🔍 Read the full analysis: Could Robots Need Their Own AI? OpenAI's Team Split Meets Runway's Praxis-1 on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
OpenAI has disbanded its stand-alone robotics team and redistributed its researchers and applied robotics staff across existing organizations, saying general-purpose models can now be used directly for robot control. Runway has released Praxis-1, an open-weight specialist model that performed less well than GPT-6 Astra in one simulation benchmark but acted much faster in reported tests. The available evidence does not settle whether robots will rely on general models, specialist models or a combination.
OpenAI has broken up its stand-alone robotics team and moved its researchers and applied robotics staff into other parts of the company, while Runway has released Praxis-1, a dedicated robot model with downloadable weights. The contrasting moves put a practical question for robotics researchers and companies in focus: whether general-purpose AI can control robots effectively, or whether robots still need models built specifically for action.
Fortune reported November 13 that OpenAI confirmed the reorganization when asked. Its robotics researchers moved into the main research organization under Chief Research Officer Mark Chen, while the applied robotics and hardware teams moved to the Applied organization under President Greg Brockman. OpenAI has not published its own announcement about the change, according to the source material.
OpenAI’s spokesperson told Fortune that the shift reflects a change in how the company thinks robot capabilities can be built. Researchers had expected to build on video models such as Sora, but the spokesperson said general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. The spokesperson described robotics as part of OpenAI’s long-term vision and said the company remains committed to general-purpose robots.
Runway released Praxis-1 two days before Fortune’s report. The source describes it as a 27-billion-parameter “world action model” released under the Apache 2.0 license, with weights and code available for fine-tuning. In the cited RoboLab-120 results, Praxis-1 scored 46.2%, the highest score among the listed world action models and open-source robot models, and fifth overall. Its reported calls took about 0.55 to 0.6 seconds for 1.6 seconds of motion.
Robot control loop · explainer
Who decides the robot’s next move?
Every robot repeats one loop: cameras see, a model decides, the arm moves. The two bets in this story put a different kind of model into the decide step.
Option 1
A general model in the loop
Example: GPT-6 Astra in NVIDIA’s simulation study
18.56 s per model call on average in the direct set-up. Inference was 97.0% of the estimated episode time; the simulation paused while the model was thinking.
Option 2
A specialist robot model in the loop
Example: Runway’s Praxis-1, a world action model with open weights
About 0.55–0.6 s per call on one H100 server, and each call returns 1.6 s of motion. Needs fine-tuning data for each robot.
General model plans the next stepslow, outside the loop
Option 3 · illustrative
Both, in two layers
The idea Runway and the NVIDIA researchers both describe
Runway calls the two approaches “complementary”: an agent plans, a fast policy carries out each step. NVIDIA’s authors say the best way to combine them “remains an open question”.
Sources: NVIDIA Seattle Robotics Lab, “Astranomy” study of GPT-6 Astra, Oct 14, 2026 (simulation, direct set-up); Runway, Praxis-1 technical report, Nov 2026 (one H100 server). Loops and timelines are schematic; the two figures come from different set-ups, not from a head-to-head test.
Graphic: Powered by Thorsten Meyer AI
Speed and Success Trade Off
The competing approaches highlight a gap between solving tasks and acting at robot speed. In a simulation study by NVIDIA researchers, GPT-6 Astra’s agentic setup achieved a 74.8% success rate on RoboLab-120, compared with 51.0% for the best trained robot policy on the leaderboard cited in the source. But in the study’s direct setup, each model call took an average of 18.56 seconds, and the simulation paused while the model processed a response.
That delay matters outside a paused simulation, where a robot must respond as it moves and conditions change. Runway’s own test on one of its robots found GPT-6 Astra took 15.6 times longer than Praxis-1 to finish three tasks and cost 51 times more per completed task. Those figures come from Runway’s test and should not be treated as a general comparison across robots or tasks. They illustrate why benchmark success alone does not establish which approach is more useful in deployment.
For developers, the immediate issue is not simply which model scores higher. A general model may handle a wider range of tasks, while a dedicated model may deliver responses quickly enough for physical control. Runway describes the approaches as complementary, and the supplied evidence leaves open whether a robot could use both—for example, a general model for planning and a specialist model for fast movements.
Products Worth Considering
LG UltraGear 24-Inch Full HD IPS Gaming Monitor, 120Hz (144Hz Overclocked), 1ms MBR, G-SYNC Compatible, FreeSync, HDR10
Logitech MX Vertical Wireless Ergonomic Mouse
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
OpenAI’s Changing Robotics Bet
OpenAI’s latest change is its second shutdown of a stand-alone robotics team. The company wound down an earlier team in 2020. It rebuilt robotics research in 2024 and partnered with humanoid maker Figure, but that partnership ended in February 2025, when Figure said it would develop its own robot AI, according to the supplied report.
The company has also said it is developing a humanoid robot, with CEO Sam Altman targeting a public demonstration in 2027. That target is a stated plan, not confirmation that a demonstration will occur. OpenAI’s reorganization also came just before the Conference on Robot Learning in Austin: its booth remained on the floor plan, but Fortune found it empty. The source says the company had already registered the booth before the recent change.
The cited comparison is based on different kinds of evidence. NVIDIA researchers tested GPT-6 Astra in RoboLab-120, a simulation of 120 tabletop tasks for a single robot arm, without robot-specific training. Runway’s task-time and cost comparison was its own test on one of its robots. The results offer a snapshot of performance under those conditions, not a general verdict on every robot or real-world setting.
“The two approaches are “complementary.””
— Runway, as described in the supplied source
Limits of the Available Tests
The benchmark results do not establish real-world performance. RoboLab-120 is a simulation of tabletop tasks for one robot arm, and the supplied material does not show how GPT-6 Astra or Praxis-1 performs across different robot bodies, settings or longer tasks. In the NVIDIA study’s direct setup, the simulation paused during model processing, so its reported success rate does not capture the consequences of those delays in continuous operation.
The Runway figures also come from a company test on one of its robots and cover three tasks. The source does not provide enough detail here to assess how broadly those results apply or how the test conditions compare with NVIDIA’s simulation. OpenAI has not released a statement of its own about the reorganization, and the full reasoning behind the internal changes is not publicly documented in the supplied material. It remains unclear whether OpenAI will publish further robotics results or how its reorganized teams will develop robot-control systems.
Robot Plans and Further Testing
The next reported milestone for OpenAI’s hardware effort is the public humanoid-robot demonstration Altman has targeted for 2027. The company has not provided a confirmed date or further details in the supplied material. More information about how the reorganized teams will work on robotics may come through future company statements, research or demonstrations.
For Praxis-1, the next test is whether results reported by Runway and benchmark performance hold across independent evaluations, additional robots and tasks outside simulation. Researchers will also need to test how general and specialist systems perform when used together. Until those results are available, the evidence supports a trade-off—higher simulated task success for GPT-6 Astra in one setup, faster reported action for Praxis-1—rather than a settled answer about which model type robots need.
Key Questions
What changed in OpenAI’s robotics organization?
According to Fortune, OpenAI moved its robotics researchers into its main research organization under Mark Chen and transferred its applied robotics and hardware teams to the Applied organization under Greg Brockman. OpenAI confirmed the reorganization when Fortune asked, but had not published its own announcement in the supplied source.
What is Runway’s Praxis-1?
Praxis-1 is a 27-billion-parameter robot model that Runway released under the Apache 2.0 license, with weights and code available for fine-tuning. The source reports a 46.2% score on RoboLab-120 and calls of roughly 0.55 to 0.6 seconds for 1.6 seconds of motion.
Did GPT-6 Astra outperform Praxis-1?
Not in a direct, like-for-like comparison across the evidence provided. GPT-6 Astra reached a 74.8% success rate in an NVIDIA researcher simulation study on RoboLab-120. Runway reported separate performance results for Praxis-1 and a company test comparing task time and cost on one of its robots. The different test settings limit direct conclusions.
Does OpenAI’s reorganization mean it is leaving robotics?
OpenAI’s spokesperson told Fortune that robotics remains part of the company’s long-term vision and that it remains committed to general-purpose robots. The team structure changed, but the supplied material does not establish what specific robotics projects or staffing levels will follow.
Will robots use general-purpose or specialist AI models?
The available results do not settle that question. General-purpose and specialist models showed different strengths in the cited tests, and Runway described the approaches as complementary. Further testing across robots and real-world tasks is needed to determine how they may be used.
Source: ThorstenMeyerAI.com
