
OpenAI has disbanded its stand-alone robotics team. Again. Fortune reported on Friday, November 13, 2026, that the team is gone and its people reassigned — and OpenAI confirmed the change when asked. The company has published no blog post of its own about it, and has not said how many people were affected.
The reason OpenAI gives is not a retreat from robotics. It is a bet: general-purpose frontier models, above all GPT-6 Astra, have become capable enough to control robots directly, so a separate team building robot-specific models is no longer the fastest way forward. Whether that bet is right will decide a lot about what the robots of the next few years actually run on.
What Actually Changed
The reorganisation, as Fortune reported it, splits the old team in two. The robotics researchers now sit inside OpenAI’s main research organisation, reporting to Chief Research Officer Mark Chen. The applied robotics and hardware staff moved into the company’s larger Applied organisation, which reports to President Greg Brockman. As a standalone unit, the robotics team no longer exists.
An OpenAI spokesperson told Fortune that robotics remains core to the company’s long-term vision, and that the plan for a first public humanoid demonstration in 2027 still stands. The framing from the company is acceleration, not withdrawal — the same discovery, it says, that motivated the consolidation.
The timing gave the story an awkward visual. The reorganisation landed during the Conference on Robot Learning (CoRL) in Austin, the same event where NVIDIA had just unveiled GR00T 2. OpenAI had registered a booth at the conference, but Fortune found booth #208 empty when the floor opened.
The Bet: The Model Is the Robot Now
The core of OpenAI’s argument is GPT-6 Astra, the frontier model it released in September 2026. Instead of training a specialised model to understand a robot’s body and tasks, OpenAI says a general model can now be connected to a robot and simply told what to do — no robot-specific training required.

The evidence OpenAI points to is a study NVIDIA ran on October 14, 2026. 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 in the comparison, Lakeside, scored 51.0%. On that measure, the general model did not just keep up with the specialists — it beat them by a wide margin.
That result is the whole case for the reorganisation in one number. If a general model can already outperform a purpose-built robot brain on manipulation tasks, the argument goes, then building a separate robot brain is solving a problem that no longer exists.
The Catch: Slow, Expensive, and Still Simulated
The headline number comes with three caveats, and they matter more than the number itself. First, RoboLab-120 is a simulation, not a physical robot. A model that picks up virtual objects in a virtual room has not yet picked up anything real. Second, each attempt cost around $36.40 in tokens. Third, each model call took about 18.56 seconds. A robot that pauses for nearly nineteen seconds between decisions — and burns thirty-six dollars per try — is not a robot you can put on a factory floor.

Dedicated robot models exist precisely because they are fast and cheap per action. A small specialised model running on the robot itself can react in milliseconds, which is what manipulation demands: by the time a cloud model has finished thinking, the scene has changed. That is the same on-device logic behind llama.cpp 1.0’s stable API — the software that wins on a robot will be the software that can think where it stands, without phoning home.
There is also follow-on evidence starting to arrive on real hardware. A StationeryBench comparison pits GPT-6 Astra, driven by an Inspect Robots agent policy, against Ai2’s MolmoAct2 vision-language-action model on five bimanual desk tasks using real YAM robot arms. Real-hardware results like these — not simulations — are what will settle whether the general-model bet actually works.
This Has Happened Before
This is the second time OpenAI has closed a robotics team. The first was in 2020 and 2021, when the company shut down the team behind Dactyl, its dexterous robotic hand, and walked away from robotics research entirely. That closure was a retreat: the work was expensive, slow, and going nowhere fast enough.
The team was rebuilt in February 2025, when OpenAI announced it was getting back into robotics and building its own humanoid. Twenty-one months later, the team is gone again — but the company insists the meaning is different this time. In 2021, robotics was abandoned because the models were not good enough. In 2026, the team is being folded away because, OpenAI says, the models finally are.
Fortune also reported, in the same story, operating losses of $10.9 billion in OpenAI’s most recent quarter, alongside a possible IPO in early 2027. That context is worth noting — a company under that kind of financial pressure has every incentive to cut duplicated effort — but it is Fortune’s reporting, not OpenAI’s stated reason, and the two should not be confused.
Two Opposite Bets in the Same Week
What makes this week remarkable is that the industry is placing opposite bets at the same time. Two days before Fortune’s report, Runway released Praxis-1, an open-weight 27-billion-parameter robot brain that anyone can download and fine-tune — the dedicated-model bet, in its purest form. And at CoRL, NVIDIA unveiled GR00T 2, an open world-action model that learns dexterous skills from human video, claiming 95% success after five demonstrations.
OpenAI’s bet is the opposite of both. No robot-specific brain, no robot training data — just the general model, applied directly. If GPT-6 Astra keeps improving faster than the specialists, the separate robot model becomes a legacy idea. If the specialists close the gap on speed and cost, or if general models plateau on physical tasks, OpenAI will have dissolved the team it needs most. The week that also saw Jensen Huang resign as NVIDIA’s CEO after 33 years is, fittingly, a week in which the robotics roadmap itself is being rewritten.
The same week also showed how fast agentic AI is moving from chat into the physical world of money: ChatGPT Wallet now lets OpenAI’s agents pay for purchases on a user’s behalf, with approval every time. Robots are the next step in that same direction — AI that does not just answer, but acts.
What to Watch Next
The honest answer is the 2027 humanoid demo. OpenAI still targets a first public demonstration that year, and whether it arrives — and what model is running it — will say more than any reorganisation announcement. If the demo runs on GPT-6 Astra or its successor, the bet is vindicated in public. If it slips, or quietly runs on something else, the reorganisation will look very different in hindsight.
Before then, watch two numbers: latency and cost per action. General models need to get dramatically faster and cheaper before they can control a robot in real time, and the StationeryBench-style real-hardware benchmarks will show whether the simulation advantage survives contact with physical arms. Watch, too, whether other labs follow. If a second frontier lab dissolves its robotics team, the general-model era has arrived. If they hire instead, OpenAI may have just made the most expensive org-chart change in robotics.
Frequently Asked Questions
Did OpenAI quit robotics?
No. OpenAI disbanded its stand-alone robotics team and moved the researchers into its main research organisation and the hardware staff into its Applied organisation. A spokesperson told Fortune that robotics remains core to the company’s long-term vision, and the target of a first public humanoid demonstration in 2027 still stands.
Why did OpenAI disband its robotics team?
OpenAI’s stated reason is that general-purpose frontier models such as GPT-6 Astra can now be applied directly to robots, so a separate team building robot-specific models is no longer the best way to move fast. The company frames the change as acceleration, not retreat.
What evidence supports using a general model to control a robot?
An NVIDIA study from October 14, 2026, used GPT-6 Astra as a robot controller on RoboLab-120, a simulated pick-and-place benchmark, with no special robot training. It scored 74.8% success, against 51.0% for the best dedicated robot model in the comparison. The benchmark was a simulation, and each attempt cost about $36.40 in tokens with roughly 18.56 seconds per model call.
Has OpenAI closed a robotics team before?
Yes. OpenAI closed its first robotics team in 2020–2021, after work on the Dactyl dexterous hand, and walked away from robotics research. It relaunched a robotics team in February 2025. The November 2026 closure is the second time the stand-alone team has been disbanded.
What are rival labs doing instead?
The opposite. Runway released Praxis-1, an open-weight 27-billion-parameter robot brain, two days before Fortune’s report, and NVIDIA unveiled GR00T 2, an open world-action model that learns manipulation from human video, at CoRL on November 9, 2026. Both bet on dedicated robot models; OpenAI is betting the general model makes them unnecessary.
Sources: ExplainX — OpenAI Disbands Robotics Team: GPT-6 Astra Bet (2026) · DigiTechBytes — Does Robotics Need Its Own AI? · Dev.to — OpenAI Disbanded Its Robotics Team. GPT-6 Astra Is the Robot Now

