OpenAI Disbands Stand-Alone Robotics Team, Betting General Models Like GPT-6 Astra Can Drive Robots

OpenAI has dissolved its stand-alone robotics team and reassigned the researchers to other groups, a reorganization the company frames as an acceleration of robotics work rather than a retreat. Fortune reported the change on November 13, 2026, citing a company confirmation and a spokesperson who said robotics "remains a core part of OpenAI's long-term vision."

The move is a concrete test of a bet spreading across the AI industry: that general-purpose models trained on language, code, and video can also function as the control layer for physical machines, making dedicated robotics-model teams unnecessary. It is also a cost decision made under IPO pressure, and roboticists are split on whether it reflects genuine technical progress or a retreat from hardware.

What OpenAI actually changed

OpenAI eliminated the stand-alone robotics org and split its people between two existing organizations. The robotics research team now sits inside the company's main research organization and reports to Chief Research Officer Mark Chen. The applied robotics and hardware teams moved into the larger Applied organization, which reports to OpenAI President Greg Brockman.

The reorganization happened recently enough that OpenAI still had a booth reserved at CoRL, a major academic robotics conference held in Austin the same week. Fortune visited booth #208 and found it empty, with no OpenAI logo displayed. Former members of the robotics team were nonetheless seen at the conference and drew a crowd despite having no demo or display.

The stated rationale: general frontier models already run robots

OpenAI says its general-purpose frontier models can now be applied directly to robot control, which is why specialized robotics researchers were folded into general model teams. According to a company spokesperson, OpenAI researchers began the year assuming the best route to robotics progress was to build on models such as Sora, the video generation system that must implicitly model physical dynamics to produce realistic footage. Since then, the company says, general frontier models including GPT-6 Astra have demonstrated direct applicability to robotics.

GPT-6 Astra is the model named in Fortune's reporting as the one researchers connected to robots and observed performing robotic tasks with surprising competence. OpenAI has separately said it is developing its own humanoid robot, with CEO Sam Altman targeting a first public demonstration in 2027.

OpenAI has built and killed a robotics team before

OpenAI's first robotics team was wound down in 2020. Co-founder Wojciech Zaremba later attributed that decision to limited robot data and to a more efficient path toward artificial general intelligence through other methods, particularly training very large language models on text, which produced ChatGPT in late 2022.

OpenAI rebuilt a stand-alone robotics team in 2024 and announced a partnership with the startup Figure to develop AI models for humanoid robots. That partnership ended in February 2025, when Figure said it would develop its own robot AI.

The November 2026 reorganization is therefore the second time OpenAI has dissolved a dedicated robotics group, and the first time it has done so while explicitly claiming that general models can absorb the work.

The benchmark number behind the claim

Fortune cites an Nvidia study in which an Astra-powered agent reached a 74.8 percent success rate on a simulated manipulation benchmark. The best specialized robotic models on that leaderboard reported 51.0 percent on the same benchmark.

The setup: researchers gave Astra camera images, robot state information, and a control interface through which the model could command the robot. In the Nvidia study the robot was simulated rather than physical, so the result measures control reasoning in a digital environment, not real-world manipulation reliability.

Why roboticists are not fully convinced

Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, told Fortune that he reads the reorganization as a signal that OpenAI now sees robotics as "no longer a special problem" — a downstream application of general intelligence, in the same way the same frontier models now power specialized coding, legal, and financial tools.

Shi also described the limits. General models' improved spatial and reasoning capabilities let them tolerate a bottle placed in a different position or a change in lighting, and let them reason about the sequence of a task or identify which step failed. But large language models are slow and must be hosted in the cloud, which means the robot or device needs a persistent network connection. Shi still sees a need for smaller specialized models that run on a robot's onboard computer.

Russ Tedrake, an MIT professor and cofounder and CEO of Walden Robotics, argued that robotics companies retain an advantage the AI labs lack: robots already accumulating work experience in real deployments such as factory floors. That deployment experience and the data it generates help robots adapt to specific workplaces and tasks. Better foundation models make site-specific data more valuable, Tedrake said, because robots can learn from it faster.

Pete Florence, cofounder of the startup Generalist, which trains general-purpose models for robots, told Fortune that "There may be different starting points for how to grow intelligence," and that it is unclear any existing model will dominate robotics in the long run.

Hacker News commenters: latency, token cost, and hierarchy

The most substantive critique in the Hacker News discussion targeted control-loop latency rather than model capability. One commenter argued that general models can run robots but do so badly, and that a layered stack works better:

General models can run robots. They are very bad at it. It is much more efficient to have a stack of models with 'dumber' but faster models taking care of lower level functions and higher level ones that are much slower taking care of high level planning. ... Sensors and actuators require millisecond level responses, that's something you're not going to get out of an LLM on the other side of a communications link.

A related comment asked whether there is a primer for using LLMs to control robots at all, noting that streaming commands at 60 frames per second looks very token-inefficient. Another asked whether the change means OpenAI has given up on robots that are not cloud-connected.

Several commenters read the reorganization as financial rather than technical. One wrote that it "looks like general cost cutting to boost some numbers before IPO," another summarized it as "money is tight," and a third said OpenAI has realized that robots with actual hardware are hard and expensive. Another questioned the gap between the company's "long-term vision" framing and the absence of anything mid-term.

One commenter offered the sharpest general framing of the trend:

The Bitter Lesson comes for everything.

That references Richard Sutton's argument that general methods leveraging compute beat hand-engineered domain specialization over time. It is the strongest steelman for OpenAI's move, and the strongest reason to expect more reorganizations of this kind across the industry.

Other comments were skeptical of the human angle, with one calling it ironic that people working to make others disposable turned out to be disposable themselves, and another asking what the change implies for OpenAI's robot research office in Richmond, California. Fortune's reporting does not address that office.

The cost context: $10.9 billion in quarterly losses and a possible 2027 IPO

OpenAI is under pressure to cut spending and concentrate on a small number of core products. Altman told the company it needed to abandon "side quests," and OpenAI has already jettisoned some efforts, including Sora, to reallocate computing resources toward more central and potentially profitable work.

Fortune reports that OpenAI's operating losses totaled $10.9 billion in the most recent quarter, and that the company is seeking to close a new funding round while preparing for a possible IPO in early 2027. The robotics reorganization lands squarely inside that window.

What the reorganization implies for robotics startups

The claim that general models can control robots without specialized "spatial intelligence" models challenges the premise on which many robotics startups have raised billions: that robots need brains of their own.

Tedrake's counterargument is that the scarce asset in robotics is not the brain but the deployment: real robots in real workplaces generating task-specific data. Under that view, general models raise the value of a deployed fleet rather than replacing the companies that own one. Shi's counterpoint is complementary: cloud-hosted general models handle high-level reasoning, while compact on-device models handle fast control.

The practical reading is a hybrid architecture, not a wholesale replacement. OpenAI's reorganization argues that the high-level reasoning layer is commoditizing toward general frontier models. It does not establish that the millisecond-scale control layer is.

Open questions the reporting does not answer

Several details remain unreported as of the November 13, 2026 story. Fortune does not say how many researchers were reassigned, whether the reorganization involved departures, or what happens to OpenAI's stated humanoid robot program and Altman's 2027 demonstration target under the new structure. The 74.8 percent versus 51.0 percent comparison comes from a simulated benchmark, not physical robots. And the company's characterization of the move as acceleration is a statement of intent, not a measured outcome.

Sources