🔍 Read the full analysis: Why Robot-Specific AI Is Back In Focus After OpenAI's Team Split on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI has disbanded its stand-alone robotics team and moved its researchers into broader research and applied groups, saying general-purpose models can now be used directly for robotics. A cited simulation study found GPT-6 Astra handled more benchmark tasks than specialist models, but a separate test showed a dedicated model was much faster and cheaper per task.
OpenAI has broken up its stand-alone robotics team and moved its researchers into broader company groups, telling Fortune the change reflects progress in applying general-purpose AI models to robot control, not a retreat from robotics. The shift comes as Runway releases a robot-specific model and published results highlight a trade-off: general models can perform strongly in simulation, while dedicated models act much faster.
Fortune reported on November 13 that OpenAI confirmed the reorganization after being asked about an empty company booth at the CoRL robotics conference in Austin. The robotics research group now sits under Chief Research Officer Mark Chen; applied robotics and hardware teams moved to the Applied organization, led by President Greg Brockman. OpenAI has not published its own announcement about the change.
According to Fortune, an OpenAI spokesperson said frontier models, including GPT-6 Astra, have shown they can be applied to robotics directly. The spokesperson described robotics as part of the company’s long-term vision and said OpenAI remains committed to bringing general-purpose robots into the world. The source material says CEO Sam Altman has targeted a public demonstration of OpenAI’s humanoid robot in 2027; the reorganization does not, by itself, establish whether that schedule has changed.
The competing approach was on display two days before Fortune’s report: Runway released Praxis-1, a 27-billion-parameter “world action model” with weights and code available under the Apache 2.0 license. In the cited RoboLab-120 benchmark, it scored 46.2%, leading the listed world-action and open-source robot models while placing fifth overall. These figures come from different research and testing settings, so they do not settle which approach works best across real-world robots.
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 Capability Pull Apart
The reorganization is a practical test of whether robotics can be developed as an application of broadly trained AI or still needs models built specifically for physical control. That choice affects how robot systems are trained, how quickly they respond, and the computing costs involved. A general model that handles a wide range of tasks could reduce the need to build a separate system for each robot or setting, but benchmark success alone does not show that it can safely and reliably operate physical machines.
The available results point to a tension rather than a clear winner. In an NVIDIA researchers’ simulation study, a GPT-6 Astra agent achieved 74.8% success on RoboLab-120 in an agentic setup, compared with 51.0% for the best trained robot policy on the cited leaderboard. But the study’s direct setup averaged 18.56 seconds per model call, with the simulation paused while the model processed each request. Runway’s own test on one of its robots found GPT-6 Astra took 15.6 times longer and cost 51 times more per completed task than Praxis-1. That comparison is Runway’s, not an independent evaluation.
Those limitations matter because robots often need to respond to changing conditions in real time. Strong task completion in a paused simulation does not necessarily translate into quick or dependable action outside it. A plausible path is a combined system: a general model could interpret goals or handle unfamiliar situations while a specialist model controls routine movements. Runway described the approaches as “complementary,” but the supplied evidence does not show that such a combined system is already proven in deployment.
Products Worth Considering
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
Bose QuietComfort Wireless Bluetooth Headphones – Active Over‑Ear Noise Cancelling with Mic, USB‑C Charging, Deep Bass, up to 24 h Playtime,
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Cypress Green
As an affiliate, we earn on qualifying purchases.
OpenAI’s Robotics Path So Far
The cited change is not OpenAI’s first retreat from a separate robotics unit. The company wound down an earlier team in 2020. The source material says co-founder Wojciech Zaremba later pointed to limited robot data and a more efficient route to artificial general intelligence through other methods. OpenAI rebuilt a robotics team in 2024 and partnered with humanoid-maker Figure; Figure said the partnership ended in February 2025 as it developed its own robot AI.
OpenAI’s latest reasoning marks a different emphasis. Earlier in 2026, according to the spokesperson’s account to Fortune, researchers expected to build on video models such as Sora, which model how the physical world behaves. The company now says general-purpose models can be applied to robotics directly. Fortune also cited pressure on OpenAI to control costs and focus on core products as business context, but the company’s stated explanation for the move was the progress of its general models.
The benchmark evidence has its own limits. The 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 a single robot arm. It evaluated direct, agentic and orchestrated approaches; the agentic setup scored 74.8%, while the direct and orchestrated setups scored 66.3% and 66.5%. These are simulation results, not a demonstration of comparable performance across physical robots.
“The two approaches are “complementary.””
— Runway
Real-World Performance Still Unclear
It is not yet clear how OpenAI’s new reporting structure will change the staffing, funding or pace of its robotics work. The company has not released a detailed statement, and the source material does not confirm whether its planned 2027 humanoid demonstration remains on schedule.
The performance evidence also leaves open whether either model approach can meet the demands of physical deployment. The NVIDIA result is from a single-arm simulation, and the reported processing delay comes from the study’s direct setup; it should not be treated as a universal response time for every way of using GPT-6 Astra. Runway’s timing and cost comparison covers three tasks on one of its robots and was conducted by the company. Independent comparisons across more tasks, robots and operating conditions are not provided.
It is also unknown whether OpenAI plans to combine general models with specialist control systems, or how it will address latency and cost if general models are used directly. The current evidence establishes a shift in organizational structure and a live technical debate, not a settled outcome for commercial robotics.
Robot Tests Will Shape the Next Phase
The next visible milestone for OpenAI may be its stated target of a public humanoid demonstration in 2027, although no updated schedule was provided in the reports. Before then, the company’s research and product work may offer clues about whether general models are being used for direct robot control or paired with faster specialist policies.
For the field, the key evidence to watch is testing beyond simulated tasks and narrow company-run comparisons: response time, cost, reliability and performance across different robots and environments. OpenAI has not publicly detailed what changes the reorganization will bring, so further information from the company or new independent evaluations will be needed to show whether its model-first strategy can meet the speed demands of physical control.
Key Questions
What did OpenAI change in its robotics organization?
According to Fortune, OpenAI moved robotics researchers into its main research organization under Mark Chen and moved applied robotics and hardware teams to the Applied organization under Greg Brockman. The company confirmed the reorganization to Fortune but has not issued its own public announcement.
Why did OpenAI say it reorganized the team?
An OpenAI spokesperson told Fortune that general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. Fortune also reported cost pressures as business context, but that was not the company’s stated reason for the change.
Does the evidence show that general AI is better for robots?
No. GPT-6 Astra scored 74.8% in an agentic RoboLab-120 simulation, but the test was not a real-world deployment. Runway reported that its dedicated Praxis-1 model was much faster and cheaper than GPT-6 Astra in a three-task test on one of Runway’s robots. The results come from different settings and do not establish a universal winner.
What is Runway’s Praxis-1?
Praxis-1 is a 27-billion-parameter robot-focused “world action model” released under the Apache 2.0 license, with weights and code for fine-tuning. Runway reported a 46.2% score on RoboLab-120; that benchmark figure should be read in the context of the listed models and test setup.
Has OpenAI changed its planned humanoid robot demonstration?
The source material says Sam Altman had targeted a public demonstration in 2027, but the reports do not say whether the reorganization affects that timeline. OpenAI has not provided an updated schedule in the cited coverage.
Source: ThorstenMeyerAI.com
