🔍 Read the full analysis: What Runway's Praxis-1 And OpenAI's Robotics Split Say About AI's Future on ThorstenMeyerAI.com
TL;DR
OpenAI has reorganized its robotics teams, saying general-purpose models can now be applied directly to robot control, while Runway has released the dedicated open-weight model Praxis-1. Published tests show a trade-off: a general model scored higher in one simulation benchmark, while specialist models acted faster. The results do not establish which approach will work best in real-world robots.
OpenAI has broken up its stand-alone robotics team and moved researchers into broader company groups, while Runway has released Praxis-1, an open-weight model designed for robot control. The developments put competing approaches side by side: OpenAI says its general-purpose models can control robots directly, while Runway is betting on a dedicated model built for physical action.
Fortune reported on November 13 that OpenAI confirmed the reorganization after the publication found the company’s CoRL conference booth empty. Robotics researchers now report to Chief Research Officer Mark Chen, while applied robotics and hardware teams have moved into the Applied organization under President Greg Brockman. OpenAI has not published its own announcement about the change.
An OpenAI spokesperson told Fortune the move reflects the company’s belief that 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 general-purpose robots. That is the company’s explanation; the available reporting does not establish that the reorganization will produce a particular robot or product.
Runway released Praxis-1 two days before Fortune’s report. The company describes it as a 27-billion-parameter world action model and makes its weights and fine-tuning code available under the Apache 2.0 license. In the RoboLab-120 results cited in the source material, Praxis-1 scored 46.2%, placing it fifth overall and ahead of the other listed world action and open-source robot models.
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 Skill Remain in Tension
The results suggest that the question is not simply whether a general model or a specialist model is better. The reported tests point to a capability-versus-latency trade-off: a general model can perform well across tasks in simulation, but slow decisions could make it unsuitable for actions that require quick, repeated control. A specialist model may be less capable on the benchmark yet respond fast enough for practical use.
In a study by NVIDIA researchers, an agent using GPT-6 Astra reached a 74.8% success rate on RoboLab-120, compared with 51.0% for the best trained robot policy on that leaderboard. But in the study’s direct setup, each model call averaged 18.56 seconds, and the simulation paused while the model responded. The comparison therefore concerns performance in a particular simulated task set, not a general demonstration of real-world robot capability.
Runway reported a different 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 figures are Runway’s own test results, not an independent evaluation. Together, the findings matter to robotics developers because speed, cost and reliability can matter as much as task success when a model must control a physical machine.
The companies’ approaches need not be mutually exclusive. Runway has called the approaches “complementary,” and the NVIDIA study also tested an arrangement in which Astra reviewed movements proposed by a trained robot model. Such combinations could divide planning and fast movement control between different systems, though the cited results do not establish that this is a settled or proven design.
Popular right now
eufy C30 Robot Vacuum and Mop
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
Ninja Air Fryer XL with MaxCrisp, 6.5 Qt, 6-in-1, Dark Grey (AF181)
As an affiliate, we earn on qualifying purchases.
A Second Robotics Reorganization
OpenAI’s shift follows an earlier change in its robotics work. The company wound down a stand-alone team in 2020, then rebuilt one in 2024 and partnered with humanoid maker Figure. Figure said in February 2025 that it would develop its own robot AI, ending that partnership. The latest reorganization is therefore another change in how OpenAI organizes robotics, rather than the first time it has moved away from a dedicated team.
The company’s explanation has also shifted from an earlier research plan. According to Fortune, OpenAI researchers had expected to build on video models such as Sora, which model how the physical world behaves. The spokesperson said newer general-purpose models, including GPT-6 Astra, demonstrated a direct path to robotics. Fortune also cited cost pressures and a focus on core products as business background, but OpenAI’s stated reason was the applicability of its general models.
The benchmark evidence is limited to specific tests. NVIDIA’s October 14, 2026, study used RoboLab-120, a simulation of 120 tabletop tasks for one robot arm. Its agentic setup used a persistent Codex session with history and additional tools; it was not simply a single model response. Results in simulation and a company’s test on its own robot do not, by themselves, show how either approach performs across varied physical settings.
„The two approaches are “complementary.”“
— Runway
Real-World Performance Is Unsettled
It remains unclear how either approach will perform across real-world environments, where lighting, objects, surfaces and unexpected events vary. The cited RoboLab-120 score comes from simulation, while Runway’s timing and cost comparison comes from a test on one of its own robots. The source material does not provide an independent, like-for-like evaluation across a broad range of hardware and tasks.
Other open questions include how reliably the models can control robots over long periods, what safeguards are needed, and whether the cost and response time can be reduced without hurting performance. OpenAI has not released a public statement detailing the reorganization, and the reported material does not specify a product timeline for the restructured teams. Its spokesperson’s commitments are company statements, not confirmation of a launch date.
Tests and Robot Plans Ahead
The next useful evidence will be further evaluations that compare general and specialist models on the same tasks, hardware and timing conditions, including tests outside simulation. Praxis-1’s open weights and code may allow other researchers and developers to inspect or adapt the model, but the provided source material does not describe a formal independent testing program or a scheduled follow-up.
OpenAI’s stated longer-term plan includes developing a humanoid robot. Fortune reported that CEO Sam Altman had targeted a first public demonstration in 2027; that is a reported target, not confirmation that a demonstration will take place. For now, the reorganized teams’ work, the models’ performance in physical settings and any public product milestones remain to be seen.
Key Questions
What changed in OpenAI’s robotics organization?
Fortune reported that OpenAI moved its robotics researchers under Chief Research Officer Mark Chen and its applied robotics and hardware teams into the Applied organization under President Greg Brockman. OpenAI confirmed the change to Fortune but had not published its own announcement as of November 14, 2026.
What is Runway’s Praxis-1?
Praxis-1 is a 27-billion-parameter model for robot action, released with weights and fine-tuning code under the Apache 2.0 license. Runway’s cited RoboLab-120 result for the model was 46.2%.
Did GPT-6 Astra outperform Praxis-1?
Not in a direct, independent comparison across the same full set of conditions. GPT-6 Astra reached 74.8% in an NVIDIA simulation study’s agentic setup, while Praxis-1 scored 46.2% on the cited RoboLab-120 results. Runway separately reported that Astra took longer and cost more on three tasks on one of Runway’s robots.
Why does model speed matter for robots?
Robots often need repeated control decisions as they move and respond to their surroundings. In the NVIDIA study’s direct setup, Astra calls averaged 18.56 seconds and the simulation paused while the model responded; slower decisions can limit practical use even when task success is high.
When might OpenAI show a humanoid robot?
Fortune reported that CEO Sam Altman had targeted a first public demonstration in 2027. That is a reported target, and the available material does not confirm a scheduled date or guarantee that a demonstration will happen then.
Source: ThorstenMeyerAI.com
