The problem
On 13 November Fortune reported that OpenAI had disbanded its stand-alone robotics team, and OpenAI confirmed the reorganization. The robotics researchers now sit in the main research organization under Chief Research Officer Mark Chen. The applied robotics and hardware teams moved into the Applied organization under President Greg Brockman. According to Fortune’s paraphrase, a spokesperson said that robotics remains a core part of OpenAI’s long-term vision, and gave the reason: at the start of the year its researchers thought the best path to robotics ran through Sora-style world models, but since then general frontier models, including GPT-6 Astra, have shown they can be applied to robots directly.
The change is recent enough that OpenAI’s booth, #208 at the CoRL robotics conference in Austin that week, sat empty when Fortune visited.
One widely shared post read it as a funeral. It said “OpenAI has officially disbanded its robotics team, just five months after relaunching it”, then “it’s over”, and passed a million impressions. OpenAI staff pushed back the same afternoon. Engineer David Schnurr wrote “robotics at openai is alive and well.” and pointed to 15 open roles. Welton Wang, whose profile reads “sora, robot researcher @ openai, cs @ stanford”, replied: “this is completely wrong (read the article), we are investing more than ever in robotics.”
Both readings argue about whether OpenAI is quitting robots. Fortune’s article doesn’t say it is, and the staff say it isn’t. The more useful question is what this org chart bets on, and what it gives up.
What I concluded
On Fortune’s account, I read this as OpenAI cutting its robotics program along the exact seam it had described as the program’s foundation.
On 31 May Sam Altman wrote: “Our world simulation research program, led by Aditya Ramesh (@model_mechanic), has evolved over the past year into OpenAI Robotics. Progress is rapid, and based on a foundation of co-design between robotics hardware and ML research.” Five and a half months later, on Fortune’s account, the ML research reports to one executive and the hardware reports to another.
That is a coherent bet, not a retreat. If a general model really can drive a robot, then robotics becomes a downstream application of that model, the way coding or legal work already is. Guanya Shi of Carnegie Mellon told Fortune he reads the move as OpenAI seeing robotics as “no longer a special problem.” In that world, a separate robot-brain team mostly duplicates the general model and competes with it for compute.
The cost is that co-design used to be one team’s job, and now it is an interface between two reporting lines. Whatever a robot needs that the general model doesn’t care about (speed, running offline, contact-rich control, uptime on a real floor) no longer has a dedicated owner on the research side. My read: the reorg is right if OpenAI makes robot tasks part of how the general model is judged, and wrong if robotics becomes a customer that files requests with a research org busy on everything else.
Constraints
Four things make this harder than “the model is good now, so fold the team in.”
Speed and connectivity. Shi also told Fortune that large language models are powerful but slow, run in the cloud, and need a constant network connection. He still sees a need for smaller specialised models on the robot’s onboard computer. A model that’s good at a task but has to phone home and wait isn’t yet a controller.
Field data lives with whoever has robots deployed. Russ Tedrake (MIT; cofounder and CEO of Walden Robotics) told Fortune that robotics companies have one advantage the labs don’t: their robots are already gaining work experience in real places, and better foundation models make that site-specific data more valuable, because robots learn from it faster. OpenAI’s own robots are not yet in public; Fortune reports Altman is targeting a first public demonstration of its humanoid in 2027.
Physical engineering still bites. OpenAI’s own 3 November post is the best evidence on this point. The Astra example there is a device users can use to talk to Codex, not a robot. Two team members designed that test device with Astra in under a week, as a side project, and a contract manufacturer built a small batch of prototypes in under a month. OpenAI writes: “Getting the device ready for manufacturing still required several rounds of review and iteration. Astra did not consistently account for the manufacturer’s production limits, tolerances, or details required for assembly. These gaps are helping us identify what future models need to do better.” My judgement: Fortune says the hardware team now sits in Applied, but this post does not say which group built the device. Robot hardware runs into the same kind of manufacturing detail, so I read the gap as a co-design problem.
Money and focus. Fortune notes the reorganization comes as OpenAI is under pressure to cut costs and focus on a few core products. It cites Altman’s earlier call to drop side quests, Sora being dropped, operating losses that reportedly reached $10.9 billion last quarter, and a possible IPO in early 2027. Those are reported figures, not OpenAI disclosures. Fortune presents this as context, not as OpenAI’s stated reason (OpenAI’s public rationale is capability). I read the cost pressure as making a separate robot-model effort harder to defend.
The options, and what I traded away
The decision: how to organise robotics inside a lab whose main product is a general model.
| Option | What it buys | What it costs | Why I did / did not pick it |
|---|---|---|---|
| Keep a stand-alone robotics team with its own model (the May setup) | One owner from motor to model; co-design inside one team; robot needs set the research agenda | Duplicates the general model; competes for compute; hard to defend under a budget that drops side quests | Would not pick it if the general model already transfers. OpenAI says it does |
| Fold research into the general model, hardware into Applied (what OpenAI did) | Every general-model gain carries over to robots; no second model to maintain | Co-design becomes a handoff between two orgs; robot-specific needs compete with every other priority in research | Would pick it, with conditions (below) |
| Exit hardware and license the model to robot makers | No factory, no hardware burn; robot companies bring the field data | OpenAI has tried a version of this. In February 2025 Adcock announced Figure was leaving the collaboration agreement, citing a breakthrough in “fully end-to-end robot AI, built entirely in-house” | Would not pick it. My speculation, which neither Adcock’s post nor Fortune states: a partner with its own plans could leave and take the field data with it |
What I trade away by picking the middle row: a single person accountable for whether the robot works end to end. The conditions exist to buy that back.
Evidence
What is solid: OpenAI confirmed the reorganization to Fortune, including both destinations and both executives. Altman’s May post is primary and names co-design as the foundation. OpenAI’s 3 November post is primary and documents that its hardware and infrastructure teams use Codex and ChatGPT for CAD, drawings, simulations and design reviews across robotics and compute infrastructure, with hardware-team token use up 23-fold since February.
What is disputed or thin: the scale. Fortune gives no headcount, and nobody has said how many people moved where. Schnurr says there are 15 open roles. On 17 November the OpenAI public job list (https://jobs.ashbyhq.com/openai) had exactly 15 postings whose team and department are Robotics. None is an ML research role. Eleven are mechanical, electrical, firmware, manufacturing, or test. The other four are a control-systems software role, a data-collection and evaluation software role, an actuators technical program manager, and a robotics safety engineer. The control-systems posting says the hire will work with machine learning engineers to implement reinforcement learning models; the role itself is real-time control, not a research seat. Wang’s post responds to the “it’s over” framing, not to Fortune’s facts, and he tells readers to read the article.
What I left out on purpose: success rates from simulated benchmarks of general models controlling robots. Simulation scores say little about the constraints above, and I would rather not lean on them.
Counter-arguments worth taking seriously. “Teams and reporting lines don’t matter; the same people do the same work.” Sometimes true, and the CoRL scene supports it: Fortune saw former team members at the empty booth still drawing a crowd. But reporting lines decide whose roadmap wins when compute is tight, which is exactly the situation Fortune describes. “Robots are just another app for the model.” Possibly, but Pete Florence of Generalist, which trains general-purpose models for robots, told Fortune: “There may be different starting points for how to grow intelligence.” He also said it was unclear any existing model will dominate robotics.
Where this stops being true
The argument assumes Fortune’s org chart describes how work actually happens. If research and Applied run a joint robotics group in practice, the seam I’m worried about may not exist. It also weakens if the general model gets fast and small enough to run on the robot, because then the speed constraint goes away. And it doesn’t apply to robot companies with deployed fleets: for them, Tedrake’s point cuts the other way, and their field data becomes more valuable as general models improve. I am not claiming OpenAI is leaving robotics, or that the reorg will fail.
What I would do differently
If I ran robotics inside OpenAI’s new structure:
Name one owner across both orgs for the robot’s closed-loop requirements: response time, behaviour when the network drops, and recovery after a failed grasp.
Put a fixed set of real-robot tasks into the general model’s release evaluation, so a robotics regression can hold up a model release the way a coding regression would.
Keep a small hardware-and-research pod on the 2027 demo itself, with a dated acceptance test written now.
If I were a robot company watching this, I would keep collecting deployment data, and I would not plan on a cloud general model as my controller until its maker publishes real-robot results with latency and offline behaviour.
What I’m watching: whether the next OpenAI model release reports any robot evaluation, whether those 15 robotics roles stay open, and whether the 2027 demonstration date holds.