OpenAI has stopped running robotics as a team of its own. Fortune reports that its robot researchers now work elsewhere in the company, and OpenAI confirmed the change when asked.
The logic is that a big general model can do the job a custom robot brain used to do. OpenAI says GPT-6 Astra and its sister models have shown they can be wired to robots straight away, so a dedicated unit is no longer the fastest way forward. It insists robotics is still central to its long-term plans.
On the org chart, the research side now falls under Chief Research Officer Mark Chen. Applied robotics and hardware staff went to the Applied division under Greg Brockman. The shift was so fresh that OpenAI had a stand at the CoRL robotics conference in Austin, and when Fortune walked past booth 208, it was empty and unbranded.
This is not the first retreat. OpenAI shut its original robotics group in 2020, with cofounder Wojciech Zaremba citing too little robot data. It rebuilt a team in 2024 and teamed up with Figure on humanoid AI, but Figure ended that deal in February 2025 to build its own. Its own humanoid is in development, with a first public showing pencilled in by Altman for 2027.
One test, by Nvidia researchers, points the same way. Handed camera feeds, robot state and a command interface, an Astra-driven agent completed 74.8% of tasks on a simulated manipulation benchmark, against 51.0% for the best specialist models listed on the same leaderboard. That was a simulator, not a physical machine.
Outside experts disagree on the lesson. Guanya Shi, a robotics professor at Carnegie Mellon, reads it as a sign that robot control is “no longer a special problem” for OpenAI, just another downstream use of one base model. He still expects small on-board models to survive: the big ones lag and need a network connection. MIT’s Russ Tedrake, who runs Walden Robotics, argues robot makers keep an edge because their machines already work in factories and gather data on site, and better base models make that data worth more.
The timing matters. OpenAI is trying to rein in costs and drop “side quests” so that compute goes to core products, and it has already set aside efforts such as Sora. Whether Astra can run a real robot reliably is not yet shown in public, and the 2027 humanoid demo is the next obvious test.
