OpenAI has dissolved its standalone robotics team, folding researchers into the core research division under Chief Research Officer Mark Chen and moving applied robotics and hardware staff into the Applied organization under President Greg Brockman3. The company's stated rationale is that general-purpose models such as GPT-6 Astra have demonstrated the ability to directly control robots, making a dedicated robotics unit redundant1,2. The thesis OpenAI is advancing is stark: foundation models are now good enough that robotics is an application layer, not a research frontier. The strongest counter-reading — that folding robotics researchers into Mark Chen's core division deepens integration rather than abandoning the field — deserves weight. But the early performance data suggests the thesis is running well ahead of the engineering reality, and the organizational signal is unambiguous: there is no longer a team whose job is to solve robotics-specific problems.
The reorganization surfaced publicly at CoRL. At the Conference on Robot Learning (CoRL) in Austin on November 13, booth 208, reserved by OpenAI, stood empty: the company's name was still printed on the floor plan, but no signage or product displays could be found on site. OpenAI confirmed the restructuring to Fortune, with a spokesperson portraying it as a sign of commitment to robotics rather than a retreat. The timing was recent enough that OpenAI had already registered the booth.
The company's argument rests on GPT-6 Astra's capabilities. OpenAI said the model can be directly applied to robotics. Unitree's founder Wang Xingxing offered partial corroboration: at the Shanghai International Industry Fair on October 12, Unitree connected GPT-6 to various robots to perform tasks, and Wang Xingxing said the results exceeded expectations. At the World Physical AI Conference on October 22, Unitree demonstrated a robot using GPT-6 to complete a painting, and Wang Xingxing said the team did not specially train for the painting task or write complex motion programs.
But the same body of evidence that supports the claim also exposes its limits. In six real-machine and simulation tests built around GPT-6 Astra, the model failed to directly grasp a Rubik's cube, simulate carrying a basket, or directly stack bowls. In Anthropic's RoboDojo Sim framework, GPT-6 Astra scored 29.8%, compared with Claude Fable 5.2's previous generation at 27.6%. Both scores sit below one-third success — a range difficult to characterize as production-ready for physical tasks where failure means dropped parts or damaged hardware. A fair objection: these are zero-shot numbers, and the dedicated-model results from DeepMind (96%) and Nvidia (95%) reflect task-specific fine-tuning, demonstration data, and in some cases real-machine reinforcement learning. The comparison is not apples-to-apples on methodology. But it is apples-to-apples on the question OpenAI's restructuring poses: can a general model, applied directly, match what dedicated pipelines deliver? At 29.8% versus 94–96%, the answer today is no.
The inference economics compound the problem. In Nvidia's research, the GPT-6 Astra agent approach consumed about 30.9 million tokens per successful task episode, at an estimated API cost of $36.40 per episode. The direct-control approach took an average of 18.56 seconds per inference. An 18-second inference loop is roughly three orders of magnitude slower than the millisecond-scale control cycles industrial robotics demands. And $36.40 per successful task episode is a cost structure no factory floor can absorb at scale.
UBTech's testing illustrates the gap. UBTech tested both general-purpose and dedicated approaches, using a coordinated system where Thinker-VLA executed operations and Thinker-WM predicted action outcomes. With a dedicated embodied model, the robot completed grasping a lucky bag, moving around obstacles, and stacking bowls. In the bowl-stacking test specifically, Astra's token consumption fell by 83.6% and estimated API costs fell by 74.2% when paired with a dedicated model. The industry is converging on hybrid architectures where foundation models handle perception and planning while dedicated models handle control — not on the pure foundation-model approach OpenAI's restructuring implies, nor on the pure dedicated approach either. OpenAI's organizational bet is that the dedicated-control layer will eventually collapse into the foundation model. The 29.8% RoboDojo score and the 83.6% token-cost reduction UBTech achieved by adding a dedicated model suggest that collapse is not imminent.
Wang Xingxing offered a useful taxonomy of the difficulty gradient. He said current top multimodal models already possess strong understanding and planning capabilities, but noted that grabbing a box can tolerate positional deviation while inserting a connector requires handling angle, friction, and contact force. He said parts insertion can get stuck in the last few millimeters. The distinction matters because the tasks where GPT-6 Astra has shown promise, such as painting and object grasping, sit on the tolerant end of that spectrum, while the precision tasks that define manufacturing value remain out of reach for direct general-model control.
The competitive landscape reinforces the point. Google DeepMind released the Gemini Robotics 2.5 vision-language-action model on November 10. DeepMind published parts-sorting results from Apollo 2 humanoid robot tests at Mercedes-Benz plants in Sindelfingen, Germany, and Kecskemét, Hungary, achieving a 96% success rate. Apptronik CEO Jeff Cardenas said Apollo 2 showed marked improvement after several weeks of training and had only a small performance drop after migrating to the Mercedes production line. Nvidia released GR00T 2 in November; across 50 tasks for which the target robot had not previously been trained, GR00T 2 had a zero-shot average success rate of 40%, but after five demonstration operations, less than three minutes of robot data per task, and about 30 minutes of fine-tuning on a single RTX 5090, its success rate rose to 95%. Nvidia also used real-machine reinforcement learning to bring charger-plugging success to 96% and cable-insertion success to 94%.
OpenAI's organizational history with robotics adds context. The company's partnership with Figure ended in February 2025. OpenAI was recruiting robotics talent at salaries of up to $500,000 as recently as two months before the dissolution. OpenAI closed an $11 billion funding round on November 6, and Yen Chi departed OpenAI on the same date. The speed of the reversal, from active high-salary recruitment to team dissolution in roughly two months, suggests either a rapid internal demonstration that changed the calculus or a resource-allocation decision driven by other pressures. The capital runway makes a resource-scarcity explanation implausible, but it also means OpenAI could afford to maintain a dedicated team and chose not to. The company has not disclosed its post-restructuring budget, headcount, or computing-power allocation for robotics.
The broader robotics funding environment remains active. Walden Robotics raised $300 million from Toyota and Deviation Capital in July. Unitree went public at a $5.9 billion valuation in August. Generalist raised $200 million from 8VC in August. Capital continues to flow into companies building dedicated robotics capabilities, not away from them — all three rounds closed in the three months before OpenAI's restructuring.
What to watch: whether OpenAI publishes robotics benchmarks from GPT-6 Astra that match or exceed the 94–96% success rates DeepMind and Nvidia have demonstrated with dedicated models on precision tasks such as cable insertion. The company's post-restructuring headcount and compute allocation for robotics remain undisclosed. The inference-cost trajectory matters: at $36.40 per successful task episode, the economics do not work for manufacturing. The next marker is whether any OpenAI robotics research appears at ICRA 2027. Two talent signals bear watching: Yen Chi's departure from OpenAI on November 6 and Zhu Mingxuan's departure from Figure in September. Where those researchers land will say more about the field's center of gravity than any press release.