When General AI Beats Specialists: What OpenAI's Robotics Shutdown Means for Builders
OpenAI shut down its dedicated robotics team after a general AI model outperformed specialist ones — a signal that purpose-built AI is under pressure from the very models it was meant to beat.
- 74.8%
- Astra's score on robot manipulation test
- 51.0%
- best specialized robot model's score
- 2nd time
- OpenAI has disbanded a robotics team
- 2027
- target year for OpenAI's first humanoid demo
OpenAI just shut down its dedicated robotics team. Not because robotics failed — but because its general-purpose AI, GPT-6 Astra, turned out to be better at the job than the specialists were. The researchers moved into OpenAI's broader research and applied teams. The company's bet is that a general-purpose model will carry robotics forward on its own.
The numbers back that up. In an Nvidia simulation study, an Astra-powered agent scored 74.8% on a robot manipulation test — picking and placing objects in a virtual environment. The best dedicated robotics models on the same leaderboard scored 51.0%. A nearly 50% improvement, with no robotics-specific training at all. These are simulated results, not a real robot, but the gap is large enough to notice.
This is the second time OpenAI has closed a robotics team. The first was in 2020. Back then, robot AI was its own discipline — dedicated data, dedicated models, dedicated staff. What has changed is that frontier models are now capable enough to erase that division. OpenAI is betting that one broadly trained model will keep improving until specialist teams are no longer necessary.
The practical question for anyone building with AI: before you invest in a specialized model or a dedicated team for a narrow task, test what a general model can already do. The gap between general-purpose and purpose-built AI is closing faster than most expect. The real edge now comes from deployment — real hardware, real data, and real environments that only come from getting out into the field.