🔍 Read the full analysis: A New Robotics AI Debate Emerges From OpenAI's Team Split And Praxis-1 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
OpenAI confirmed to Fortune that it has disbanded its stand-alone robotics team, moving researchers and applied robotics staff into other company groups. The change comes as Runway releases Praxis-1, a downloadable specialist robot model: published results suggest general models can perform well in simulation, while specialist models act faster.
OpenAI has broken up its stand-alone robotics team and moved its researchers and applied robotics staff into other parts of the company, Fortune reported November 13, with OpenAI confirming the change to the outlet. The reorganization coincides with Runway’s release of Praxis-1, a specialist robot model, renewing a practical question for the field: whether general-purpose AI can control robots effectively or dedicated models are still needed for fast, physical action.
According to Fortune, OpenAI moved its robotics researchers into its main research organization under Chief Research Officer Mark Chen. Its applied robotics and hardware teams were moved to the Applied organization under President Greg Brockman. Fortune reported that an OpenAI spokesperson described the reorganization as a sign of commitment to robotics, not a retreat. OpenAI had not published its own announcement about the change as of November 14, according to the supplied reporting.
The explanation the spokesperson gave Fortune was that general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. An earlier research plan had looked to build on video models such as Sora. That explanation is the company’s stated rationale; the available reporting does not establish whether other business considerations also influenced the decision.
Praxis-1, released by Runway two days before Fortune’s report, is a 27-billion-parameter world action model. Runway made its weights and fine-tuning code available under the Apache 2.0 license. On the RoboLab-120 results cited in the source material, Praxis-1 scored 46.2%, the highest among the listed world action models and open-source robot models, and fifth overall. Runway says a call takes about 0.55 to 0.6 seconds for 1.6 seconds of motion. These figures are reported benchmark results, not proof of performance across real-world robots and settings.
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 Versus General Capability
The organizational change and model release highlight a design trade-off, rather than a settled contest between two approaches. General-purpose models may handle varied tasks without robot-specific training, while specialized systems may respond quickly enough for continuous control. For robots acting in physical environments, delay is not just a benchmark detail: a model that takes many seconds to choose a movement may be less useful for tasks requiring prompt adjustments.
The distinction could shape how robotics companies allocate research and computing resources. If general models can take on more robot tasks, companies may rely less on separate model teams. But the evidence presented so far is limited: the strongest general-model result cited comes from simulation, while the speed comparison from Runway was conducted on one of its own robots and across three tasks. Neither result alone establishes broad real-world superiority.
OpenAI’s spokesperson framed the move as a way to apply the company’s general models to robotics. Runway, by contrast, has released a specialized model with downloadable weights and commercial-use terms. That makes Praxis-1 available for outside adaptation, while the OpenAI results described here concern a model tested through a study rather than an equivalent open release.
Products Worth Considering
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Bose TV Speaker (2nd Gen) – Compact Soundbar with Bluetooth and HDMI‑ARC Connectivity, Clear Dialogue, Dolby Atmos, Deep Bass, Remote Contro
LG UltraGear evo 27-inch 5K Mini-LED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How OpenAI Reached This Structure
OpenAI has shifted its approach to robotics before. It wound down an earlier stand-alone team in 2020. The source material says co-founder Wojciech Zaremba later pointed to limited robot data and a more efficient path to artificial general intelligence through other methods. OpenAI rebuilt a robotics 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.
The latest change became visible at the Conference on Robot Learning in Austin. Fortune reported that OpenAI’s booth remained listed on the event floor plan, but was empty when the outlet visited. The timing suggests the reorganization was recent; the empty booth by itself does not explain the decision.
The research cited in the debate was published by researchers at NVIDIA’s Seattle Robotics Lab on October 14, 2026, as a CoRL 2026 late-breaking result. In the study, GPT-6 Astra received camera images, robot information and a control interface, then selected movements without robot-specific training. RoboLab-120 is a simulation benchmark of 120 tabletop tasks for a single robot arm. The test setting matters: simulated success rates are not direct evidence of equivalent results on physical hardware.
What the Results Cannot Establish
The published comparisons leave several limits. In NVIDIA’s simulation study, GPT-6 Astra’s agentic setup achieved a 74.8% success rate, compared with 51.0% for the best trained robot policy on the cited leaderboard. But the supplied material says a direct model call averaged 18.56 seconds, and the simulation paused while the model was processing. The reported success advantage therefore does not settle whether the model can act quickly or reliably on physical robots.
Runway reported that, on one of its robots and across three tasks, GPT-6 Astra took 15.6 times longer than Praxis-1 to complete them and cost 51 times more per completed task. That comparison is attributed to Runway and covers a narrow test; the source does not provide enough detail to generalize it to other hardware, tasks or operating conditions. The exact basis and conditions behind every comparison are not fully set out in the supplied material.
It is also unclear how OpenAI will organize robotics work in practice, what research or hardware programs will continue, and whether it plans a public demonstration of its own humanoid robot as previously targeted by CEO Sam Altman for 2027. Fortune mentioned pressure on OpenAI to control costs and focus on core products as business background, but the stated reason for the move was the applicability of general-purpose models. The available reporting does not confirm cost pressure as the cause.
Signals to Watch in 2027
The next useful evidence will be results on physical robots that report both task success and response time under comparable conditions. Researchers and users will also need clearer comparisons across multiple robot platforms and tasks to judge whether the simulation gap carries over to real-world use. Praxis-1’s open weights and fine-tuning code may allow outside groups to test and adapt the model, but the source material does not report independent follow-up evaluations yet.
For OpenAI, the next indicators are how the redistributed teams operate and whether the company publishes more detail about its robotics work. The supplied reporting notes that Altman had targeted a first public demonstration of an OpenAI humanoid robot in 2027; it does not confirm that a demonstration is scheduled or that the target remains unchanged. Further company statements, research releases and hardware demonstrations could clarify whether the reorganization leads to a different development path.
Key Questions
What changed in OpenAI’s robotics organization?
Fortune reported that OpenAI moved its robotics researchers into the main research organization under Mark Chen and its applied robotics and hardware teams into the Applied organization under Greg Brockman. OpenAI confirmed the reorganization to Fortune.
What is Praxis-1?
Praxis-1 is Runway’s 27-billion-parameter world action model for robot control. Runway released its weights and fine-tuning code under the Apache 2.0 license.
Did GPT-6 Astra outperform Praxis-1?
The cited results do not provide a single like-for-like verdict. GPT-6 Astra scored 74.8% in an agentic setup on the RoboLab-120 simulation benchmark, while Praxis-1 scored 46.2% on the reported leaderboard. Runway separately reported a speed and cost comparison on one of its robots across three tasks. The different setups and limited scope make broad conclusions uncertain.
Why might a specialist robot model still be useful?
The figures in the source material suggest a potential speed advantage: Runway says Praxis-1 takes about 0.55 to 0.6 seconds per call for 1.6 seconds of motion. GPT-6 Astra’s calls averaged 18.56 seconds in the direct setup of the cited simulation study. These timings come from different reported settings and should not be treated as a universal comparison.
Has OpenAI said it is ending work on robotics?
No. The company told Fortune that robotics remains part of its long-term vision and that it remains committed to general-purpose robots. The reported change is a move of teams into other OpenAI organizations, not a confirmed shutdown of all robotics work.
Source: ThorstenMeyerAI.com
