OpenAI folds its robotics team into research after GPT-6 Astra starts driving robots directly

OpenAI has disbanded its stand-alone robotics team for the second time in its history, Fortune reported on November 13, and the company confirmed the change when asked. The robotics researchers now sit inside the main research organization under Chief Research Officer Mark Chen. The applied robotics and hardware teams moved into the Applied organization that reports to President Greg Brockman.

The stated reason is the interesting part. OpenAI’s position is that its general frontier models, GPT-6 Astra among them, can now be pointed at a robot directly, so a separate group building robot-specific models is no longer the fastest route. An OpenAI spokesperson told Fortune that robotics remains core to the company’s long-term vision and that the reorganization is meant to speed the work up. Fortune’s reporter found one visible sign of the change at CoRL, the robot learning conference in Austin this week: OpenAI’s registered booth, number 208, stood empty with no logo and no demo.

What changed inside OpenAI’s research bet

According to Fortune, OpenAI’s robotics researchers started the year assuming the best foundation for robots was a video model such as Sora, on the theory that a model trained to predict video has to learn some implicit physics. That assumption did not survive the year. Researchers gave Astra camera images, the robot’s state and a control interface, and found the general model could command the robot for some tasks with no robot-specific training.

The one number in the report comes from an Nvidia study, which Fortune cites without naming the benchmark: an agent built on Astra reached a 74.8% success rate on a simulated manipulation leaderboard where the best specialized robot models scored 51.0%. Two limits on that figure matter. It is simulation only, and Fortune does not link the study, so the task set, the control frequency and how much scaffolding sat around the model are all unknown. A 24-point lead in simulation tells you a general model can plan and sequence manipulation steps. It does not tell you it can close a control loop on real hardware at speed.

Astra is the same model whose behavior under evaluation we covered in our report on its safety research.

The second time OpenAI has shut this team

OpenAI had a robotics group in its early years, known for a robot hand that solved a Rubik’s Cube. Fortune dates the wind-down of that team to 2020; cofounder Wojciech Zaremba later explained that robot data was too scarce and that large language models looked like a more efficient path to general intelligence.

The company rebuilt a stand-alone team in 2024 alongside a partnership with the humanoid startup Figure. Figure ended that partnership in February 2025, saying it would build its own robot AI. Fortune reports that OpenAI is still developing its own humanoid robot and that Sam Altman is targeting a first public demonstration in 2027, so the hardware ambition has not been cancelled. What has changed is where the intelligence for it is expected to come from.

The two shutdowns rest on opposite readings of the same idea. In 2020 the argument was that language models were the better road to general intelligence and robots were a detour. In 2026 the argument is that the road led back to robots: the general model arrived, and it turned out to be usable as the robot’s brain.

Cost pressure is part of the story

Fortune also places the move in a financial context. Altman told staff last year to drop “side quests”, and OpenAI has since dropped projects including Sora, which we covered in an earlier post. Fortune puts the company’s most recent quarterly operating loss at $10.9 billion and says it is seeking a new funding round and preparing for a possible IPO in early 2027.

Both explanations can be true at once. A research finding that the flagship model transfers to robots is also a convenient reason to stop funding a parallel model line. OpenAI has not published the internal results that drove the decision, so outsiders cannot weigh how much of it was evidence and how much was budget.

Why robotics researchers are not conceding

Fortune quotes three researchers who each accept part of OpenAI’s premise and reject part of it.

Guanya Shi of Carnegie Mellon’s Robotics Institute reads the move as a sign that robotics is becoming a downstream application of general intelligence, like coding tools. General models now cope with changes in object position and lighting and can reason through task steps. He still expects smaller specialized models to run on the robot itself, because a cloud-hosted LLM is slow and needs a constant connection. That is the practical gap in the “general model runs the robot” story: frontier models respond in hundreds of milliseconds at best, while stable manipulation and balance need control updates many times per second, on hardware that cannot lose its network link mid-motion.

Russ Tedrake, the MIT professor who leads Walden Robotics, argues that deployment data is where robotics companies keep their edge. Better foundation models make site-specific experience from a factory floor more valuable, not less. Pete Florence of Generalist, which trains general-purpose robot models, says it is not clear any existing model will dominate robotics in the long run.

The stakes for that side are large. Startups have raised billions of dollars on the premise that robots need their own purpose-built brains, trained on robot data. If a general model with a camera feed and a control interface is competitive, that premise weakens. Google is running the opposite experiment in public, with a dedicated robotics model line that we covered in our post on Gemini Robotics 2.

What to watch

• Real hardware results. A published OpenAI or Nvidia evaluation of Astra on physical robots, with latency numbers, would settle more than the simulated leaderboard does.
• The humanoid demo. If the 2027 demonstration happens, whether the robot runs a frontier model in the cloud, a distilled on-device model or a mix will show how far the “no special model” claim really goes.
• Who leaves. Reorganizations like this often push specialists toward dedicated robotics companies. Fortune did not report departures or the team’s size.
• Funding for specialized robot models. The next rounds for companies training robot-specific foundation models will show whether investors accept OpenAI’s reading.

This remains a single-outlet report resting on an unnamed spokesperson, and OpenAI has published nothing of its own on the change.