OpenAI's massive robot bet: general frontier models will win over physical AI

OpenAI's massive robot bet: general frontier models will win over physical AI

John KoetsierNovember 16, 2026
ShareXLinkedInFacebook
For the second time in six years, OpenAI has disbanded its standalone robotics team, betting that general-purpose AI models can run robots better than any dedicated robotics stack. The early returns aren't promising, but if Astra grows fast enough, the company might be right.

OpenAI dissolved its standalone robotics team on November 13, 2026 for the second time in six years. The big bet: frontier models will zero-shot or one-shot robot training.

The research half of the group moved to the core research organization, now led by Chief Research Officer Mark Chen. The applied robotics and hardware staff shifted to the Applied organization under President Greg Brockman. No layoffs were confirmed, and OpenAI did not disclose headcount figures for the team.

Essentially, OpenAI believes its general-purpose frontier models, primarily GPT-6 Astra, can run robots without a dedicated robotics AI stack.

A spokesperson told Fortune that robotics "remains a core part of OpenAI's long-term vision" and framed the reorganization as acceleration rather than retreat.

This is a big bet on frontier foundation models

GPT-6 Astra, released in September 2026, is the linchpin of this strategy. An Nvidia study tested Astra as a robot controller on RoboLab-120, a simulated pick-and-place benchmark. Astra achieved a 74.8% success rate with no specialized robotics training data. The best dedicated robotics models on the same leaderboard hit only 51%.

That is a substantial gap ... in simulation.

Guanya Shi, an assistant professor at CMU's Robotics Institute, told Fortune he reads the reorganization as a signal that OpenAI sees robotics as "no longer a special problem" and instead as a downstream application of general intelligence. That flies in the face of where most in the industry are voting with their dollars: that physical AI is a special case and requires special training and models.

Apparently, CEO Sam Altman has set a target: a first public humanoid demonstration in 2027. That is the year he has previously described as when AI moves from the intellectual realm into the physical world.


Interested in the humanoid robot components and technology making all this possible? Humanoid Technology Expo is Europe's first conference dedicated to humanoid robot components, technology, and supply chains.


Simulation vs. the real world

The simulation numbers are impressive. The real-world numbers ... not so much.

At least so far.

In physical robot testing, Astra completed just 7 of 100 bimanual desk tasks. Each attempt cost $36.40. Latency ran 18.56 seconds per action. At those economics, commercial deployment is not a 2027 problem, or even a 2028 one.

The gap between simulated benchmarks and physical robot performance is a recurring theme across the robotics industry. It does not invalidate OpenAI's strategic thesis, but it does raise the question of timeline. Reorganizing toward general-purpose models might make sense if Astra gets substantially better over the next 12 to 24 months. That is a reasonable bet, given the pace of model improvement.

But the current numbers suggest a demo in 2027 would be just that: a demo.

Of course, Altman knows OpenAI's pace of progress, and if Astra or other frontier models increase their pace of development via recursive self-improvement ... maybe all bets are off.

A pattern of restarts ... is OpenAI right this time?

OpenAI first dissolved its robotics team in 2020, citing limited access to robot training data. The team was rebuilt in 2024, alongside a partnership with Figure AI.

That partnership ended in February 2025. Figure AI CEO Brett Adcock later said the collaboration had provided "very little" value beyond brand recognition. The split reportedly became inevitable once OpenAI signaled intent to build its own humanoid robots, turning partner into competitor.

In May 2026, OpenAI shelved plans to spin out its robotics and consumer hardware units, reportedly because they would still have consolidated on the company's financial statements regardless. The November dissolution completes that retreat from robotics as a standalone organizational bet.

What this means for the ecosystem

For the humanoid robotics industry, this move is both a threat and a validation.

The threat: if GPT-6 Astra or its successors can genuinely run robots as a general capability, the specialized robot AI market shrinks. Companies like Physical Intelligence, Covariant, and Skild AI all bet on the value of purpose-built robot foundation models. Others like Figure AI that are spending billions on physical AI will have wasted money.

OpenAI is betting against that thesis.

The validation: OpenAI is committing to robotics, just through a different door. The world's most prominent AI lab believes robots powered by general intelligence are coming. That is an enormous signal for the whole sector, regardless of the organizational chart.

The question is whether Astra will close the gap between 74.8% in simulation and 7% on a real desk before anyone else does.

Sources

Share this article

ShareXLinkedInFacebook

Written by

John Koetsier
John Koetsier

I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”