🔍 Read the full analysis: How OpenAI's Robotics Split And Runway's Praxis-1 Put Robot AI In Focus on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI has moved its robotics researchers and applied robotics teams into other parts of the company, saying general-purpose models can now be used directly for robot control. Runway, meanwhile, released Praxis-1, an open-weight specialist model; published results suggest general models can perform well in simulation, while specialist models respond much faster.
OpenAI has disbanded its stand-alone robotics team and moved its researchers into its broader research organization, while Runway has released Praxis-1, a dedicated robot-control model with downloadable weights. The developments, reported around the CoRL robotics conference in Austin, sharpen a debate over whether robots should rely on general-purpose AI or models built specifically for physical tasks.
Fortune reported on November 13 that OpenAI confirmed the reorganization after the outlet found the company’s registered booth at CoRL empty. The robotics researchers now report to Chief Research Officer Mark Chen, while applied robotics and hardware teams have moved to the Applied organization, led by President Greg Brockman. OpenAI has not published its own announcement about the change.
An OpenAI spokesperson told Fortune that the move reflects the company’s commitment to robotics, not a retreat from the field. The spokesperson said general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. Fortune reported that OpenAI had initially expected to build on video models such as Sora. OpenAI has also been developing a humanoid robot, with CEO Sam Altman previously targeting a public demonstration in 2027.
Runway released Praxis-1 two days before Fortune’s CoRL report. The company describes it as a 27-billion-parameter “world action model” and offers its weights and fine-tuning code under the Apache 2.0 license, which permits commercial use. On the RoboLab-120 benchmark, Praxis-1 scored 46.2%, according to the reported results: the highest score among the listed world action and open-source robot models, and fifth overall. Its reported call time was 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
Robot AI Faces a Speed Trade-off
The competing approaches highlight a practical divide between task performance and response time. In a simulation study by researchers at NVIDIA’s Seattle Robotics Lab, an agent using GPT-6 Astra achieved a 74.8% success rate on RoboLab-120. That exceeded the best trained robot policy listed on the benchmark, according to the study. But in the direct setup, each model call took an average of 18.56 seconds, and the simulation paused while the model generated its response.
Runway’s published figures point to a different advantage: Praxis-1 is designed to produce actions more quickly. Runway also said that, in its own test on one of its robots, GPT-6 Astra took 15.6 times longer than Praxis-1 to complete three tasks and cost 51 times more per completed task. Those are company-reported results from a limited test, not a broad independent comparison.
For robotics developers, latency and cost matter because machines must respond to changing surroundings, not just select successful actions in a paused simulation. The results do not establish that either approach is better across real-world applications. They do show why companies may pursue both: general models may offer adaptable reasoning, while specialist policies may provide quicker control.
Products Worth Considering
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
Bose QuietComfort Wireless Noise-Cancelling Headphones, Cypress Green
LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
OpenAI’s Robotics Team History
This is the second time OpenAI has ended a stand-alone robotics team. The company wound down an earlier team in 2020. Cofounder Wojciech Zaremba later pointed to limited robot data and the view that other approaches offered a more efficient route to artificial general intelligence. OpenAI rebuilt a robotics team in 2024 and partnered with humanoid-robot maker Figure. Figure said in February 2025 that it would develop its own robot AI, ending that partnership.
The current reorganization follows a different explanation from OpenAI: the company says its general-purpose models have become usable for robotics directly. The broader business environment is also part of the reported backdrop. Fortune noted pressure on OpenAI to control costs and focus on core products, but the company’s stated reason for the change was its assessment of general-purpose models.
The NVIDIA study, published October 14 and identified as a CoRL 2026 late-breaking result, tested GPT-6 Astra on a simulation of 120 tabletop tasks involving a single robot arm. Its agentic setup scored 74.8%; a direct-policy setup scored 66.3%, while an orchestrated setup in which Astra reviewed another model’s proposed movements scored 66.5%. These are simulation results, not proof of performance across physical robots or operating conditions.
“The two approaches are complementary.”
— Runway
Real-World Performance Still Unclear
The reported scores come from different kinds of evidence: NVIDIA’s benchmark is a simulation study, while Runway’s speed and cost comparison is the company’s own test on one of its robots. The available information does not establish how Praxis-1 and GPT-6 Astra compare across the same range of physical tasks, robots, environments, and safety requirements.
It is also unclear how OpenAI will organize its robotics research and hardware work over time, or whether it will publish further detail about the reorganization. OpenAI has not issued a public statement of its own on the change, according to the source reporting. The benchmark figures alone do not show whether general models can meet the response-time needs of physical robots without additional tools or specialist components.
Evidence From Robots in Practice
The next useful milestone will be evidence from comparable tests on physical robots, including task completion, response times, costs, and performance when conditions change. Runway’s reported three-task comparison offers an initial company test, but broader results will be needed to judge how well either approach works beyond its current setup.
For OpenAI, attention will turn to what the reorganized teams build and whether the company shares more about its general-purpose approach to robot control. The reported target for a public demonstration of its humanoid robot is 2027, though the source material does not provide a more specific date or describe what the demonstration will include. Until then, the extent to which OpenAI’s general models can control robots reliably outside simulation remains an open question.
Key Questions
What changed in OpenAI’s robotics organization?
Fortune reported that OpenAI ended its stand-alone robotics team. Its researchers moved into the research organization under Mark Chen, and its applied robotics and hardware teams moved to the Applied organization under Greg Brockman.
What is Runway’s Praxis-1?
Praxis-1 is a 27-billion-parameter robot-action model released with weights and fine-tuning code under the Apache 2.0 license. Runway says it can be used commercially.
Did GPT-6 Astra outperform Praxis-1?
Not in a direct, matched comparison across the evidence reported here. GPT-6 Astra scored 74.8% in an NVIDIA simulation study on RoboLab-120; Praxis-1 scored 46.2% on the benchmark results described in the source. Runway separately reported a speed and cost comparison on one of its robots. The tests do not establish which model performs better across real-world robotics tasks.
Why might a robot need a specialist model?
A specialist model may respond more quickly to control requests. In the reported figures, Praxis-1 took about 0.55 to 0.6 seconds per call for 1.6 seconds of motion, while GPT-6 Astra averaged 18.56 seconds per call in the NVIDIA study’s direct setup. The setups differ, so those numbers should not be treated as a complete head-to-head test.
Source: ThorstenMeyerAI.com
