OpenAI Disbands Robotics Team, Betting Its General AI Models Can Run Robots
The company has folded its standalone robotics group into its core research and applied-product teams, arguing that general frontier models such as GPT-6 Astra can already act as a robot's controller without dedicated engineering built around them.

OpenAI has eliminated the standalone robotics division it rebuilt only two years ago, splitting its researchers between the company's core research group and its applied-products team. The company told Fortune, which first reported the reorganisation on Friday, that the change reflects a bet that its general-purpose frontier models can now act as a robot's control system without a dedicated engineering group built around them.
Under the new structure, robotics research sits inside OpenAI's main research organisation, reporting to chief research officer Mark Chen, while applied robotics and hardware work has moved into the applied division run by president Greg Brockman. An OpenAI spokesperson confirmed the change and said robotics remains part of the company's long-term plans, framing the move as a way to accelerate rather than deprioritise the work. OpenAI has not published its own account of the reorganisation or said how many staff were affected.
A second dissolution in six years
The reorganisation is notable chiefly because it is the second time OpenAI has unwound a dedicated robotics group since 2020. The company shut its original robotics team that year, after cofounder Wojciech Zaremba concluded that physical robots simply did not generate enough data to train on, compared with text and code. OpenAI rebuilt the team in 2024 with a narrower goal: supplying AI models for other companies' robots rather than building its own hardware. That approach produced a partnership with the humanoid-robot startup Figure, which TechCrunch reported Figure itself ended in February 2025, with Figure's chief executive saying the company needed to build its own, more tightly integrated AI rather than rely on an outside model.
Since then, OpenAI has said it is developing its own humanoid robot in-house, with chief executive Sam Altman targeting a public demonstration in 2027. That hardware ambition appears intact; what has changed is the organisational chart underneath it. Robotics researchers no longer have their own reporting line, management structure or, for now, a presence at industry gatherings: Fortune's reporters found OpenAI's registered booth, number 208, empty and unbranded at the Conference on Robot Learning, the field's leading academic gathering, which the conference's own program confirms ran November 9 through 12 in Austin, Texas.
A 75 percent success rate, at a price
OpenAI's rationale rests on a shift in what its general models can do without any robot-specific retraining. A study published by Nvidia's Seattle Robotics Lab, which tested OpenAI's GPT-6 Astra model as a direct controller for simulated manipulation tasks, found that running Astra inside a persistent coding-agent session produced a 74.8 percent success rate on a 120-task pick-and-place benchmark called RoboLab-120. The best dedicated robotics model evaluated alongside it, called Lakeside, managed 51.0 percent, according to the Nvidia Research study, titled Astranomy.
That gap comes with real costs. The same study found that the configuration producing Astra's best results consumed roughly 30.9 million tokens per task episode, at an estimated cost of about $36.40, and took an average of more than 13 seconds between actions — far too slow for most physical robots operating in real time. Nvidia's broader case for the benchmark, laid out in a separate technical blog post from the company, argues that existing robotics benchmarks are too easily saturated or too statistically noisy to show whether a model is genuinely generalising rather than memorising, which is part of why the RoboLab platform behind the study was built with capability-tagged, frequently refreshed tasks rather than a fixed task list.
Researchers split on what it means
The reorganisation has drawn mixed reactions from academics who work on the same problem. Guanya Shi, an assistant professor at Carnegie Mellon's Robotics Institute, told Fortune the shift suggests robotics is "no longer a special problem" in AI research, though he added that he still expects a continuing role for small, specialised models that can run onboard a robot without a constant connection to the cloud. Pete Florence, a cofounder of the robotics startup Generalist, was more cautious, telling Fortune that "there may be different starting points for how to grow intelligence" and that it remains unclear which approach will ultimately dominate. Russ Tedrake, a robotics professor at MIT, argued that companies building their own specialised robot models still hold an advantage because they can gather real-world deployment data that a general-purpose lab like OpenAI does not have direct access to.
The stakes extend beyond academic debate. A wave of robotics startups have raised venture funding over the past two years on the premise that robots need purpose-built AI "brains" distinct from general-purpose chatbot models — Figure among them. OpenAI's move, even if it does not kill that market, challenges the assumption underlying it, and gives ammunition to investors weighing whether to back specialist robotics-model companies or wait for frontier labs like OpenAI, Google DeepMind and Nvidia itself to commoditise the layer those startups were built to own.
What happens next
OpenAI is not stepping back from robotics as a business line; if anything, the reorganisation suggests the opposite intent, folding the work more tightly into the teams building its flagship models rather than treating it as a side project. Fortune's report also noted that the company has been working to cut costs elsewhere, including shelving other initiatives, after reporting a large operating loss in its most recent quarter — a backdrop that may have made a dedicated, slower-moving robotics unit harder to justify organisationally even as the underlying technical goal continues.
The near-term test will be whether OpenAI's merged structure can turn benchmark results like Astra's into something that works outside a simulator. Nvidia's researchers were explicit that their findings describe simulated manipulation under laboratory conditions, not deployment on physical hardware in homes or warehouses, and the latency and cost figures they reported are well outside what most commercial robots could tolerate today. Altman's 2027 demonstration target gives outside observers a concrete date to judge the bet against. Until then, the clearest signal of how seriously rivals are taking the reorganisation may be less what OpenAI says about it than how quickly competing robotics companies feel the need to respond.

OpenAI Lets Outside Researchers Embed Inside Its Labs to Hunt for Rogue AI Behavior
Meta, Signal and Tailscale tell Canadian senators Bill C-22 would force them to build surveillance tools

Alibaba Launches Qwen4-Max, Betting Lower Prices Can Challenge US AI Leaders
