🔍 Read the full analysis: Runway Opens Praxis-1: Could OpenAI's Robotics Split Reflect A Wider Shift? on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Fortune reported that OpenAI has disbanded its standalone robotics team, moving researchers into its main research group and applied robotics and hardware teams into its Applied organization. Runway released Praxis-1, an open-weight robotics model, as published tests suggest general models can succeed on more simulated tasks while specialist models act much faster. OpenAI says the reorganization reflects its belief that general-purpose models can be applied directly to robotics.
OpenAI has disbanded its standalone robotics team and moved its researchers and applied robotics staff into other company groups, Fortune reported November 13, while Runway released Praxis-1, an open-weight model built for robot actions. The moves frame a live technical question: whether general-purpose AI can control robots effectively, or whether dedicated models remain necessary for speed and cost.
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. The change was not announced in a public OpenAI statement, according to the source material. Fortune found the company’s previously listed booth empty at the CoRL robotics conference in Austin.
An OpenAI spokesperson told Fortune that general-purpose frontier models, including GPT-6 Astra, have shown they can be used directly for robotics. The spokesperson described robotics as part of the company’s long-term vision and said OpenAI remains committed to bringing general-purpose robots into the world. Fortune reported that the company is developing a humanoid robot, with CEO Sam Altman targeting a public demonstration in 2027.
Praxis-1 takes the specialist route. Runway describes it as a 27-billion-parameter “world action model” and released its weights and fine-tuning code under the Apache 2.0 license. On RoboLab-120, a simulated benchmark of 120 tabletop tasks, it scored 46.2%, the highest result among the world action models and open-source robot models listed in the material, and fifth overall. Runway 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
Speed and Capability Pull Apart
The results point to a practical trade-off rather than a clear winner. In a simulation study by NVIDIA researchers, an agent using GPT-6 Astra achieved a 74.8% success rate on RoboLab-120 in an agentic setup, above the listed best trained robot policy at 51.0%. But the study’s direct setup averaged 18.56 seconds per model call, with the simulation paused while the model reasoned. A high task score in that setting does not establish that the model can control a physical robot at useful speed.
Runway reported a separate comparison 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 Runway’s results from its own test, not an independent comparison across robots or operating conditions. Still, the figures show why speed, inference cost and reliability matter alongside benchmark success. A robot that reasons well but responds slowly may be unsuitable for tasks requiring continuous movement.
For companies building robotic systems, the distinction could shape where investment goes: into broadly capable models, specialized real-time policies, or systems that combine both. Open weights also give developers access to Praxis-1 for use and modification under its license. The published figures do not settle which approach will work best in homes, factories or other real-world settings.
Products Worth Considering
Apple Watch Series 12 GPS 42mm
Bose TV Speaker (2nd Gen) – Compact Soundbar with Bluetooth and HDMI‑ARC Connectivity, Clear Dialogue, Dolby Atmos, Deep Bass, Remote Contro
Samsung 32-inch Odyssey G5 G51F QHD 180Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
OpenAI’s Robotics Path
OpenAI’s decision changes the organizational home of its robotics work, but the source material does not establish that the company has abandoned robotics. Its spokesperson framed the shift as a response to progress in general-purpose models. The company’s earlier approach, according to Fortune, was to build on video models such as Sora, on the premise that models that represent how the physical world behaves could support robot control.
This is the second time OpenAI has closed a standalone robotics team. The company wound down its first team in 2020; co-founder Wojciech Zaremba later cited limited robot data and a more efficient route to artificial general intelligence through other methods. OpenAI rebuilt a robotics team in 2024 and partnered with humanoid-robot maker Figure. Figure said it would develop its own robot AI after that partnership ended in February 2025.
The benchmark evidence has limits. NVIDIA researchers tested Astra in simulation on a single robot arm and a set of tabletop tasks, not across a range of physical robots or environments. They evaluated direct control, an agentic setup with persistent history and tools, and an orchestrated setup in which Astra reviewed another policy’s proposed actions. The 74.8% result came from the agentic setup; it should not be treated as a general measure of performance in the physical world.
“The two approaches are complementary.”
— Runway
Real-World Performance Remains Open
It remains unclear whether GPT-6 Astra’s simulation results will translate into reliable, fast physical control, or how performance would change with different robots, tasks and operating conditions. The benchmark result and Runway’s robot test use different setups, so they do not provide a direct, independently verified head-to-head comparison.
OpenAI has not publicly detailed the new reporting structure or explained how the reorganization will affect its humanoid robot plans. The source material also does not provide independent confirmation of Runway’s time and cost comparison, or enough detail to establish how representative its three-task test is. Runway’s open release terms are specified as Apache 2.0, but the longer-term support and deployment plans for Praxis-1 are not stated.
Watch for Physical Robot Tests
The next useful evidence will be results from real-world robot deployments that report speed, task success, safety and cost under comparable conditions. Further testing could clarify whether general models can handle responsive control directly, or whether they work best when paired with specialist policies such as Praxis-1.
OpenAI’s stated target for a first public humanoid demonstration is 2027. The source material gives no date for additional details on the reorganization or for a specific Praxis-1 deployment milestone. Until those results arrive, the published evidence supports a distinction between broad simulated capability and faster specialist action, not a settled verdict on which model strategy will dominate.
Key Questions
What did OpenAI change?
Fortune reported that OpenAI ended its standalone robotics team structure. Researchers moved into the main research organization under Mark Chen, while 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 model for robot actions. Runway released its weights and fine-tuning code under the Apache 2.0 license.
Which approach performed better on RoboLab-120?
In the cited simulation results, the GPT-6 Astra agentic setup scored 74.8%, while Praxis-1 scored 46.2%. The tests and setups differ, and simulation scores do not establish performance on physical robots.
Why might a specialist model still be useful?
Praxis-1’s reported calls took about 0.55 to 0.6 seconds, while an NVIDIA study’s direct GPT-6 Astra setup averaged 18.56 seconds per call. Faster responses may matter for real-time movement, though the figures come from different tests.
Has OpenAI said it is leaving robotics?
No. The spokesperson told Fortune that robotics remains part of OpenAI’s long-term vision and described the organizational changes as a response to the capabilities of its general-purpose models. OpenAI had not posted a public statement about the reorganization in the source material.
Source: ThorstenMeyerAI.com
