OpenAI disbands its robotics team for the second time, betting GPT-6 Astra can run robots directly
Fortune reported November 13 that OpenAI has disbanded its standalone robotics team for the second time in six years, moving researchers into chief research officer Mark Chen's main research org and applied and hardware staff into president Greg Brockman's Applied group. OpenAI confirmed the reorganization, saying general-purpose models including GPT-6 Astra can now be applied to robot control directly. A humanoid demo is still targeted for 2027.
The News
Fortune reported November 13 (exclusive) that OpenAI has disbanded its standalone robotics team for the second time in six years. Researchers are moving into chief research officer Mark Chen's main research organization, and applied and hardware staff are moving into president Greg Brockman's Applied group. OpenAI confirmed the reorganization to Fortune, saying general-purpose models including GPT-6 Astra can now be applied to robot control directly. OpenAI is a private company; Microsoft, its backer, is the public read-through ticker here.
The evidence OpenAI points to comes from an Nvidia study. In it, an Astra-powered agent scored 74.8% on RoboLab-120, a simulated benchmark that tests robot manipulation — picking things up and moving them in a virtual environment — versus 51.0% for the best specialist model built specifically for that task. But the caveat is important: 74.8% is the study's most scaffolded 'agentic' setup, with persistent history and extra tools; the direct-control setup scored 66.3%, and the lowest Astra configuration scored about 50%, roughly level with the specialist. The test was in simulation only, not on a physical robot, and the simulation was paused while the model thought, so it does not test real-time control. On cost and speed, the two figures come from different setups: the agentic approach cost an estimated $36.40 per successful episode, while the direct setup averaged 18.56 seconds per model call. Either way, that is far too slow and too expensive for a robot doing real work today.
A humanoid demo is still targeted for 2027. The reorganization does not cancel OpenAI's robot ambitions — it changes how they are organized. Newer real-hardware evidence is mixed: on OpenAI's StationeryBench test with real dual-arm robots, Astra completed 7 of 100 desk tasks versus 0 for the specialist rival MolmoAct2 — a win, but a long way from reliable. The bet is that the model itself, rather than a dedicated robotics team, is now the right tool for the job. Microsoft shares closed Friday at $523.88, up 0.17%.
OpenAI is betting the model is the robot team now — but the proof is a simulation that costs $36 a try.
Why It Matters
OpenAI's bet is that the model has become the robotics team. For years, building a robot meant building a specialist system: custom software for grasping, custom software for moving, custom software for planning. If a general-purpose model like GPT-6 Astra can do all of that directly — and the 74.8% versus 51.0% result says it can, at least in simulation — then a separate robotics group is duplicating work the main research org already does.
The second disbanding in six years also tells you something about how OpenAI works. The company reorganizes around its models' capabilities: when a model crosses a threshold, the team built for the old way of doing things gets folded in. That is efficient if the model really is ready, and it is also how OpenAI has operated before — this is not a new management philosophy, it is the same one applied to a new capability.
The reason for caution is the gap between simulation and the real world. A simulated robot does not drop things, does not deal with bad lighting, and does not cost $36 and 19 seconds every time it moves. The 2027 humanoid demo target is the real test: if OpenAI shows a physical robot doing useful work, the reorganization looks visionary. If the demo slips or leans on simulation results, the disbanded team will look like a team that was needed after all.
Analyst's Take
Our call is Mixed. There is a real signal here: a general-purpose model beating the best specialist model on a manipulation benchmark, 74.8% to 51.0%, is the kind of result that changes how a company organizes itself. If the model really can handle robot control, a separate robotics team becomes overhead — and folding it into the main research org puts the best model researchers on the problem. OpenAI has also done this before, so the move fits a pattern of the company reorganizing around whatever its models can now do.
But the evidence is thinner than the headline suggests. RoboLab-120 is a simulation, not a real robot in a real room. The cost and speed figures come from different setups: about $36.40 per successful episode for the most scaffolded agentic approach, and 18.56 seconds per model call for the direct setup. Either way, the Astra-powered approach is far too slow and expensive for practical robot work — a warehouse robot cannot pause for nearly twenty seconds every time it picks something up. A simulation win is a research result, not a product.
Reader poll: is OpenAI right to fold robotics into its main research org? Vote: Smart bet, Too early, or Wait and see.
Are you bullish or bearish on this story?
By The Numbers
FAQ
Why would OpenAI disband its robotics team?
OpenAI's argument is that it no longer needs a separate team. It says general-purpose AI models — including its GPT-6 Astra model — have become good enough to control robots directly, so robot work can happen inside the main research organization rather than in a dedicated robotics group. The company confirmed the reorganization to Fortune.
What is the RoboLab-120 benchmark evidence?
An Nvidia study cited in the reporting tested an Astra-powered agent on RoboLab-120, a simulated test of robot manipulation — in plain terms, picking things up and moving them in a virtual environment. The Astra-powered agent scored 74.8% versus 51.0% for the best specialist model — but 74.8% is the study's most scaffolded 'agentic' setup; the direct-control setup scored 66.3%. The test was in simulation only, not on a real robot, and the simulation paused while the model thought. On cost and speed, the figures come from different setups: about $36.40 per successful episode for the agentic approach, and 18.56 seconds per model call for the direct setup — far too slow and expensive for real-world use today.
Is this the first time OpenAI has done this?
No — it is the second time in six years that OpenAI has disbanded a standalone robotics team. The company has gone back and forth between treating robotics as its own group and folding it into the main research effort.
Is OpenAI still building a humanoid robot?
A humanoid demo is still targeted for 2027, according to the reporting. The reorganization changes how the work is organized, not whether it happens — though with no dedicated team, the demo will now be built by researchers inside the main org and the Applied group.
👁
Source: ExplainX (summarising Fortune); figures from Nvidia Astranomy study