OpenAI Disbands Its Robotics Team — and Bets GPT-6 Astra Can Run the Robots Instead
OpenAI has disbanded its standalone robotics team, Fortune reported in an exclusive on Friday 13 November 2026, and moved its people into other parts of the company. The robotics research team now sits inside OpenAI's main research organisation, reporting to Chief Research Officer Mark Chen, while the applied robotics and hardware teams have moved into the larger Applied organisation, which reports to President Greg Brockman. An OpenAI spokesperson confirmed the reorganisation to Fortune, and the company says robotics remains a core part of its long-term vision — this is a restructuring, not a retreat, and a humanoid robot demonstration is still targeted for 2027. The reason OpenAI gives is striking: general-purpose models such as GPT-6 Astra can now be applied to robots directly, so a separate team building robot-specific models is no longer the fastest way forward.
A general-purpose model is a frontier AI model trained on broad data — text, images, video, code — rather than a model trained specifically to move a robot arm. The bet is that the same model that answers questions and writes code can also look at a scene, work out what to pick up, and tell a robot how to do it, without a dedicated "robot brain" in between. If that bet is right, the robotics team does not need to exist as a separate unit; the model team is the robotics team.
Where the people went
The reorganisation splits the old team in two directions. Research goes to research; hardware goes to product. The robotics researchers — the people working on how models understand and control physical movement — now report to Mark Chen inside the main research organisation, alongside the teams building OpenAI's frontier models. The applied robotics and hardware staff — the people building and testing physical systems — move to the Applied organisation under Greg Brockman, which turns research into products.
| Group | New home | Reports to |
|---|---|---|
| Robotics research team | Main research organisation | Chief Research Officer Mark Chen |
| Applied robotics and hardware teams | Applied organisation | President Greg Brockman |
| Standalone robotics team | Disbanded as a unit | — |
What OpenAI has not said matters as much as what it has. The company has published no blog post about the change, and it has not said how many people were affected. There is no new robotics product, no new robot, and no new capability announcement attached to the move. Everything the public knows comes
![]() |
from Fortune's reporting and the company statement given to it.
The benchmark behind the bet
The evidence OpenAI is pointing to is a study run by Nvidia, cited by Fortune. Researchers used GPT-6 Astra as a robot controller on RoboLab-120, a simulated pick-and-place benchmark, with no special robot training at all. It scored 74.8% success; the best dedicated robot model, Lakeside, scored 51.0%. The gap — nearly 24 percentage points — is the number OpenAI's case rests on.
But the number needs its warning label. RoboLab-120 is a simulation, not a physical robot. The 74.8% figure is a success rate inside a simulated environment, on simulated manipulation tasks. It says nothing directly about how GPT-6 Astra would perform on a real robot arm in a real room, with real friction, lighting and objects that slip and break. Simulated benchmarks are useful for comparing approaches quickly, but they are not proof that a model can run a physical robot. Anyone repeating the 74.8% figure without the word "simulated" is overstating what was measured. The underlying paper, "An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond", frames the work as early evaluations — not a finished robot system.
Fortune also reports that OpenAI's researchers first assumed the best route to robotics was building on video models such as Sora, and later found that general frontier models could be applied directly to robot control. That shift is the real story: the team was built on the assumption that robots needed their own models, and the new evidence suggests they may not.
The second time OpenAI has done this
This is the second time OpenAI has closed a robotics team. The first team was wound down in 2020, when co-founder Wojciech Zaremba cited limited robot data and a more efficient path to general intelligence through other methods. OpenAI then rebuilt a standalone team in 2024 and announced a partnership with Figure to develop AI models for humanoid robots. Now, two years later, that team has been folded back into the wider company.
The history matters because it stops the simple readings in both directions. It is not true that OpenAI has "given up on robots" — the company says robotics remains core, it is still developing its own humanoid robot, and chief executive Sam Altman is targeting a first public demonstration in 2027, according to Fortune. But it is also not a routine reshuffle: a standalone team rebuilt with fanfare in 2024 has been dissolved after two years, because the model has overtaken the team. The first closure in 2020 was a retreat from robotics; this one is presented as the opposite — a claim that robotics no longer needs its own team because the general model has caught up.
![]() |
AI-generated illustration (generated 14 Nov 2026)
There was one telling detail in Fortune's reporting. The company had registered a booth at the Conference on Robot Learning (CoRL) in Austin, but booth 208 was empty when Fortune's reporter visited — the public-facing presence was booked before the decision, and nobody was left to staff it.
What it means for the robotics industry
Fortune frames the move as part of a bet gaining traction across the AI industry: general-purpose models will become the brains of robots, replacing the specialised vision-language-action models robotics startups have spent years building. If OpenAI is right, the value shifts from whoever builds the best robot-specific model to whoever builds the best general model — a contest OpenAI, Google and a handful of others are already fighting. It also puts pressure on well-funded humanoid startups whose pitch is a purpose-built robot brain: their moat looks thinner if a general model can do the job without robot-specific training.
But the bet is still unproven where it counts — on physical hardware. A simulated pick-and-place score is a long way from a humanoid robot working reliably in a home or warehouse, and OpenAI has shown no public physical demonstration to match the benchmark claim. The company's own timeline gives it until 2027 to show one. Until then, OpenAI has reorganised around a hypothesis, and the evidence for it is a simulation.
There are also open questions Fortune's reporting does not answer: how many researchers were affected, whether existing external research collaborations change, or how OpenAI will release or document its robotics work going forward. A robotics team inside the main research organisation may publish less about robots specifically, and more about general models that happen to work on robots.
Background: Claude Duet, GPT-6 Astra benchmark scores, OpenAI Decisions API, OpenAI offline training.
![]() |
AI-generated illustration (generated 14 Nov 2026)
Key Takeaways
- OpenAI disbanded its standalone robotics team on 13 Nov 2026, Fortune reported; an OpenAI spokesperson confirmed it. Researchers moved to Mark Chen's main research organisation, and applied robotics and hardware staff moved to Greg Brockman's Applied organisation.
- OpenAI says robotics remains core to its long-term vision and a humanoid robot demonstration is still targeted for 2027 — presented as acceleration, not a retreat.
- The trigger cited is an Nvidia-run study in which GPT-6 Astra scored 74.8% on RoboLab-120, against 51.0% for the best specialised robot model, Lakeside — but RoboLab-120 is a SIMULATED benchmark, not a test on physical robots.
- This is the second closure: OpenAI wound down its first robotics team in 2020, rebuilt one in 2024 and partnered with Figure, and has now folded it back into the wider company.
- OpenAI has published no blog post about the change and has given no headcount; no new robotics product was announced, and its CoRL conference booth in Austin stood empty.
Conclusion
OpenAI's robotics reorganisation is a bet stated in organisational form: the general model is the robot now, so the robot team can dissolve into the model team. The 74.8% simulated benchmark score is genuinely striking, but it is simulated, and a simulation is not a robot working in the real world. The test is the humanoid demonstration OpenAI says is coming in 2027. If a general-purpose model can run a physical robot reliably by then, the closure of the standalone team will look prescient. If it cannot, OpenAI will have dismantled the team it needs to find out why. Watch the demo, not the org chart.
Sources
- Fortune — OpenAI disbands robotics team as its general models start running robots (Wen Shao exclusive, 13 Nov 2026)
- explainx.ai — Summary and cross-check of Fortune's reporting on the robotics reorganisation (14 Nov 2026)
- dev.to — OpenAI Disbanded Its Robotics Team. GPT-6 Astra Is the Robot Now (14 Nov 2026)
- letsdatascience — OpenAI disbands standalone robotics team, folding staff into main research and Applied (14 Nov 2026)



Comments
Post a Comment