🔍 Read the full analysis: What The OpenAI Robotics Split And Runway Praxis-1 Could Mean For AI on ThorstenMeyerAI.com
Get audio and creator gear 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 its researchers into broader research and applied groups. The company says general-purpose models can now be used directly for robot control, while Runway’s open Praxis-1 results point to a continuing role for specialist models that act faster. The available tests suggest the approaches may complement each other, but comparisons are limited and do not settle how well either works on physical robots.
OpenAI has broken up its stand-alone robotics team and moved its researchers into broader company groups, while Runway has released Praxis-1, an open-weight model built specifically for robot action. The developments sharpen a practical question for robotics: whether general-purpose AI models can take over robot control, or whether specialist systems remain necessary for fast, reliable movement.
OpenAI confirmed the reorganization after Fortune asked about it, according to the outlet’s November 13 report. Its robotics researchers now sit within the main research organization under Chief Research Officer Mark Chen. Applied robotics and hardware teams have moved to the Applied organization under President Greg Brockman. The change came after OpenAI had registered a booth at the CoRL robotics conference in Austin; Fortune found the booth empty.
An OpenAI spokesperson told Fortune that the change reflects the company’s view that general-purpose frontier models, including GPT-6 Astra, can be applied directly to robotics. 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. OpenAI has not published its own announcement about the reorganization.
Runway released Praxis-1 two days before Fortune’s report. The company describes it as a 27-billion-parameter “world action model” and released 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%, the highest result among the world action models and open-source robot models listed, and fifth overall. Runway says it takes about 0.55 to 0.6 seconds per call to generate 1.6 seconds of motion.
A separate study by researchers at NVIDIA’s Seattle Robotics Lab tested GPT-6 Astra in simulation as a robot policy, using camera images, robot information and a control interface. In an agentic setup, Astra achieved a 74.8% success rate across RoboLab-120’s 120 simulated tabletop tasks. The researchers said that result was 23.8 percentage points above the benchmark’s leading policy, Lakeside. A direct setup scored 66.3%, while a version that used Astra to review another policy’s proposed movements scored 66.5%.
The higher simulated success rate came with a substantial speed limitation: in the direct setup, each model call took an average of 18.56 seconds, and the simulation paused while Astra processed its response. In a test on one of Runway’s robots, Runway said GPT-6 Astra took 15.6 times longer than Praxis-1 to complete three tasks and cost 51 times more per completed task. That result is Runway’s own comparison, not an independent evaluation.
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 Control Faces a Speed Gap
The results point to a trade-off between general capability and real-time control. A model that succeeds across a broad set of simulated tasks may be useful for planning or handling unfamiliar situations, but a robot also has to respond quickly enough to act safely and complete tasks in the physical world. The NVIDIA study’s long pauses per call make its strong simulation result different from a demonstration of fast, continuous robot control.
Praxis-1’s faster action generation, meanwhile, does not establish that it is more capable overall. Its benchmark score was below Astra’s best reported score, and the available robot comparison comes from Runway. The two releases therefore do not show that one model class has won. They suggest a possible division of labor: a general model could plan or interpret a task, while a faster specialist handles movement. Runway itself has described the approaches as complementary.
For companies and researchers, the distinction affects what they build and measure. Open weights and commercial licensing make Praxis-1 available for developers to test and adapt, while OpenAI’s reorganization places robotics work closer to its general-model teams. But benchmark success, inference speed, cost and performance on real hardware remain separate measures; improvements on one do not automatically establish improvements on the others.
Products Worth Considering
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
OpenAI’s Robotics Path So Far
The reorganization follows earlier changes in OpenAI’s robotics work. The company wound down a stand-alone robotics team in 2020. It rebuilt a 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. Fortune reported that OpenAI has also been developing a humanoid robot, with CEO Sam Altman targeting a first public demonstration in 2027.
OpenAI’s latest explanation marks a shift from an earlier research expectation. According to the spokesperson’s account to Fortune, researchers initially expected to build on video models such as Sora, whose training involves modeling how the physical world behaves. The company now believes frontier models such as GPT-6 Astra can be applied to robotics more directly. Fortune also described pressure on OpenAI to control costs and focus on core products, but the company’s stated reason for the team move was its assessment of general-purpose models.
The NVIDIA result cited in coverage is a simulation study published October 14 and labeled a CoRL 2026 late-breaking result. RoboLab-120 tests one robot arm on 120 tabletop tasks. Its scores measure task success in that environment; they do not by themselves show how reliably a model performs across different robots, physical settings or extended use.
“The two approaches are complementary.”
— Runway
Simulation Is Not a Robot Trial
Real-world performance remains unclear. The cited NVIDIA results come from a simulator, and the source material does not provide an independent, matched test of GPT-6 Astra and Praxis-1 across a range of physical robots and tasks. The Runway comparison covers three tasks on one of its robots and was conducted by the company itself.
It is also not yet clear how OpenAI will organize or fund robotics research after the move, whether it will continue developing its own humanoid hardware, or how the shift affects the target for a public robot demonstration in 2027. OpenAI has not posted a detailed public explanation, and the reported benchmark results do not establish that general-purpose models can replace specialist control systems in everyday deployments.
Watch for Hardware and Evaluation
The next useful evidence will be independent, like-for-like testing that reports task success, response time and cost on the same hardware and tasks. Results from physical robots will also be needed to show whether simulated performance carries over to real settings, where delays and mistakes can have different consequences.
OpenAI’s stated longer-term plan includes a targeted public demonstration of a humanoid robot in 2027, though the reorganization’s effect on that schedule is not known. Developers can also evaluate Praxis-1 using its released weights and code. Until further results arrive, the clearest conclusion is limited: general models have posted a higher score in one simulation benchmark, while a specialist model has shown faster action generation in the reported tests.
Key Questions
What did OpenAI change?
According to Fortune, OpenAI ended its stand-alone robotics team structure. Robotics 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 designed for robot action. Runway released its weights and fine-tuning code under the Apache 2.0 license, allowing commercial use.
Did GPT-6 Astra outperform Praxis-1?
They were reported in different tests. Astra scored 74.8% in an agentic RoboLab-120 simulation, while Praxis-1 scored 46.2% on the same benchmark. Runway separately reported that Praxis-1 completed three tasks on one of its robots faster and at lower cost than Astra. These figures do not amount to a broad independent comparison.
Can general-purpose AI models control robots in real time?
The available results do not settle that question. In the NVIDIA study’s direct setup, Astra took an average of 18.56 seconds per call, while Runway reported Praxis-1 could generate 1.6 seconds of motion in about 0.55 to 0.6 seconds. Performance in broader real-world conditions remains unclear.
Does OpenAI’s move mean it is leaving robotics?
OpenAI’s spokesperson told Fortune that robotics remains part of the company’s long-term vision and said it is committed to general-purpose robots. The company has not published a detailed statement explaining how the reorganization will affect its robotics programs.
Source: ThorstenMeyerAI.com
