🔍 Read the full analysis: OpenAI's Robotics Shake-Up And Runway's Praxis-1: What They Signal For AI 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, moving researchers into its main research organization and applied robotics and hardware teams into its Applied organization. The company says general-purpose models can now work directly on robotics; Runway’s new open-weight Praxis-1 offers a contrasting specialist approach, with published results highlighting a trade-off between simulated task success and response speed.
OpenAI has broken up its stand-alone robotics team, moving researchers into its main research group and applied robotics and hardware teams into its Applied organization, according to Fortune, which said the company confirmed the reorganization. The shift comes as Runway releases Praxis-1, an open-weight model designed for robot actions, putting two different approaches to robotics AI in focus: general-purpose models adapted for control and specialist models built for it.
Fortune reported November 13 that OpenAI’s robotics researchers now sit under Chief Research Officer Mark Chen, while its applied robotics and hardware teams have moved to the Applied organization under President Greg Brockman. The company confirmed the change to Fortune when asked. An OpenAI spokesperson characterized the move as a sign of continued commitment to robotics, not a retreat. OpenAI had not published its own announcement as of November 14, according to the source material.
OpenAI’s stated rationale is that general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robot control. That differs from the approach researchers expected earlier in the year, when they anticipated building on video models such as Sora. The company’s spokesperson told Fortune that robotics remains part of OpenAI’s long-term vision and that it remains committed to general-purpose robots.
Runway released Praxis-1 two days before Fortune’s report. The company describes it as a 27-billion-parameter “world action model.” Its weights and fine-tuning code are available under the Apache 2.0 license, which permits commercial use. In the published RoboLab-120 results cited in the source material, Praxis-1 scored 46.2%, leading the listed world-action and open-source robot models and ranking fifth overall.
A separate NVIDIA Seattle Robotics Lab study, published October 14 as a CoRL 2026 late-breaking result, tested GPT-6 Astra on RoboLab-120, a simulation of 120 tabletop tasks for one robot arm. Its agentic setup achieved a 74.8% success rate, above the 51.0% reported for the leaderboard’s best trained robot policy. In the direct setup, however, each Astra call took an average of 18.56 seconds; the simulation paused while the model processed its response.
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
Two Routes to Robot Control
The developments sharpen a practical question for robotics developers: should a robot use a broad AI model that can reason across many tasks, or a specialist model designed to issue actions quickly? The available figures suggest a trade-off, not a clear winner. Astra’s higher simulated success rate in one setup does not by itself show that it can control a physical robot at a useful speed. Praxis-1’s faster action cycle, meanwhile, does not establish that it will match a general-purpose model across a wider range of tasks.
That gap matters because robots must respond to changing conditions while operating in the physical world. A model that performs well in a paused simulation may still face latency, reliability, and cost limits when controlling real equipment. Conversely, a faster specialist may need support when a task requires broader planning or interpretation. The published comparisons make speed, task range, and operating cost central questions for companies deciding how to build robot systems.
The organizational change is also a signal about how OpenAI sees the work: rather than keep robotics as a separate research unit, it is placing the researchers alongside broader model development and the applied organization. That structure may help connect model research with hardware and deployment, but the reorganization alone does not establish whether OpenAI’s robots will reach the market sooner or perform better.
Popular right now
Dyson V15 Detect Cordless Vacuum
De’Longhi Classic Matte Black Espresso Machine with Milk Frother
KitchenAid Artisan 5-Qt Tilt-Head Stand Mixer, KSM150PS, Almond Cream
As an affiliate, we earn on qualifying purchases.
OpenAI’s Robotics Reorganization
The robotics move follows an earlier cycle at OpenAI. The company wound down its first stand-alone robotics team in 2020. It rebuilt a team in 2024 and partnered with humanoid-maker Figure; Figure said the partnership ended in February 2025 as it moved to develop its own robot AI. Fortune also described cost controls and a narrower product focus as part of OpenAI’s business backdrop, but the reason the company gave for this latest change was the progress it sees in general-purpose models.
OpenAI has said it is developing a humanoid robot, and CEO Sam Altman has targeted a public demonstration in 2027. That is a stated target, not confirmation that a demonstration is scheduled or that a commercial robot will be ready then. Runway’s release, by contrast, makes a specialist robot model’s weights and code available for outside use. The two developments are related by the broader debate over model design, but the source material does not establish a direct partnership or competition between the companies.
From Simulation to Real Robots
The benchmark results are simulation findings, not evidence that either model is ready for general deployment on physical robots. The source material does not provide a direct, independently verified comparison of Astra and Praxis-1 across the same tasks, hardware, and operating conditions. It does cite a Runway test on one of its robots in which 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, and details sufficient to generalize beyond that test are not provided here.
OpenAI has not publicly detailed the new reporting structure or published a statement of its own. It is also unclear how the change affects staffing, research priorities, or the humanoid robot effort. The available results do not settle whether general-purpose models can meet real-time control needs, or whether specialist models can provide comparable performance across a broad set of tasks.
Evidence From Deployment
The next useful evidence will come from more transparent comparisons that report task success alongside response time, cost, and hardware conditions, including tests outside simulation. Further details from OpenAI about its reorganization and robot plans could clarify how research and applied teams will work together. OpenAI’s previously stated target for a public humanoid demonstration is in 2027, though no specific date or demonstration details are confirmed in the source material. Until then, claims about either approach should be judged against the limits of the tests reported so far.
Key Questions
What changed in OpenAI’s robotics organization?
Fortune reported that OpenAI ended its stand-alone robotics team structure. Robotics researchers moved under Chief Research Officer Mark Chen, while applied robotics and hardware teams moved into the Applied organization under President Greg Brockman.
Why did OpenAI reorganize its robotics teams?
OpenAI told Fortune that general-purpose models, including GPT-6 Astra, have become usable for robotics directly. The company framed the reorganization as part of its continued commitment to robotics; it has not published a separate statement on the change.
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 permits commercial use.
Does the benchmark show that GPT-6 Astra is better than Praxis-1?
No direct, fully matched comparison is provided. The cited NVIDIA simulation study reports a 74.8% success rate for Astra’s agentic setup, while Praxis-1 scored 46.2% on the benchmark. The results come from different reported evaluations and do not settle performance on physical robots.
What is the main limitation of the reported results?
The headline success-rate result for Astra comes from simulation, where the environment paused while the model processed actions. Its direct setup averaged 18.56 seconds per call. More testing is needed to show how the models perform in real-world conditions, including speed, reliability, and cost.
Source: ThorstenMeyerAI.com
