OpenAI Disbands Robotics Team Again: GPT-6 Astra Completed 74.8% of Tasks in Simulation Versus 51%
OpenAI is disbanding its dedicated robotics team once again, betting that the general-purpose GPT-6 Astra can control a robot directly, without a narrow "robot" model. In an Nvidia study, an agent built on Astra completed 74.8% of tasks in simulation, while the best specialized models managed 51%. But the test ran on the virtual DROID robot, not on hardware. For that reason, general models are worth trying as the planning "brain," while narrow models should stay in charge of fast actions.
AI-processed from The Decoder; edited by Hamidun NewsAccording to Fortune, OpenAI is disbanding its dedicated robotics team for the second time. The reasoning is that a separate "robot" neural network may turn out to be unnecessary: the general-purpose GPT-6 Astra may be able to control a robot directly. To put it simply: previously, a robotic arm got a narrow model that learned only to move the manipulator.
Now the same large model that writes code and answers questions looks at the scene itself and decides where to reach. In an Nvidia study, an agent built on Astra completed 74.8% of tasks, while the best specialized models completed 51%.
An important caveat: the tests ran on the virtual DROID robot, that is, in simulation, not on real hardware. For those building robots or agents, this is a signal of a possible market shift, but not yet proof.
What exactly is changing inside OpenAI
Robotics researchers are moving into the main research team led by chief research officer Mark Chen. Applied robotics and hardware groups are joining Applied, which is led by Greg Brockman. The decision was made so abruptly that the company had already booked a booth at the Conference on Robot Learning (CoRL) in Austin, and it stood empty.
In 2020, OpenAI had already shut this work down. Back then, according to co-founder Wojciech Zaremba, there was not enough training data, and large language models seemed a faster path to AGI. The team was reassembled in 2024. Now the reason is different: the company is not abandoning robots, but wants to speed up its work by dissolving specialists into other teams.
What the tests show
In addition to the Nvidia study, there is a new benchmark, StationeryBench. In it, Astra was compared with Ai2's MolmoAct2 on five tasks involving stationery items on a desk. The tasks were performed by YAM dual-arm robots, for a total of 200 attempts. Astra fully completed 7 of the 100 tasks, while MolmoAct2 completed none. The average progress score was 46 for Astra and 12 for MolmoAct2. Astra has an advantage, but reliable performance is still far off: Astra did not finish 93 of the 100 tasks.
In another test, Astra on a humanoid robot moved around a kitchen and tidied up. Yoav Artzi of Cornell University and Google DeepMind calls the model a major leap in spatial understanding. According to him, on the still unpublished REMAP benchmark, Astra shows accuracy close to human level. He suggests OpenAI could have trained it on a large volume of 3D data, for example scenes from Blender. How much robotics data OpenAI and other labs used is unknown. Progress on robotics tasks has also been noted in large models from Anthropic and Google.
What has not been proven yet
Narrow models may retain advantages in cost, reliability, speed and power consumption, even if they generalize worse. For a robot in a warehouse or factory, this is critical: a huge general-purpose model is more expensive to compute and reacts more slowly. Simulation results do not guarantee the same on physical robots, and some benchmarks have either only just appeared or have not yet been published.
There is also a business context. Sam Altman reportedly aims for a first public robot demonstration in 2027, and in the long term OpenAI wants to build its own household robots. But the company is under heavy financial pressure and needs to focus on its core AI business, where it has clearly lost ground to Anthropic. The reorganization may therefore be as much a way to conserve resources as a technological bet.
What this means for practitioners
If you are building a robotics product, there is no need to rush to discard specialized models: on price, speed and predictability, they may still win. But it already makes sense to experiment with general-purpose models as the "brain" for planning and scene understanding, while keeping narrow models for fast low-level actions. The key question for the next year is whether the simulation results will hold up on real robots, and how much this approach will cost to operate.
- Nov 21, 2026Amazon Releases SGLang-Neuron: gpt-oss-120b Runs on Trainium Chips Without Changing the API
- Nov 21, 2026Factory Cut Metal Pricing from 4–8 Hours to 2–3 Minutes by Removing AI from the Arithmetic
- Nov 21, 2026Jensen Huang Steps Down as NVIDIA CEO After 33 Years, CFO Colette Kress Becomes Interim CEO
- Nov 21, 2026Gondola Used Codex to Ship 25% More Code Without Hiring and Doubled Revenue
Key Questions Answered
What exactly is changing inside OpenAI?
Robotics researchers are moving into the main research team led by chief research officer Mark Chen. Applied robotics and hardware groups are joining Applied, which is led by Greg Brockman. The decision was made so abruptly that the company had already booked a booth at the Conference on Robot Learning (CoRL) in Austin, and it stood empty. In 2020, OpenAI had already shut this work down.
What the tests show?
In addition to the Nvidia study, there is a new benchmark, StationeryBench. In it, Astra was compared with Ai2's MolmoAct2 on five tasks involving stationery items on a desk. The tasks were performed by YAM dual-arm robots, for a total of 200 attempts. Astra fully completed 7 of the 100 tasks, while MolmoAct2 completed none. The average progress score was 46 for Astra and 12 for MolmoAct2. Astra has an advantage, but reliable performance is still far off: Astra did not finish 93 of the 100 tasks.
What has not been proven yet?
Narrow models may retain advantages in cost, reliability, speed and power consumption, even if they generalize worse. For a robot in a warehouse or factory, this is critical: a huge general-purpose model is more expensive to compute and reacts more slowly. Simulation results do not guarantee the same on physical robots, and some benchmarks have either only just appeared or have not yet been published. There is also a business context.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.