OpenAI Drops Its Robot Team as Runway Goes Open

Minimal geometric illustration of a robot arm and a puzzle piece, with the words "NO MORE ROBOT TEAM."

OpenAI has disbanded its stand-alone robotics team. The news, reported by Fortune on 13 November 2026 and confirmed by OpenAI when asked, is the second time the company has closed a robotics unit — and this time it says the reason is success, not failure. General-purpose AI models, it argues, can now run robots directly.

Two days earlier, Runway made the opposite bet, releasing Praxis-1 — a robot model built specifically for controlling machines, with its weights openly available. The two moves frame the biggest open question in robotics: do robots need their own AI, or will one general model do the job?

What OpenAI actually did

Under the reorganisation, OpenAI's robotics researchers now sit inside the company's main research team, reporting to Chief Research Officer Mark Chen. The applied robotics and hardware teams have moved into the larger Applied organisation, under President Greg Brockman. The stand-alone team no longer exists, and its people were reassigned rather than let go.

OpenAI has published no announcement. Fortune reported its booth at the CoRL robotics conference in Austin was empty when visited. A spokesperson told Fortune the change reflects a new approach to robotics, not a withdrawal — and that robotics remains a core part of its long-term vision.

The company also says it is developing its own humanoid robot, with chief executive Sam Altman targeting a first public demo in 2027.

Why OpenAI says general models are enough

At the start of the year, OpenAI's researchers assumed the route to robotics ran through video models like Sora. Since then, the spokesperson said, general-purpose frontier models including GPT-6 Astra have shown they can be applied to robotics directly. That finding prompted the change.

The strongest data point comes from an Nvidia study. Researchers gave GPT-6 Astra camera images, robot state information and a control interface so it could command a robot. An Astra-powered agent reached a 74.8 percent success rate on a simulated manipulation benchmark, RoboLab-120, against 51.0 percent for the best specialised robot models.

Two cautions belong with that number. The result was in simulation, not on physical robots, and the study — published 14 October 2026 as a CoRL late-breaking result — was cited by Fortune without a link to the paper. Simulation success and real-world reliability are different things.

Runway's opposite bet: Praxis-1

Runway took the specialist route two days before Fortune's report. Praxis-1 is a 27-billion-parameter "world action model" built specifically to control robots, released with its weights and fine-tuning code under the Apache 2.0 licence, which permits commercial use.

On the same RoboLab-120 benchmark, Praxis-1 scored 46.2 percent — the highest world action model listed, but below Astra's 74.8 percent in simulation. That looks like a clear win for the general model until you look at what each number measures. Praxis-1 is built for real-time control on real hardware and runs on a single H100 server; a general model scoring higher in simulation may be slower and costlier on a robot that must react in milliseconds.

It has happened before

This is the second time OpenAI has closed a robotics team. The first, in 2020–2021, was a retreat: the company wound the team down and stepped back from robotics. The difference is the stated reason. In 2020, robotics was a distraction from the language models that became GPT. In 2026, OpenAI says those models have grown into the robot's brain.

That claim will be tested in public. The humanoid demo targeted for 2027 is a harder test than a simulation benchmark. If a general model can drive a physical humanoid through real tasks, the reorganisation will look prescient.

What it means for the robotics race

The practical effect is a split in strategy. OpenAI is betting that scale wins: one frontier model, applied to everything, including robots. Runway is betting that purpose-built models win where timing, cost and reliability matter.

Key takeaways

  • OpenAI has disbanded its stand-alone robotics team, Fortune reported on 13 November; researchers moved under Chief Research Officer Mark Chen, applied teams under President Greg Brockman.
  • OpenAI's stated reason: general-purpose models like GPT-6 Astra can now be applied directly to robots, so a separate robot-specific team is no longer the fastest route.
  • An Nvidia study cited by Fortune found an Astra-powered agent reached 74.8 percent on a simulated manipulation benchmark, vs 51.0 percent for the best specialised models — simulation, not physical robots.
  • Runway released Praxis-1 two days earlier: a 27-billion-parameter open-weight robot model under Apache 2.0, scoring 46.2 percent on the same benchmark — the best world action model listed.
  • OpenAI says it is still building its own humanoid, with a first public demo targeted for 2027. This is the second time it has closed a robotics team; the first was in 2020–2021.

Conclusion

OpenAI's message is that the robot's brain no longer needs to be built separately — the general model is the brain. Runway's message is that robots still need models built for their bodies, timing and budgets. Both may be partly right. The deciding evidence will come from robots doing real work, in real rooms, without a researcher standing by.

FAQ

Has OpenAI quit robotics?

No, according to the company. It disbanded its stand-alone robotics team and reassigned the staff into its main research and Applied organisations. It is still building its own humanoid, with a first public demo targeted for 2027.

What is Runway Praxis-1?

Praxis-1 is a 27-billion-parameter "world action model" for controlling robots, released by Runway in November 2026 with open weights and fine-tuning code under Apache 2.0. It scored 46.2 percent on RoboLab-120 — the highest world action model listed.

Why does OpenAI think general models can run robots?

OpenAI says general-purpose frontier models including GPT-6 Astra can be applied directly to robotics. An Nvidia study cited by Fortune found an Astra-powered agent reached 74.8 percent on a simulated benchmark, against 51.0 percent for the best specialised models. The result was in simulation, not on physical robots.

Researched with AI assistance and reviewed for accuracy.

Sources

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp for one tech and AI update every day: Daily Tech & AI WhatsApp channel.

Comments