OpenAI no longer has a standalone robotics team. Fortune reported on November 13 that the team has been folded into the rest of the company, and OpenAI confirmed the change when Fortune asked about it. The robotics researchers now sit inside the main research group. The people who built and tested physical systems now sit on the product side of the company.
That is the whole announcement, and OpenAI made it in the quietest way available. The company has published no blog post and no statement of its own about the move. It has not said how many people were affected, and it has not described any layoffs. What exists is a confirmation given to a reporter, plus a dispute about the word used to describe it.
The Announcement
MadRobot, which followed up on the Fortune report, sets out where the staff went. The robotics research team now sits inside the main research team, reporting to Chief Research Officer Mark Chen. The applied robotics and hardware teams have moved into the larger Applied organisation, which reports to President Greg Brockman.
A company spokesperson told Fortune that robotics remains a core part of the long term vision at OpenAI, that the company is still committed to bringing general purpose robots into the world, and that moving the researchers is the best way to speed up that work. OpenAI then pushed back on the framing. The company told MadRobot that it reorganised the team to work more closely with research and product, and that the robotics programme and the roadmap are unchanged.
Staff said the same in public. Michelle Fradin, who works on robots at OpenAI, called the Fortune headline misleading and said robotics research had been merged into main research after Astra proved “a breakthrough in robotics intelligence”. Scott Cory, who works on robot intelligence there, posted a longer version of the same point.
astra and our newer models show much better generalization and scaling on robotics than other approaches so we are leaning into that and working more closely with the rest of research. same program, roadmap, goals.
The timing left one visible loose end. OpenAI had already booked a stand at CoRL, a leading robotics conference in Austin, Texas, this week. When Fortune visited booth 208, it was empty, with no OpenAI sign in sight, though former members of the team still drew a crowd there.
Important Details
| Item | Where it stands |
|---|---|
| Reported | November 13, 2026, by Fortune, reporter Wen Shao |
| Robotics research | Now inside main research, under Mark Chen |
| Applied robotics and hardware | Now inside Applied, under Greg Brockman |
| Statement from OpenAI itself | None published when checked on November 15 |
| People affected | Number not disclosed |
| Stated reason | General models such as GPT-6 Astra can be applied to robots directly |
| Humanoid plan | First public demo targeted for 2027, per Fortune |
Why OpenAI Changed Course
OpenAI told Fortune that at the start of 2026 its researchers assumed the best route to capable robots ran through models like Sora, its video generator, which has to grasp something of how the physical world works to make realistic clips. Since then, the company says, its general frontier models have turned out to work directly on robots. GPT-6 Astra, released in September, is the model named in that account.
Outside work supports part of that claim, and only part of it. On October 14, researchers at the Nvidia Seattle Robotics Lab published a study called Astranomy. They ran GPT-6 Astra as the controller for a simulated robot arm on RoboLab-120, a benchmark of 120 tabletop tasks in simulation, with no training done for robot work at all. The model read camera views and robot state, and it chose the movements.
The best setup in that study put Astra inside a persistent Codex session, with its full history kept between steps. That setup succeeded on 74.8 percent of tasks. The best model on the same leaderboard that was trained for robot work, a system called Lakeside, scored 51.0 percent. A simpler setup, where Astra answered each step on its own with no history kept, scored 66.3 percent. A full walkthrough of the Astranomy report shows that the gap between those two Astra scores came from how the model was called, not from any change to the model itself.
RoboLab-120 success rate, simulated tasks
Source: Nvidia Seattle Robotics Lab, Astranomy, October 14, 2026. Simulation only, with the simulation paused while the model responded. Lakeside is the strongest policy on the leaderboard that was trained for robot work.
The Catch In The Same Study
The study that helps OpenAI also prices the idea, and the price is high. The agentic setup that scored 74.8 percent used about 30.9 million tokens for each successful episode, at an estimated cost of about 36.40 dollars per episode. Even the direct setup took an average of 18.56 seconds for each model call. The simulation paused while Astra responded, so the test never asked the model to keep up with a room that keeps moving while it thinks.
That gap matters for physical work. A robot arm in a lab or a factory needs answers in a fraction of a second, not after a long pause, and it needs those answers at a cost a workplace can absorb. Guanya Shi, an assistant professor at the Robotics Institute at Carnegie Mellon University, told Fortune that the reorganization suggests OpenAI sees robotics as no longer a special problem, though he still sees a need for small models that run on the robot itself, with no constant connection to the cloud.
Other benchmarks point in a less tidy direction. Vector Wire, in an analysis of the move, notes that in the RoboDojo Sim framework Astra scored 29.8 percent, a result far below the RoboLab figure and below one third success. The two tests use different tasks and different setups, so neither number cancels the other. Together they show a field with no shared test that every lab accepts, where each side can point to a chart that favours its own approach.
The contrast arrived in the same week from the other direction. Runway released Praxis-1, a model built only for robot control, with weights that anyone can download. On RoboLab-120 it scored 46.2 percent, below Astra, but it was built to answer fast and run as a dedicated control model. OpenAI is betting that a dedicated model is a step the field can skip. Runway just shipped the opposite bet. The pattern is familiar from our report on Grok Bot using Mac apps directly, where the open question was also whether a general model should act in the world on its own, or whether the acting needs a layer built for the job.
- 2020OpenAI winds down its first robotics team, with cofounder Wojciech Zaremba later citing a shortage of robot data
- 2024OpenAI rebuilds a standalone robotics team and partners with humanoid maker Figure
- Feb. 2025Figure ends the partnership to build its own robot AI
- Oct. 14, 2026Nvidia Seattle Robotics Lab publishes Astranomy, its study of GPT-6 Astra as a robot controller
- Nov. 13, 2026Fortune reports the standalone team has been folded into research and Applied, and OpenAI confirms it
What’s Next
Confirmed steps are narrow. Robotics research now lives inside main research, hardware work now lives inside Applied, and OpenAI says the programme and the roadmap have not changed. Fortune reports that OpenAI is building its own humanoid robot, and that Chief Executive Sam Altman is targeting a first public demonstration in 2027. That target now has to be met by staff who sit in two different organisations.
Open questions are the larger part of the story. OpenAI has not disclosed how many people moved, what budget or computing power robotics now holds inside the larger groups, or whether its data collection work and its prototype projects change under the new structure. It has also not published results of its own from a physical robot under the new setup. Until it does, the strongest public evidence for the move remains a simulation study run by another company, with a pause button pressed every time the model needed to think.
Russ Tedrake, a professor at MIT who cofounded a robotics company, told Fortune that robot makers still hold one edge the large labs lack: their machines are already gathering experience on real factory floors, and better base models make that data more valuable, not less. That is the part of the argument OpenAI has not answered. A general model may well be the right brain for a robot. The robots that teach it what a real room feels like still belong, for now, to someone else.
Sources
- Fortune, exclusive by Wen Shao on the OpenAI robotics reorganization, November 13, 2026. The original report of the reorganization, the confirmation from OpenAI, the destinations under Mark Chen and Greg Brockman, the empty CoRL booth, the Sora account, and the 2027 humanoid target. Fortune is cited here by name and date; OpenAI has published no announcement of its own.
- MadRobot, OpenAI reorganises its robotics team as GPT-6 Astra runs robots, November 13, 2026, updated November 14. The statement OpenAI gave MadRobot disputing the word disbanded, and the public posts from Michelle Fradin and Scott Cory.
- Nvidia Seattle Robotics Lab, Astranomy study, October 14, 2026, as walked through in Astranomy: An Empirical Study of GPT-6 Astra, checked November 15, 2026. RoboLab-120 scores for the agentic, orchestrated and direct setups, the Lakeside score, token use, cost per episode, and the 18.56 second average call time. All figures are from simulation.
- Vector Wire, OpenAI Dissolves Robotics Team, Betting Foundation Models Suffice, November 2026, checked November 15, 2026. The RoboDojo Sim score of 29.8 percent for Astra and the argument that current data does not yet support the claim that dedicated robot models are no longer needed.
- Facts checked November 15, 2026. Every image in this article is a Dirty Static illustration in the house style. No image is a screenshot or a photograph of an OpenAI robot.
END OF TRANSMISSION
Stay curious. Question everything.
