🔍 Read the full analysis: The Case For Robot AI? OpenAI Splits Its Team As Runway Opens Praxis-1 on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI has dissolved its stand-alone robotics team, moving researchers into its main research group and applied robotics and hardware staff into its Applied organization. Runway released the open-weight Praxis-1 robot model two days before the report, highlighting a trade-off in published tests: general models can score higher in simulation, while specialist models respond faster.
OpenAI has broken up its stand-alone robotics team and moved its researchers into other company groups, Fortune reported on November 13, while Runway released a dedicated robot model, Praxis-1, two days earlier. The developments put two approaches to robot control in focus: using general-purpose AI models or building models specifically for robots. Published tests show a trade-off between performance in simulation and the time needed to act.
Fortune said OpenAI confirmed the reorganization after the outlet asked about it. Robotics researchers now report to Chief Research Officer Mark Chen, while applied robotics and hardware teams moved to the Applied organization under President Greg Brockman. Fortune found the company’s registered booth at the CoRL robotics conference in Austin empty. OpenAI had not published its own announcement about the change as of November 14.
An OpenAI spokesperson told Fortune the move reflects a commitment to robotics, not a retreat. The spokesperson said general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. OpenAI had initially expected to build on video models such as Sora, according to the account. The company has also said it is developing a humanoid robot; CEO Sam Altman has targeted a public demonstration in 2027.
Runway released Praxis-1 as a 27-billion-parameter world action model with open weights and code for fine-tuning, under the Apache 2.0 license. On RoboLab-120, a simulation benchmark of 120 tabletop tasks, it scored 46.2%, according to the source material. That was the highest result among the world action models and open-source robot models listed, and fifth overall. The same material says Praxis-1 takes about 0.55 to 0.6 seconds per call to produce 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
The Trade-Off Between Reach and Speed
The competing developments matter because robots must both choose useful actions and respond quickly enough to carry them out. A high success rate in a simulation does not, by itself, show that a model can control a physical robot reliably in real time. Conversely, faster action does not establish that a specialist model can handle as broad a range of tasks as a general-purpose system.
In an NVIDIA researchers’ simulation study, GPT-6 Astra reached 74.8% success in an agentic setup on RoboLab-120, compared with 51.0% for the best trained robot policy on the leaderboard cited in the material. But in the direct setup, each call averaged 18.56 seconds, and the simulation paused while the model reasoned. Runway’s reported timing for Praxis-1 was much shorter, but its 46.2% benchmark score was lower. These are results from different systems and test configurations, not proof that one model is better for every use.
Runway also compared the models on one of its robots, saying GPT-6 Astra took 15.6 times longer than Praxis-1 to complete three tasks and cost 51 times more per completed task. Those figures are Runway’s own test results; the source material does not provide enough detail to establish how broadly they apply. Together, the results point to a practical possibility: general models may help with flexible planning, while faster specialist models handle immediate control.
Popular right now
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
KitchenAid Artisan 5-Qt Tilt-Head Stand Mixer, KSM150PS, Almond Cream
Dyson V8 Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
OpenAI’s Robotics Team, Reorganized
This is the second time OpenAI has ended a stand-alone robotics team. It wound down its first team in 2020. The company rebuilt a robotics group in 2024 and partnered with humanoid maker Figure; Figure said in February 2025 that it would develop its own robot AI, ending that partnership.
The latest change follows a shift in OpenAI’s stated technical rationale. According to Fortune, researchers had expected earlier in 2026 to build on video models such as Sora. The spokesperson’s explanation was that newer frontier models can be applied to robotics directly, leading OpenAI to move robotics researchers into the main research organization. Fortune also described pressure on the company to control costs and focus on core products, but that was background in its report, not the reason OpenAI gave for the reorganization.
The benchmark evidence is not a direct contest between equivalent products. NVIDIA researchers tested GPT-6 Astra as a robot policy without robot-specific training, using simulated camera images, robot information and a control interface. Its agentic setup used a persistent Codex session and additional tools. Praxis-1 is a specialist model designed for robot action. Different methods, hardware and test conditions limit what can be concluded from comparing their headline scores.
“The two approaches are complementary.”
— Runway
What the Tests Do Not Show
OpenAI has not published a detailed statement explaining the reorganization, and it is not clear how the new reporting structure will affect staffing, research priorities or the company’s robot development plans. The source material says the company has targeted a 2027 public demonstration of its humanoid robot, but gives no further schedule or technical details.
The reported benchmark results are simulation findings, not evidence that either system is ready for broad real-world deployment. The source material does not provide enough detail about the full methodology behind Runway’s comparison on its robot to independently assess its timing and cost figures. It also remains unclear how well either approach will perform across different robots, environments and tasks, or whether combining them will improve results.
The Next Test Is Real-World Control
The next milestones are whether OpenAI shares more information about its reorganized robotics work and whether it makes progress toward the public robot demonstration targeted for 2027. More details on the system, its capabilities and how it is tested would help clarify what the organizational change means in practice.
For both companies, further evidence will need to show how models perform beyond benchmark simulations. Relevant tests would include task success, response time, reliability and cost across multiple robots and physical settings. Until those results are available, the published figures support a comparison of specific tests, not a firm conclusion that general-purpose or specialist models are the better route to robot control.
Key Questions
What changed in OpenAI’s robotics organization?
According to Fortune, OpenAI moved its robotics researchers into the main research organization under Chief Research Officer Mark Chen. Applied robotics and hardware teams moved to the Applied organization under President Greg Brockman.
What is Runway’s Praxis-1?
Praxis-1 is a 27-billion-parameter robot model released with open weights and code for fine-tuning under the Apache 2.0 license. Runway describes it as a world action model.
Which model scored higher on RoboLab-120?
In the reported results, GPT-6 Astra scored 74.8% in an agentic simulation setup, while Praxis-1 scored 46.2%. The systems used different approaches, so the figures are not a like-for-like measure of performance in every setting.
Does the research show that general AI models can control robots in the real world?
No. The cited GPT-6 Astra result came from a simulation study. The results do not establish how reliably it or Praxis-1 would control robots across real-world environments and tasks.
Source: ThorstenMeyerAI.com
