OpenAI has dissolved its stand-alone robotics team, betting that its general-purpose frontier models can now control robots directly — and making a dedicated robotics group redundant.

The company confirmed the reorganization to Fortune. A spokesperson framed it not as a retreat but as commitment: robotics remains, in the company's words, a core part of OpenAI's long-term vision, and reshuffling researchers onto general AI teams is the fastest route to general-purpose robots.

The mechanics are straightforward. Robotics research now reports to Chief Research Officer Mark Chen inside the main research organization. The applied robotics and hardware teams moved into the larger Applied organization under President Greg Brockman.

The timing was awkward. OpenAI had already registered a booth at CoRL, a top academic robotics conference held this week in Austin. When Fortune visited booth #208, it sat empty. No logo, no demo, no display.

Why kill a team you rebuilt two years ago?

The history matters here. OpenAI ran a robotics team in its early years, then wound it down in 2020. Cofounder Wojciech Zaremba later cited two reasons: limited robot data, and a more efficient path to artificial general intelligence — the human-like AI OpenAI has chased since its founding — through other methods. That bet produced ChatGPT in late 2022.

In 2024, OpenAI rebuilt a stand-alone robotics group and announced a partnership with startup Figure to build AI models for humanoid robots. Figure walked away in February 2025, saying it would develop its own robot AI. OpenAI says it is now building its own humanoid robot, with CEO Sam Altman targeting a first public demonstration in 2027.

The company's reasoning for the latest move has shifted since the start of the year, the spokesperson said. Researchers initially assumed the best robotics path ran through Sora, the video generation model that must implicitly understand physics to render realistic video. Since then, general-purpose frontier models — including GPT-6 Astra — have shown they can be applied to robotics directly. That discovery motivated the disbanding.

What does the evidence look like?

The numbers are striking. In an Nvidia study, an Astra-powered agent hit a 74.8% success rate on a simulated manipulation benchmark — a test where the model commands a simulated robot rather than a real one. The best specialized robotic models on the same leaderboard reported 51.0%.

Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, said he interprets the reorganization as a signal that OpenAI sees robotics as "no longer a special problem." Instead, robotics becomes a downstream application of general intelligence — much as the same frontier models now power specialized tools for coding, legal, and financial work.

Shi said general models' improved spatial and reasoning capabilities let them handle tasks that previously required dedicated training. They cope when a bottle moves or lighting changes, and they can reason about task sequences or diagnose which step went wrong.

What does this mean for robotics startups?

Plenty. The thinking challenges the premise behind many robotics startups, which have raised billions on the assumption that robots need purpose-built "spatial intelligence" models of their own.

Not everyone is convinced the generalists win. Russ Tedrake, an MIT professor and cofounder and CEO of Walden Robotics, said robotics companies retain one edge: their machines are already gaining work experience in real places like factory floors. That deployment data helps robots adapt to specific workplaces, and better foundation models make that data more valuable because robots learn from it faster.

Shi also flagged a hardware constraint. Large language models are powerful but slow and cloud-hosted, meaning a robot needs a constant network connection. He still sees demand for smaller, specialized models running on a robot's onboard computer.

"There may be different starting points for how to grow intelligence," said Pete Florence, cofounder of Generalist, a startup training general-purpose AI models for robots. He said it's unclear any existing model will dominate robotics long-term.

Is cost pressure part of the story?

Almost certainly. The reshuffle lands as OpenAI works to rein in costs and focus on a few core products. Last year Altman told staff to abandon "side quests," and the company has jettisoned efforts including Sora to reallocate computing resources to more profitable endeavors. OpenAI's operating losses reportedly totaled $10.9 billion in the most recent quarter, even as it seeks to close a new funding round and prepares for a possible IPO in early 2027.

Back at CoRL, a few former OpenAI robotics researchers still drew a crowd at booth #208 — no display, no demo. The conference's hottest question was how much robotics expertise stays valuable as general models improve. Their redeployment offered one answer.