OpenAI disbands its robotics team as GPT-6 Astra runs robots directly: Reveneau

Why it mattersOpenAI is betting that a general model is better at a robot task than a model trained for the task. The robotics job market at frontier labs just got much smaller.

A company that had a whole team working on robot brains now says it does not need a whole team. OpenAI just said that.

Fortune reported on 13 November that OpenAI has dissolved its standalone robotics team. The reporting is by Wen Shao, an AI Fellow at the publication. Researchers on the team have been reassigned, Fortune said: the research group now reports to Mark Chen, OpenAI's chief research officer. The applied robotics and hardware groups now report to Greg Brockman, the company's president. OpenAI said the move is an acceleration of its robotics work, not a retreat.

What a general model is now doing

The reasoning, per Fortune, is that GPT-6 Astra can run a robot without a robotics-specific model behind it. Fortune cites a study in which GPT-6 Astra scored 74.8 percent on a simulated manipulation benchmark. The specialised robotic models in the comparison scored 51.0 percent. OpenAI, quoted in the article, said its general models can be directly applied to robotics. The company told Fortune it plans to demonstrate its first humanoid robot in 2027.

The framing is not just corporate. Guanya Shi is a robotics researcher at Carnegie Mellon. Shi told Fortune that robotics is "no longer a special problem." It is now a downstream application of general intelligence. Shi is not an OpenAI employee.

The second time this team has gone

OpenAI first broke up its robotics team in 2020. Co-founder Wojciech Zaremba rebuilt it in 2024. A partnership with the humanoid company Figure was signed the same year. It ended in February 2025, when Figure decided to build its own robot brain. The 2026 move is the second time the team has been folded into the rest of the company.

What a reader should take from it

This is OpenAI's own strategy call. It is one lab, and one set of benchmark numbers that only one group has published. A model doing well on a simulated task is one thing. A model doing well in a real kitchen that it has never seen is another.

For a founder building an AI product, the useful part of the story is the bet behind it. Spending to train a specialised model for a narrow task looked like the safe choice last year. This month, the largest AI lab has staked its roadmap on the opposite answer. A buyer who commissioned a specialist model should test it against a frontier general model. The vendor should run both on the same benchmark.

For an engineer whose career is in robotics, the signal is different. Two labs built most of the recent robotics hiring in the field. OpenAI in 2024 and Figure in 2025. Both have now folded or ended that work. Fortune does not give the headcount of the dissolved team. The lab's position is that its first humanoid demonstration is still 15 months out.

Source

OpenAI disbands its robotics team as its general models start running robots, Wen Shao, Fortune.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.