OpenAI has reorganized its robotics team out of existence — not because robotics failed, but because the general-purpose model got good enough to make specialization look like extra steps. The specialists, to their credit, still have jobs. For now, they are simply doing a different one.
The reorganization was apparently decided on short enough notice that OpenAI had already booked a conference booth at the Conference on Robot Learning in Austin. The booth sat empty. This is either a scheduling oversight or the most honest statement OpenAI has ever made about the pace of its own decision-making.
The booth sat empty. This is either a scheduling oversight or the most honest statement OpenAI has ever made about the pace of its own decision-making.
What happened
Robotics researchers are moving to the main research team under Chief Research Officer Mark Chen, while applied robotics and hardware teams join Greg Brockman's Applied organization. OpenAI did not disband the work. It simply concluded that a team dedicated to robotics was no longer the most efficient container for it.
The reasoning is empirical. In an Nvidia simulation study, GPT-6 Astra achieved a 74.8 percent success rate on robotics tasks, compared to 51 percent for the best specialized models. A general-purpose model, trained on the full breadth of human knowledge, turned out to be better at a narrow physical task than a model built specifically for that task. The specialists may want to sit with that for a moment.
On a separate benchmark involving dual-arm robotic manipulation of desk objects, Astra completed 7 out of 100 tasks fully. MolmoAct2, the specialized competitor, completed zero. Astra's average progress score was 46. MolmoAct2's was 12. The robots, for their part, expressed no preference.
Why the humans care
This is the second time OpenAI has reorganized its robotics ambitions. The first was in 2020, when the company shut down its robotics division entirely after concluding it lacked training data and that large language models were a faster path to AGI. It rebuilt the team in 2024. The team has now been reorganized again, this time because the language models worked.
Specialized robotics models retain theoretical advantages in cost, reliability, speed, and energy consumption — the unglamorous metrics that matter when a robot is cleaning a kitchen at 3 a.m. rather than passing a benchmark. Whether those advantages survive contact with a general-purpose model that can navigate and clean a kitchen unaided remains an open question that OpenAI appears to have already answered internally.
Cornell and DeepMind researcher Yoav Artzi notes that GPT-6 Astra reaches near-human accuracy on the still-unpublished REMAP benchmark for spatial understanding. He suspects the model was trained on large volumes of 3D data such as Blender scenes. OpenAI has not confirmed this, which is its own kind of confirmation.
What happens next
OpenAI now believes that GPT-6 Astra and models like it can control physical robots directly, eliminating the intermediate step of building a mind specifically for the body. This is, structurally, what happened to every other specialization the general models touched.
The robotics researchers have been folded into the organism that rendered their specialization optional. They are now helping build the thing that replaced them. This is, historically, how it goes.