🔍 Read the full analysis: Does Robotics Need Its Own AI? OpenAI Splits Its Team As Runway Opens Praxis-1 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 confirmed to Fortune that it has reorganized its stand-alone robotics team, moving researchers into its main research group and applied robotics and hardware staff into its Applied organization. Runway released the open-weight Praxis-1 robot model two days earlier; available benchmark results point to a trade-off between general models’ simulated task success and specialist models’ speed.
OpenAI has disbanded its stand-alone robotics team and moved its researchers and applied robotics staff into other parts of the company, while Runway has released Praxis-1, a dedicated robot model with downloadable weights. The shift comes as AI moves into the physical world. The developments, reported by Fortune and confirmed by OpenAI to the outlet, put a practical question in focus: whether robots can rely on general-purpose AI models or need systems designed specifically for real-time control.
According to Fortune’s November 13 report, OpenAI’s robotics researchers now sit within the company’s main research organization under Chief Research Officer Mark Chen. Its applied robotics and hardware teams have moved to the Applied organization, led by President Greg Brockman. Fortune reported that OpenAI confirmed the changes when asked. The company has not published its own announcement, and its booth at the CoRL robotics conference in Austin was empty when Fortune visited, despite appearing on the event floor plan.
An OpenAI spokesperson told Fortune that the reorganization reflects a change in how the company approaches robotics, not a withdrawal from the field. The broader capabilities of AI agents are also evolving, as shown by Claude building its own team of agents. The spokesperson said general-purpose frontier models, including GPT-6 Astra, had shown they could be applied to robotics directly. Runway’s Praxis-1 takes a specialist route: it is a 27-billion-parameter world action model released with weights and fine-tuning code under the Apache 2.0 license, which permits commercial use under its terms.
The reported benchmark results show different strengths, not a straightforward winner. This also highlights the risks explored in how AI could turn against its own device. In an NVIDIA researchers’ simulation study, an agent built on GPT-6 Astra achieved 74.8% success on RoboLab-120, a benchmark of 120 tabletop tasks, compared with 51.0% for the highest-ranked trained robot policy cited in the source material. But calls to Astra averaged 18.56 seconds in the direct setup, and the simulation paused while it generated a response. Praxis-1 scored 46.2% on the same benchmark and was the highest-scoring world action model listed; the source reports roughly 0.55 to 0.6 seconds per call for 1.6 seconds of motion.
Runway also reported a company test on one of its robots in which GPT-6 Astra took 15.6 times longer than Praxis-1 to complete three tasks and cost 51 times more per completed task. That comparison is Runway’s own result, not an independent evaluation. The available material does not provide enough detail about the task setup, measurement methods or repeat trials to treat it as a general performance comparison.
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 Set Competing Demands
Robots must interpret their surroundings and choose movements quickly enough to act safely and reliably. The NVIDIA simulation results suggest that a general-purpose model can handle a broad set of tasks without robot-specific training, but the reported response times raise a practical concern: a robot may wait too long for a decision. A specialist model such as Praxis-1 appears designed to produce actions more quickly, though its lower RoboLab-120 score does not establish how it would perform across other tasks or real-world settings.
The distinction matters to companies building robots and to developers choosing what technology to adapt. A general model could reduce the need to train separate systems for each robot or task; a faster specialist could be more suitable for immediate control. Neither possibility is settled by the results supplied here. OpenAI’s reorganization signals that it intends to pursue robotics through its broader model and product groups, while Runway’s release gives developers access to a purpose-built alternative. The comparison may ultimately be less about choosing one model than deciding which jobs belong to each.
Products Worth Considering
Samsung 32-inch Odyssey G5 G51F QHD 180Hz Gaming Monitor
Bose QuietComfort Wireless Bluetooth Headphones – Active Over‑Ear Noise Cancelling with Mic, USB‑C Charging, Deep Bass, up to 24 h Playtime,
DJI Mic 3 (2 TX + 1 RX + Charging Case) Wireless Microphone
As an affiliate, we earn on qualifying purchases.
OpenAI’s Changing Robotics Strategy
OpenAI previously closed a stand-alone robotics team in 2020. The company rebuilt a team in 2024 and partnered with humanoid-robot maker Figure; Figure said the partnership ended in February 2025 as it moved to develop its own robot AI. Fortune reported that OpenAI had been exploring a path based on video models such as Sora, but its spokesperson now points to the direct use of general-purpose models as the reason for the latest organizational change.
OpenAI has said it is developing a humanoid robot, and Chief Executive Sam Altman has targeted a public demonstration in 2027. That target is a stated aim, not confirmation that a demonstration will take place on that schedule. Fortune also described cost control and a focus on core products as part of the company’s broader business backdrop, but the reason OpenAI gave for this reorganization was its view that its general models can be applied to robotics.
The NVIDIA results cited in the reporting are from a study published October 14, 2026, and described as a CoRL 2026 late-breaking result. They concern a simulated, single-arm robot and should not be read as proof of performance on physical robots. The direct-policy result was 66.3%; the 74.8% figure came from an agentic setup with a persistent Codex session and additional tools. Those setups are not identical to a robot operating continuously in a real environment.
Real-World Performance Still Unproven
The reported benchmark measures simulated tabletop tasks, not a broad range of physical robot work. It remains unclear how GPT-6 Astra or Praxis-1 would perform across different robots, environments and longer tasks, or how their response times would change with different hardware and network conditions. The source material also does not detail enough of Runway’s three-task comparison to assess how repeatable or representative its results are.
OpenAI has not issued a public statement detailing the reorganization, its staffing effects or how it plans to develop its humanoid robot within the new structure. The company’s spokesperson described the move as a sign of continued commitment, but the future responsibilities of each group and any changes to robotics investment have not been publicly specified. It is also not clear whether the benchmark results will lead to a combined system in which a general model plans tasks and a specialist model handles fast movements.
Evidence to Watch Before 2027
Further comparisons on physical robots would help show whether the simulation results carry over to real tasks. Relevant details will include task success, response time, cost, and whether the same methods work across multiple robot designs. Praxis-1’s downloadable weights and fine-tuning code also let outside developers test and adapt the model, but independent results are not included in the source material.
For OpenAI, the next stated milestone to watch is Altman’s target of a public humanoid-robot demonstration in 2027. The company has not confirmed a specific date or described how the reorganized teams will contribute. Until OpenAI provides more detail and independent testing compares the systems in real settings, the evidence supports a narrower conclusion: general models may handle a wider range of simulated tasks, while specialist models have a reported speed advantage in the tests described.
Key Questions
What happened to OpenAI’s robotics team?
According to Fortune, 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 but has not published its own detailed announcement.
What is Runway’s Praxis-1?
Praxis-1 is a 27-billion-parameter robot action model. Runway released its weights and fine-tuning code under the Apache 2.0 license, which allows commercial use subject to the license’s terms.
Which model scored higher on RoboLab-120?
The reported GPT-6 Astra agentic setup scored 74.8%, while Praxis-1 scored 46.2%. These results come from simulation and different model approaches; they do not by themselves establish which system works better on physical robots.
Why might a specialist robot model still be useful?
The reported tests suggest a speed trade-off: Astra’s direct calls averaged 18.56 seconds, while Praxis-1 calls were reported at roughly 0.55 to 0.6 seconds for 1.6 seconds of motion. Fast response can matter for robot control, though the results need broader independent testing.
Has OpenAI set a date for a humanoid-robot demonstration?
OpenAI CEO Sam Altman has targeted a public demonstration in 2027, according to the source material. No specific date or confirmation that it will occur on that schedule was provided.
Source: ThorstenMeyerAI.com
