Skip to content
HomeBriefFor youTrackersTopicsCompaniesGlossaryNewsletterSecurityHealth & medicineFinance & marketsRoboticsModels & researchPolicy & regulationEnterprise & businessDeveloper tools
Robotics · Nov 14, 2026 · 3 min read

OpenAI disbands its standalone robotics team for the second time, betting GPT-6 Astra's 74.8% simulation score makes a separate robot brain unnecessary

The researchers move under Mark Chen, the hardware staff under Greg Brockman, and the empty booth at CoRL tells robot makers what OpenAI thinks the moat is: the general model, not the robot.

Why it matters
ExecutivesIf you are evaluating a humanoid or manipulation vendor, ask what happens to its differentiation if a general frontier model matches its policy next year. The honest answers involve deployment data and hardware, not model architecture.
ML engineers74.8% versus 51.0% on a simulated manipulation leaderboard is a strong signal and a narrow one. Latency, onboard compute and sim-to-real transfer are where specialised stacks still win; the interesting designs split the task between a general planner and a small controller.
FinanceOpenAI wound down robotics in 2020 for lack of data and is reorganising it in 2026 after a $10.9 billion quarterly loss and Sora's closure. Treat 'robotics is core' and 'robotics has no dedicated team' as both true and price the ambiguity.

OpenAI has dissolved its standalone robotics team, Fortune reported on Thursday, and the company confirmed it. The robotics researchers now report into the main research organisation under chief research officer Mark Chen; the applied robotics and hardware engineers move into the Applied organisation under president Greg Brockman. OpenAI's statement is that robotics remains a core part of its long-term vision and the restructuring will speed it up. The visible evidence was an empty booth, number 208, at the CoRL robotics conference in Austin this week, registered under OpenAI's name and left without branding or a demo.

What changed in the thesis

Early in 2026, according to Fortune's reporting, OpenAI's robotics researchers expected progress to come from building on video models like Sora, the generation system OpenAI has since discontinued. The company now says its general-purpose frontier models, GPT-6 Astra in particular, can be applied to robots directly. The evidence it points to is an Nvidia study in which an Astra-powered agent reached a 74.8% success rate on a simulated manipulation benchmark where the best specialised robotics models scored 51.0%. If a general model beats purpose-built policies in simulation, the argument runs, a separate robotics research group is organisational overhead.

This is the same bet OpenAI made in coding and in agents: the frontier model plus a harness beats a specialised model, and the harness is a product question rather than a research one, which is why the hardware people went to Applied.

The history

OpenAI has been here before. Its first robotics team, famous for a robot hand that solved a Rubik's cube, was wound down in 2020; co-founder Wojciech Zaremba said the data was not there and other routes to general intelligence were faster. The company rebuilt a robotics team in 2024 and partnered with Figure on humanoid models until Figure walked away in February 2025 to build its own. OpenAI says it is developing a humanoid of its own and Sam Altman has targeted a first public demonstration in 2027. The timing of this reorganisation, after a reported $10.9 billion quarterly loss, Sora's closure and ahead of a possible 2027 listing, suggests cost discipline is at least half the story.

The counter-argument

Robotics researchers quoted by Fortune were not persuaded that the general model is the whole answer. Carnegie Mellon's Guanya Shi read the move as OpenAI deciding robotics is 'no longer a special problem', but argued that small onboard models are still required for control. MIT's Russ Tedrake pointed to real-world deployment data as the asset robotics companies hold and labs do not. Generalist's Pete Florence said it is unclear which model class wins in the long run. A simulation leaderboard measures planning and perception under benign conditions; it does not measure a 50-millisecond control loop on a battery-powered arm, or what happens when the gripper slips. This week's other robotics stories, J.P. Morgan's note that hands and actuators are the binding constraint and Skild's and Nvidia's simulation releases, all point to bottlenecks a frontier model does not touch.

What to watch

Whether OpenAI publishes Astra robot results outside Nvidia's benchmark, and on physical hardware. Whether Figure, Agility, Apptronik and the Chinese humanoid makers respond by licensing frontier planners while keeping their own controllers, which would validate OpenAI's thesis and theirs at the same time. Whether the 2027 humanoid demo date survives the reorganisation. And whether Nvidia, whose GR00T 2 world action model competes directly with the general-model approach, keeps publishing leaderboards where OpenAI's model wins.

Key terms in this story

  • World action model A robot foundation model that predicts how the world will change and chooses actions from that prediction, rather than mapping images straight to motor commands.
  • Frontier model A model at or near the state of the art at the time, typically the newest top tier from a major lab.
  • Benchmark A standard test used to compare models, such as a set of coding tasks or exam questions.
  • Agent An AI system that pursues a goal by taking actions with tools, in a loop, rather than answering a single question.
All 99 terms →
roboticsmodelsagentsOpenAI
Written in our own words from the linked sources · updated 2026-11-14 · corrections: corrections@aibriefing.io

Get the brief in your inbox

One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.

Published daily by AI Briefing. We never share your address.
Like this? Get tomorrow's brief by email, free.