OpenAI has reorganized its robotics team, and the company's framing matters. OpenAI says it is not abandoning robotics but accelerating it by folding researchers into other groups. 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. The company has since clarified that the change was about working more closely with research and product, and that robotics work continues.

The strategic claim underneath the reorganization is the important part. OpenAI apparently believes general-purpose models like GPT-6 Astra can control robots directly, reducing the need for specialized robotics models. The evidence offered is promising but thin. An Nvidia study reported a 74.8 percent success rate for an Astra-powered agent on simulated robotics tasks, against 51 percent for the best specialized models. The report notes that these tests used a simulated DROID robot, not a physical one, which is a significant limitation for any claim about the real world.

The StationeryBench results are more revealing, and more sobering. Across 200 trials, Astra fully completed 7 out of 100 tasks, while Ai2's MolmoAct2 completed none. Astra's average progress score was 46, compared with 12 for MolmoAct2. Those are real gains, but a model that fully completes 7 percent of desk-object tasks is not yet a dependable household or factory worker. A kitchen-cleaning demonstration on a humanoid robot is encouraging, but a single demonstration is not a reliability record.

Readers should also keep the unanswered questions in view. The report says it remains unclear how much robotics data OpenAI and others used in training, and Cornell and Google DeepMind researcher Yoav Artzi suspects large amounts of 3D training data such as Blender scenes may have played a role. Specialized models may still win on cost, reliability, speed and energy use, even if they generalize less well. Those are practical advantages that matter to anyone who would actually deploy a robot.

My view is that OpenAI should be judged on demonstrated physical performance, not simulated scores or the appeal of a unified model. The company's history here is instructive: it shut its robotics division in 2020, partly because it lacked training data, and rebuilt the team in 2024. Sam Altman is reportedly targeting a first public robot demo for 2027. That is a reasonable milestone, and the company should publish results on physical hardware, with failure rates, before claiming that general models make specialized robotics unnecessary.