🔍 Read the full analysis: OpenAI Splits Its Robotics Team; Runway Opens Praxis-1. What's Next For Robot AI? 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 has disbanded its stand-alone robotics team, moving researchers into its main research organization and applied robotics and hardware staff into its Applied organization. Runway released Praxis-1, an open-weight robot model; available tests suggest general-purpose models can perform well in simulation, while specialist models act faster. The comparison does not settle how the two approaches will be used in deployed robots.
OpenAI has broken up its stand-alone robotics team, moving its researchers into the company’s main research organization and its applied robotics and hardware teams into Applied, according to Fortune, which said OpenAI confirmed the change. The shift came as Runway released Praxis-1, an open-weight model built for robot control, putting a fresh focus on whether robots need specialist AI or can use general-purpose models.
Fortune reported the reorganization on November 13, after finding OpenAI’s registered booth empty at the Conference on Robot Learning in Austin. The company’s robotics researchers now report to Chief Research Officer Mark Chen, while the applied robotics and hardware teams moved to the Applied organization under President Greg Brockman. OpenAI did not publish a separate announcement about the change, according to the source material.
An OpenAI spokesperson told Fortune the move reflects the company’s belief 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. That explanation is the company’s stated rationale; the reporting also noted broader pressure on OpenAI to control costs, but did not establish that cost-cutting caused this 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. Runway’s announcement frames Praxis-1 as a dedicated robot model, contrasting with OpenAI’s stated direction toward using general-purpose systems.
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
The Trade-Off Between Skill and Speed
For robotics developers, the key issue is not simply which model scores highest, but whether it can make decisions quickly and reliably enough to control a physical machine. In an NVIDIA simulation study, GPT-6 Astra reached a reported 74.8% success rate on RoboLab-120 when used in an agentic setup. The best trained robot policy on that benchmark’s leaderboard scored 51.0%, according to the source material. The result suggests a general model can handle a broad set of simulated tasks, but it does not establish performance in real-world deployment.
Speed was a major limitation in the study’s direct setup: GPT-6 Astra averaged 18.56 seconds per model call, and the simulation paused while it generated a response. Runway reported Praxis-1 achieved 46.2% on RoboLab-120, placing it fifth overall and first among the world-action and open-source robot models listed. It took about 0.55 to 0.6 seconds per call for 1.6 seconds of motion. Those figures point to a practical divide: general models may be capable across tasks, while specialist policies may be better suited to rapid control.
Runway also compared the systems on one of its robots, reporting that GPT-6 Astra took 15.6 times longer than Praxis-1 to complete three tasks and cost 51 times more per completed task. That is a company-run comparison, not an independent evaluation, and the small task set cannot establish how the models perform across other robots or conditions. Still, the results help explain why the debate matters for the cost, responsiveness and design of future robot systems.
Popular right now
Ninja AF101 4-Quart Air Fryer, 4-in-1, Grey
Vitamix Touch & Go Advance 32-Ounce Bar Blender, White
Vitamix VX1 Professional-Grade Blender, 64 oz, Black
As an affiliate, we earn on qualifying purchases.
OpenAI’s Earlier Robotics Efforts
This is the second time OpenAI has closed a stand-alone robotics team. The company wound down an earlier effort in 2020. Cofounder Wojciech Zaremba later cited 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-robot maker Figure; Figure said in February 2025 that it would develop its own robot AI, ending that partnership.
The current reorganization follows a different explanation from the one cited after the 2020 closure. Fortune reported that OpenAI’s researchers had expected earlier in 2026 to build on video models such as Sora, but a spokesperson said progress in general-purpose models changed that view. OpenAI has also said it is developing a humanoid robot, with CEO Sam Altman targeting a public demonstration in 2027. That target is an announced plan, not confirmation that a demonstration will occur on schedule.
The evidence often used to support general-purpose robot control remains bounded. NVIDIA researchers tested GPT-6 Astra in RoboLab-120, a simulation of 120 tabletop tasks for one robot arm. The study compared different system setups, including a direct policy, an agentic policy with persistent history and tools, and a setup in which Astra reviewed another model’s proposed movements. Simulated task success is useful for comparison, but it does not by itself show how a system handles changing environments, hardware failures or safety demands in physical settings.
“The two approaches are complementary.”
— Runway
What the Tests Do Not Establish
The reported benchmark results do not settle whether general-purpose models can control robots reliably outside simulation. RoboLab-120 covers tabletop tasks with a single simulated arm, while physical robots face differences in sensors, hardware, surroundings and the consequences of mistakes. The available material does not provide a matched, independent real-world evaluation of GPT-6 Astra and Praxis-1.
The comparisons also have limits. NVIDIA’s findings come from its researchers’ study, and Runway’s speed and cost figures come from a test on one of its own robots involving three tasks. The source material does not give enough detail to determine whether every measurement used identical operating conditions, or how results would change across other tasks and hardware. Praxis-1’s open weights make it available for testing and adaptation, but do not alone establish safety or performance in commercial deployments.
OpenAI has not publicly detailed the new teams’ responsibilities, staffing levels or how its robotics work will proceed after the reorganization. The company’s stated commitment and reported 2027 demonstration target do not clarify what product, robot or technical approach will be shown. It is also unclear whether OpenAI plans to combine general models with specialist control systems, as Runway says may be possible.
Evidence to Watch in 2027
The next useful milestones are independent tests on physical robots, including comparisons that use the same hardware, tasks and measurement methods for general-purpose and specialist models. Such evaluations could show whether the simulated success rates translate into reliable control, and whether the speed and cost gaps reported so far persist beyond the setups described.
For OpenAI, readers can watch for a direct company statement, more detail on the responsibilities of its reorganized teams, and developments toward Altman’s stated 2027 public demonstration target. For Runway, Praxis-1’s use by outside researchers and developers may produce additional results, though comparisons will depend on the robots and tasks chosen. At present, the confirmed development is a shift in organizational structure alongside the release of a specialist model; which approach will prove more useful in deployed robotics remains open.
Key Questions
What changed in OpenAI’s robotics organization?
According to Fortune, OpenAI disbanded its stand-alone robotics team. Robotics researchers moved into the main research organization under Mark Chen, while applied robotics and hardware teams moved to Applied under Greg Brockman.
What is Runway’s Praxis-1?
Runway describes Praxis-1 as a 27-billion-parameter world action model for robotics. It released the model’s weights and fine-tuning code under the Apache 2.0 license.
Did GPT-6 Astra outperform Praxis-1?
In the reported RoboLab-120 results, GPT-6 Astra’s agentic setup scored 74.8%, while Praxis-1 scored 46.2%. The figures come from different reported testing work and do not establish which model performs better under matched real-world conditions.
Why might a specialist robot model still be useful?
The reported tests show a speed difference: GPT-6 Astra averaged 18.56 seconds per call in NVIDIA’s direct simulation setup, while Runway reported about 0.55 to 0.6 seconds per Praxis-1 call. Faster responses can matter for robot control, although the figures come from different setups.
What is still unknown about OpenAI’s plans?
OpenAI has not publicly detailed how its reorganized teams will work, what robot system they will develop or whether it will combine general-purpose and specialist models. A public demonstration in 2027 is a target attributed to Sam Altman, not a confirmed event.
Source: ThorstenMeyerAI.com
