🔍 Read the full analysis: Praxis-1, OpenAI, And The Question Of Building AI For Robots on ThorstenMeyerAI.com
Get the latest gadgets 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 moved researchers into its broader research and applied organizations, telling Fortune that general-purpose models can now be used directly for robotics. Runway released Praxis-1, an open-weight robot model, as benchmark results point to a trade-off: general models performed better on a simulation benchmark, while Praxis-1 acted much faster. The available evidence does not establish which approach works best on physical robots at scale.
OpenAI has broken up its stand-alone robotics team and moved its researchers into other parts of the company, while Runway has released Praxis-1, an open-weight model designed for robot control. The changes put a practical question at the center of this week’s robotics conference, CoRL: whether robots can use general-purpose models directly or still need specialized systems that respond faster. OpenAI confirmed the reorganization to Fortune; the available benchmark results show different strengths, not a clear winner for real-world deployment.
According to Fortune, OpenAI moved its robotics researchers into its main research team under Chief Research Officer Mark Chen. Its applied robotics and hardware teams moved to the Applied organization, which reports to President Greg Brockman. Fortune reported the change on November 13 after finding the company’s registered CoRL booth empty. OpenAI has not issued a public statement about the reorganization, according to the source material.
An OpenAI spokesperson told Fortune that the change reflects a shift in how the company approaches robotics: its general-purpose frontier models, including GPT-6 Astra, have shown they can be applied to robot control directly. The spokesperson described robotics as part of OpenAI’s long-term vision and said the company remains committed to bringing general-purpose robots into the world. The spokesperson’s explanation is distinct from Fortune’s reporting about broader pressure on OpenAI to control costs; the company gave the model-capability shift as its reason.
Runway released Praxis-1 two days before Fortune’s report. The 27-billion-parameter model is described as a “world action model” and is offered with weights and fine-tuning code under the Apache 2.0 license. In the reported RoboLab-120 results, GPT-6 Astra’s agentic setup scored 74.8%, while Praxis-1 scored 46.2%. The same results also show a speed gap: Astra’s model calls averaged 18.56 seconds in the study’s direct setup, while Praxis-1 took about 0.55 to 0.6 seconds per call 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
The Trade-Off Between Generality and Speed
The comparison matters because a robot must do more than choose a plausible action: it must respond quickly enough to act in a changing physical environment. Higher simulated task success may show that a general model can reason across a range of tasks, but long delays between decisions can limit its usefulness for live control. By contrast, a model that returns actions quickly may be easier to use for immediate movement, even if its benchmark score is lower.
The reported numbers are not a controlled, like-for-like contest across every condition. The NVIDIA study tested GPT-6 Astra in several simulation setups, including an agentic configuration with persistent history and extra tools. Runway’s reported comparison on one of its robots measures task completion time and cost for three tasks. That test found GPT-6 Astra took 15.6 times longer than Praxis-1 to complete them and cost 51 times more per completed task, according to Runway. Those figures are company-reported results from a limited test, not evidence of performance across all robots or tasks.
For developers, researchers and companies choosing how to build robot systems, the practical question may not be which model category wins outright. The results leave room for hybrid systems, in which a general model handles planning or unusual situations and a faster specialist handles continuous movements. Runway has described the approaches as complementary, but their relative roles in deployed systems remain to be established.
Products Worth Considering
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How OpenAI Reached This Reorganization
OpenAI had previously built a separate robotics team, wound it down in 2020, and rebuilt one in 2024. It partnered with humanoid maker Figure, but Figure said in February 2025 that it would develop its own robot AI, ending that partnership. The source material says OpenAI is also developing a humanoid robot, with CEO Sam Altman targeting a public demonstration in 2027; that target is a plan, not a confirmed demonstration date.
Earlier in 2026, OpenAI researchers expected to build on video models such as Sora, according to the explanation given to Fortune. The company now says general-purpose models can be applied to robotics directly, prompting the move into broader research and applied groups. Separately, NVIDIA researchers published their RoboLab-120 study on October 14 as a CoRL 2026 late-breaking result. The benchmark simulates 120 tabletop tasks for a single robot arm, so its scores describe simulated performance rather than a broad record of robots operating in homes, factories or public spaces.
“The two approaches are complementary.”
— Runway
What the Tests Cannot Yet Show
The published results do not settle whether a general-purpose model or a specialist model is more effective on physical robots in everyday conditions. RoboLab-120 is a simulation benchmark, and the NVIDIA study’s reported call time comes from a particular setup in which the simulation paused while the model processed its response. Runway’s comparison involved three tasks on one of its robots; the source material does not provide enough detail to treat that test as a general performance measure.
It is also unclear how OpenAI will organize robotics work after the change, whether it will continue developing its own humanoid hardware, or how its models would be paired with specialist control systems. OpenAI has not published a statement of its own about the reorganization, and the source material does not provide detailed information about the costs, task conditions or repeatability behind each reported result. The 2027 public demonstration target remains a stated goal, not a confirmed event.
Evidence From Robots in Practice
The next useful evidence will come from tests that compare models under clearly described, repeatable conditions on physical robots, including how often tasks succeed, how quickly systems respond and what each completed task costs. Further results from the research teams and robot developers could clarify whether general models can maintain their simulated capabilities outside the benchmark.
For OpenAI, readers will be watching for more detail on the new reporting structure and progress toward the company’s stated humanoid-robot demonstration target in 2027. For Runway, Praxis-1’s open weights and fine-tuning code give researchers and developers a way to test the specialist model, but its broader performance and adoption are not yet established. Until those results arrive, the evidence supports a trade-off between task performance and response time rather than a decisive answer on which approach robots need.
Key Questions
What changed in OpenAI’s robotics organization?
According to Fortune, OpenAI moved its robotics researchers into the main research team under Mark Chen and its applied robotics and hardware teams into the Applied organization under Greg Brockman. The company confirmed the reorganization to Fortune.
What is Praxis-1?
Praxis-1 is Runway’s 27-billion-parameter model for robot actions. Runway released its weights and fine-tuning code under the Apache 2.0 license.
Which model scored higher on RoboLab-120?
In the reported results, GPT-6 Astra’s agentic setup scored 74.8%, while Praxis-1 scored 46.2%. RoboLab-120 is a simulation benchmark, and the figures do not establish which model will perform better across physical robots and tasks.
Why might a robot use a specialist model if a general model scores higher?
The reported tests point to a speed difference. Astra’s calls averaged 18.56 seconds in the NVIDIA study’s direct setup, while Praxis-1 took about 0.55 to 0.6 seconds per call for 1.6 seconds of motion. Fast response can matter for controlling movement, though results depend on the setup and task.
Has OpenAI confirmed when it will demonstrate a humanoid robot?
The source material says CEO Sam Altman targeted a public demonstration in 2027. That is a stated target, not confirmation of a scheduled event or guarantee that it will take place then.
Source: ThorstenMeyerAI.com
