Team, Betting General Models Can Now Run Robots

OpenAI has dissolved its stand-alone robotics team for the second time in six years, reassigning researchers to its main research organization and applied robotics staff to its broader Applied division. The company confirmed the reorganization to Fortune on November 13, framing it as an acceleration of its robotics work rather than a retreat.

The robotics research group now reports to Chief Research Officer Mark Chen, while applied robotics and hardware teams move under President Greg Brockman. A spokesperson said robotics "remains a core part of OpenAI's long-term vision" and that the change reflects a discovery: the company's general-purpose frontier models, including GPT-6 Astra, can be applied directly to robotics without a dedicated robotics stack.

A second reset

This marks the second time OpenAI has disbanded a stand-alone robotics team in six years. The company ran a robotics team in its early years, then wound it down in 2020. Cofounder Wojciech Zaremba later cited limited robot data and a more efficient path to artificial general intelligence through large language models—a pivot that produced ChatGPT in late 2022.

OpenAI rebuilt a stand-alone team in 2024 and announced a partnership with humanoid startup Figure. That partnership ended in February 2025 when Figure said it would develop its own robot AI. OpenAI has since said it is building its own humanoid, with CEO Sam Altman targeting a first public demonstration in 2027.

The timing of the latest reorganization was visible at CoRL, a top robotics conference in Austin this week. OpenAI had registered booth #208, but Fortune found it empty, with no company logo on display. A few former team members milled about, drawing a crowd despite having no demo.

The bet: general models as robot brains

The move reflects a bet gaining traction across the AI industry: that models built for language, code, and vision can also drive "embodied AI"—robots, driverless cars, drones—obviating the need for separate teams building specialized robot brains.

OpenAI told Fortune that at the start of the year, researchers assumed the best path was to build on video models such as Sora, which must hold an implicit understanding of physical dynamics to generate realistic footage. Since then, the spokesperson said, general-purpose frontier models have shown they can be applied to robotics directly. That finding motivated the reorganization.

The strongest evidence cited comes from an Nvidia study. Researchers gave GPT-6 Astra camera images, robot state information, and a control interface to command a simulated robot. The Astra-powered agent achieved a 74.8% success rate on a simulated manipulation benchmark, compared with 51.0% reported by the best specialized robotics models on that leaderboard.

Fortune noted the test was in simulation, not on a physical robot. An analysis by explainx.ai cautions that a single simulation benchmark does not cover contact-rich real-world tasks.

Expert reactions are mixed

Guanya Shi, an assistant professor at Carnegie Mellon's Robotics Institute, read the move as a signal that OpenAI sees robotics as "no longer a special problem" but as a downstream application of general intelligence—similar to how the same frontier models now power coding, legal, and financial tools. He said advances in spatial and reasoning capabilities let general models cope with a bottle in a different position or changing lighting, and reason about which step went wrong.

Shi still sees a need for smaller, specialized models that run on a robot's onboard computer. Large language models, however, are slow and typically cloud-hosted, requiring a constant network connection.

Russ Tedrake, an MIT professor and CEO of Walden Robotics, argued that robotics companies retain an edge: their robots are already accumulating work experience in real places such as factory floors. That deployment data helps robots adapt to specific workplaces, and better foundation models make such site-specific data more valuable because robots can learn from it faster.

Pete Florence, cofounder of Generalist, a startup training general-purpose AI models for robots, said it is unclear that any existing model will dominate robotics in the long run.

Business pressure backdrop

The reorganization comes as OpenAI faces pressure to rein in costs and focus on core products. Altman told staff last year to abandon "side quests," and the company has scaled back other efforts, including Sora, to reallocate computing resources. Fortune reported that operating losses reportedly totaled $10.9 billion in the most recent quarter as OpenAI seeks to close a new funding round and prepares for a possible IPO in early 2027.

OpenAI's statement emphasized the research motivation, not cost pressures. The spokesperson said the reassignment was motivated by the research finding that general models can be applied directly to robotics.

What remains unknown

The number of affected researchers was not disclosed. The Nvidia result is a simulation benchmark; real-world deployment is a different test.

For robotics startups that have raised billions on the premise that robots need bespoke brains, OpenAI's bet poses a direct challenge. The counterargument is that real-world deployment data—and the ability to run models locally—remain advantages that general-model labs do not yet have.