On November 13, 2026, Fortune reported in an exclusive that OpenAI has disbanded its stand-alone robotics team and moved the researchers to other parts of the company. OpenAI confirmed the reorganization when asked. Its reason: 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 best way to move fast.
This is the second time OpenAI has closed a robotics team. The first, in 2020 to 2021, was a retreat. This one, by OpenAI's account, is a bet that robots are now just another application of a frontier model. Here is what is reported, what the evidence is, and what it means for robotics builders.
TL;DR
| Question | Answer |
|---|---|
| What changed? | The stand-alone robotics team no longer exists as a unit. |
| Who confirmed it? | An OpenAI spokesperson, to Fortune. |
| Where did people go? | Researchers to Mark Chen's research org. Applied robotics and hardware to Greg Brockman's Applied org. |
| Stated reason? | GPT-6 Astra and other general models can be applied to robotics directly. |
| Is robotics abandoned? | OpenAI says no. A humanoid demo is still targeted for 2027, per Fortune. |
| Hard evidence? | An Nvidia study: 74.8 percent on a simulated manipulation benchmark, vs 51.0 percent for the best specialized models. |
What Fortune reported
Wen Shao's report says OpenAI reassigned its robotics research team to the main research team, reporting to Chief Research Officer Mark Chen. The applied robotics and hardware teams moved to the larger Applied organization, which reports to President Greg Brockman. A spokesperson portrayed the change as a sign of commitment to robotics, not a retreat.
The timing was visible. OpenAI had already registered booth #208 at CoRL, a top academic robotics conference in Austin, this week. Fortune says the booth was empty with no OpenAI logo, though a few former team members still drew a crowd.
The history: a third act for OpenAI robotics
The story has context. As Fortune recounts, OpenAI had a robotics team in its early days and wound it down in 2020. Cofounder Wojciech Zaremba later cited limited robot data and a more efficient path to AGI through other methods. The Robot Report covered that earlier decision when it happened.
OpenAI rebuilt a stand-alone team in 2024 and announced a partnership with Figure to build models for humanoid robots. That ended in February 2025, when Figure said it would develop its own robot AI. Our look at Figure's F.02 decommission shows where that company went next. OpenAI has since said it is building its own humanoid, with Sam Altman targeting a first public demo in 2027.
So this is the third structure in six years: no team, a stand-alone team, and now no stand-alone team with the work spread out.
Why OpenAI says it did this
The spokesperson told Fortune that at the start of the year researchers assumed the best path on robotics was to build on work from models such as Sora. A video model must hold some implicit picture of how the physical world works. Since then, the spokesperson said, general-purpose frontier models including GPT-6 Astra have shown they can be applied directly to robotics. That finding motivated moving the researchers onto teams building more general models.
Fortune adds business context. OpenAI is under pressure to cut costs and focus on core products. Altman told staff last year to drop "side quests," and Sora was among the efforts cut so compute could go to core work. Fortune reports operating losses of $10.9 billion in the most recent quarter and a possible IPO in early 2027. That is Fortune's reporting, and OpenAI did not tie the robotics move to costs in the statement quoted.
The evidence: general models on robots
The strongest data point in the piece comes from Nvidia. Researchers gave GPT-6 Astra camera images, robot state information and a control interface so it could command a robot. In the study, an Astra-powered agent reached a 74.8 percent success rate on a simulated manipulation benchmark. The best specialized robot models on that leaderboard reported 51.0 percent.
Two cautions apply. It was a simulation, not a physical robot. And one benchmark does not cover contact-rich real-world tasks. Still, other recent work points the same way. See our posts on a world-action model that compared itself with GPT-6 Astra on 60 tasks and a benchmark of robot value models. Those results show that general models do reasonably on some tasks and fail on others.
What outside experts say
Fortune quotes three robotics voices, and they do not fully agree.
- Guanya Shi, assistant professor at Carnegie Mellon's Robotics Institute, reads the move as a signal that OpenAI sees robotics as "no longer a special problem." He says general models can already cope when a bottle moves or lighting changes, and can reason about which step went wrong. He still sees a need for smaller specialized models that run on the robot without a constant network link, because large models are slow and cloud-hosted.
- Russ Tedrake, MIT professor and CEO of Walden Robotics, says robotics companies keep an edge: their robots already work in real places such as factory floors. That deployment data becomes more valuable as foundation models improve.
- Pete Florence, cofounder of Generalist, says it is unclear that any existing model will dominate robotics in the long run.
What this means for robotics startups and builders
Many robotics startups have raised money on the premise that robots need their own brains. OpenAI's move challenges that. If a general model plus a control interface works, specialized models face a harder case. The counter-argument from Tedrake is data from real deployments, and from Shi, latency and offline operation.
For builders, three practical points follow:
- Try the general model first. The Nvidia result suggests that a frontier model with images and a command interface is a strong baseline. Measure it before you train a custom policy. Tools such as an open robot-use agent paired with DeepSeek show this works zero-shot on some tasks.
- Keep a local fallback. Latency and connectivity limits are real. Plan for a small on-board model for reflexes and a cloud model for planning.
- Collect your own deployment data. It is the asset general models do not have. Tesla's Optimus production plans rest on the same logic of scale and field data.
How to test a general model on your own robot
You do not need an OpenAI-scale team to check the claim on a small task. The setup in the Nvidia study, as Fortune describes it, has three parts: camera images, robot state information, and a control interface that lets the model issue actions. A sensible way to copy the idea:
- Start in simulation. Pick one manipulation task, such as pick and place, and run a frontier model through a simulator. Count successes over at least 50 trials so the result is not luck. The Nvidia number is a success rate, so measure yours the same way.
- Expose a small action set. Give the model a handful of named commands (move to, grasp, release) rather than raw joint angles. Fewer choices cut errors and make logs readable.
- Log every step and every failure. Shi's point that models can reason about "which step went wrong" only helps if the model sees the failure. Feed back the last observation after each action.
- Perturb the scene. Move the object, change the lighting. Fortune notes general models handle these better than older policies, and this is where you will see whether that holds for your task.
- Time it. Record latency per decision. If the loop is too slow for your task, that is the case for a small on-board model next to the large one.
This is a test plan, not a result. Any success rate you measure will depend on the model version, the simulator and the task, so report them with the number.
Why a reorganization can matter as much as a model launch
Org charts tell you where a lab thinks value comes from. When a lab keeps a team separate, it signals a distinct problem with its own methods. When it folds the team in, it signals the problem is now solved by the shared stack. OpenAI had the same view about coding and about tools for legal and financial work, where the same frontier models now power specialized products. Robotics is being moved into the same category.
That does not mean specialized work disappears. Smaller models for on-board control, safety layers and hardware design remain. It does mean the center of gravity shifts toward the lab's main model line, and toward whoever has the best general model plus the best real-world data. For readers following the field, the signals to watch are the 2027 humanoid demo, the next Nvidia or academic benchmarks using general models, and whether other labs make the same structural move.
What is not known
- OpenAI has not published a blog post on the change. The source is Fortune plus a company statement given to it.
- Headcount, which people moved, and whether anyone left are not reported.
- The Nvidia study is cited by Fortune. The report does not link the paper, so we could not read the method.
- No date or spec for the 2027 humanoid demo was given beyond what Fortune attributes to Altman.
For the model behind the claim, see how to pick among OpenAI's GPT-6 models.
Related reading
- Figure F.02 decommission
- Tesla Optimus production starts: what is verified
- Liber-0 world action model vs GPT-6 Astra
- RoboValue benchmark for robot value models
- Open robot-use agents with DeepSeek
- How to pick a GPT-6 model
- Sources: Fortune exclusive, The Robot Report on the 2021 decision
Details are accurate as of November 13, 2026. Claims are as reported by Fortune and OpenAI's spokesperson.