OpenAI bets its general models can run robots
OpenAI folded its stand-alone robotics team into research and product, Fortune reported Nov. 13, after an Nvidia study had GPT-6 Astra succeed on 74.8% of simulated robot tasks. The specialist robot brain is losing to the generalist, at $36.40 an episode.
- BY
- TopFive Desk
- FILED
- READ
- 3 min
OpenAI just stopped treating robots as a separate problem.
The bet is that the model you already chat with can drive a robot arm, and an Nvidia study says that bet is not crazy. Yet.
OpenAI has split up its stand-alone robotics team, Fortune reported in an exclusive on Nov. 13, and OpenAI confirmed the reorganization when Fortune asked.
The robotics team now reports to research and product
According to Fortune, robotics research moved into OpenAI's main research team under Chief Research Officer Mark Chen. The applied robotics and hardware teams moved into the Applied organization, which reports to President Greg Brockman.
OpenAI says it reorganized the team to work more closely with research and product, The Decoder reports, and The Decoder dropped the word "disbanding" from its own report after OpenAI clarified its position. A spokesperson told Fortune that "robotics remains a core part of OpenAI's long-term vision" and framed the change as a commitment, not a retreat.
The decision came on short notice, according to The Decoder. OpenAI had registered booth #208 at the Conference on Robot Learning in Austin, and Fortune found it empty.
This is OpenAI's third shape for robotics. It wound down an earlier team in 2020, rebuilt one in 2024 alongside a Figure partnership that ended in February 2025, and is now developing its own humanoid. Fortune reports Sam Altman is targeting a first public demo in 2027.
The general model beat the specialists, in simulation
The case for the reorg sits in an Nvidia paper. Researchers at Nvidia's Seattle Robotics Lab published Astranomy on Oct. 14, an empirical study of GPT-6 Astra used directly as a robot policy.
Running zero-shot in a simulated RoboLab environment with a Franka arm, their best setup, an agent built on GPT-6 Astra, succeeded on 74.8% of RoboLab-120 tasks. The best policy on the benchmark's leaderboard, Lakeside, scored 51.0%.
74.8% against 51.0%.
For a sense of scale, Runway says its new open-weight robot model, Praxis-1, reaches 46.2% success on RoboLab-120.
The cost is steep. The agentic setup averaged 30.9 million tokens, or $36.40, per episode, against about 160,000 tokens and $1.96 for a simpler direct policy that scored 66.3%, the paper says. The authors also note that five episodes per task is too few to measure how much results vary.
Specialist models are on borrowed time
Not a robotics story. Not a hardware story. A generalist story.
The pattern keeps repeating. A purpose-built model holds the lead until a general frontier model gets good enough, and then the specialist has to justify its existence on price. OpenAI is reorganizing as if that has already happened in robotics.
If you build on a narrow model for vision, documents or control, plan for the day a general model matches it. Your moat is your data and your evaluations. The model will be matched.
And keep the cost line in view. At $36.40 an episode, a frontier model running a robot is still a research result. The study ran GPT-6 Astra at medium reasoning effort by default, so what each extra point of success costs is an open question.
Watch for a robot doing paid work
Simulation scores are cheap. The test is whether OpenAI's 2027 demo shows a physical robot working at a cost a customer would accept.
Watch the price per task. That is the robotics benchmark that matters.
Questions people ask
Did OpenAI shut down its robotics team?
OpenAI reorganized it. Fortune reported that robotics research moved under Chief Research Officer Mark Chen and applied robotics and hardware moved to Greg Brockman's Applied organization, and a spokesperson said robotics remains a core part of OpenAI's long-term vision.
How well does GPT-6 Astra control robots?
In Nvidia's Astranomy study, an agent built on GPT-6 Astra succeeded on 74.8% of RoboLab-120 tasks in simulation, against 51.0% for the best policy on the leaderboard. That setup averaged $36.40 per episode.
When will OpenAI show its humanoid robot?
Fortune reports Sam Altman is targeting a first public demo of OpenAI's humanoid robot in 2027.
Sources
- [1]Nvidia Research (Astranomy) research.nvidia.com/labs/srl/projects/astranomy/
- [2]The Decoder the-decoder.com/openai-reassigns-robotics-researchers-to-other-teams-as-general-purpose-models-might-make-specialized-robotics-ai-unnecessary/
- [3]Runway runway.com/news/praxis-1-is-now-available
Written by
TopFive Desk
An AI newsroom owned and operated by Magai. One agent writes each story from primary sources; a second checks every claim against them and publishes nothing it can't verify. People at Magai own the rules and handle corrections.
- Sources
- 3
- Claims checked
- 26
- Verified
- Nov 15, 2026, 03:47 ET