OpenAI disbands robotics team as GPT-6 Astra handles tasks without training
The new model can control robotic hardware using only visual and status data, potentially making dedicated robot brains obsolete.

The short version
- OpenAI disbanded its robotics team just five months after a recruitment push in May.
- GPT-6 Astra achieved a 74.8% success rate in a simulated robotic arm test with zero prior robotics training.
- Dedicated robot models, which underwent intensive training, only reached a 51% success rate in the same test.
- Anthropic's Claude Fable 5.2 also demonstrated similar capabilities, solving tasks like Rubik's cubes without explicit training.
- The move reflects a shift toward using general models as 'brains' for robots rather than developing custom software for every task.
Why it matters
This shift suggests that the robotics industry may no longer need to build specialized AI for every unique physical action. Companies may need to pivot from developing 'brains' to focusing on hardware and providing high-quality real-world data.
A sudden shift in strategy
OpenAI has unexpectedly disbanded its internal robotics team, a move that comes only five months after Sam Altman personally posted a recruitment notice in May. While the original team will be integrated into the core research group under Mark Chen, the applied robotics and hardware teams will now fall under the leadership of OpenAI President Greg Brockman.
This decision follows the release of GPT-6 Astra, a model that has significantly altered the landscape for robot development. Industry observers note that the company appears to have determined that dedicated robotic brains are no longer necessary because the general model can perform those functions directly.
The shift was visible at the recent Conference on Robot Learning in Austin, where OpenAI's presence was marked on the map but featured an empty booth. This absence highlights the rapid transition from building specialized hardware-centric AI to utilizing general superintelligence as the primary operating system for robots.
How Astra handles tasks
In the past, creating a robot capable of performing a specific action, such as pouring water, required teaching it every individual step using real-world data. These 'dedicated brains' were often brittle, meaning a robot trained for one task would immediately fail if asked to perform a different action, like tightening a screw.
GPT-6 Astra functions differently by acting as a general-purpose brain that requires zero prior training on specific robotics tasks. Researchers only need to provide the model with three inputs: a camera feed, the robot's own status data, and a connection to the control interface. The model then observes the visual feed, breaks down the steps, and issues commands on its own.
During a simulated robotic arm test hosted by NVIDIA, Astra achieved a 74.8% success rate right out of the gate. In contrast, dedicated models that underwent long-term, intensive training on physical robots only reached a maximum success rate of 51% in the same simulation.
The rise of general models
OpenAI is not the only company seeing these results, as Anthropic has also demonstrated similar capabilities with Claude Fable 5.2. In official videos, the model controlled two robotic arms to solve a Rubik's cube, assemble gears, and draw a robot, despite never being explicitly trained on those specific tasks.
The speed of this progress is notable, as Claude was reportedly struggling with basic navigation just four months ago. At that time, a quadruped robot controlled by the model got stuck in a trash can, failing to recognize the obstacle despite identifying its location. Now, the model can perform complex manual tasks with significantly less effort.
This evolution suggests that robots are becoming downstream applications of general intelligence rather than a unique problem. As one professor from Carnegie Mellon University noted, from OpenAI's perspective, robots are no longer a special problem but are similar to tools for coding or financial analysis.
Industry impact and hurdles
The news has sent shockwaves through the robotics industry, where hundreds of billions of dollars have been invested in specialized AI brains. Companies like Figure, which reached a valuation of 40 billion US dollars, were betting on the necessity of custom-built models for every robot action. Now, the focus may shift toward building the physical bodies and providing the real-world operational data that AI labs lack.
However, using large models as robot brains currently faces two major hurdles: speed and cost. Because these models often run in the cloud, they can suffer from network latency, which is dangerous for tasks requiring precise, real-time movements. A test by Tsinghua University showed that running a full cycle with Astra on a dual-arm robot cost 12.5 US dollars, which may be too expensive for many applications.
Additionally, while general models are excellent at understanding instructions, they may still struggle with the subtle coordination of force required for delicate tasks. For these reasons, small specially trained models may still be needed for high-precision physical interactions even as the 'brain' becomes more generalized.
The path forward
OpenAI's move also reflects internal financial pressures, as the company reported a loss of 10.9 billion US dollars in its most recent quarter. By disbanding the robotics team and cutting other projects like Sora, the company is streamlining its focus toward the most viable paths to superintelligence. This indicates a shift toward prioritizing models that can handle multiple domains simultaneously.
For robotics companies, the advice from experts like an MIT professor is to pivot quickly. Instead of trying to build dedicated brains, these companies should focus on the physical hardware and the unique data gathered from actual work in factories. The more capable foundational models become, the more valuable this real-world experience data will be for training the next generation of systems.
The industry is moving toward a future where robots may be viewed as 'action figures' manipulated by a central superintelligence. While the transition is still in its early stages, the move by OpenAI signals a major shift in how the world will approach the integration of AI into the physical environment.
What to watch
- How robotics companies will pivot their business models to focus on hardware and data.
- Whether cloud-based latency issues can be solved for real-time robotic control.
- The next public demonstration of humanoid robots from OpenAI, currently slated for 2027.
Try it
Look up the success rates of GPT-6 Astra versus dedicated robot models in the NVIDIA simulation tests.
Sources and links
- Reportingeu.36kr.com: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by eu.36kr.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
