Reported 13 Nov 2026 · Labs A one-minute read

OpenAI dissolved its robotics team, saying its general models can now run robots.

Fortune reported on November 13 that OpenAI folded its stand-alone robotics group into its core research and applied teams. Caveat: OpenAI published nothing itself, and the benchmark behind the claim is simulated, not physical.

Until now

Robots need specialist models trained on robot data.

Now

A frontier laboratory says the general model is the robot.

Why it matters

This is the second time OpenAI has dissolved a robotics team. In 2020 the reason was too little robot data. Now the reason is the opposite: one general model scored 74.8% where the best specialist managed 51%. The economics disagree. Each attempt cost about $36.40 and took 18.56 seconds.

Established

Fortune reported the reorganization; an OpenAI spokesperson confirmed it in that article.
Researchers moved under Mark Chen; hardware and applied work under Greg Brockman.
OpenAI said robotics remains a core part of its long-term vision.

Not established

How many people moved. No headcount was disclosed to anyone.
That general models can run real robots. The benchmark was simulated.
An exit from robotics. A humanoid demonstration is still targeted for 2027.

The five-minute readThe record, the timeline, the weight and 5 sources

The record

Reported, not verified. The Fortune piece is a paywalled exclusive, OpenAI published no post of its own, and the confirmation is a spokesperson's quote inside that article. We read 4 independent summaries of it and found nothing in OpenAI's own newsroom. The dated event that would move it: an OpenAI post or org page naming the change, by December 15, 2026.

Fortune published an exclusive by Wen Shao on November 13, 2026, headlined "OpenAI disbands its stand-alone robotics team as its general-purpose AI models start running robots." Per that report and 4 independent summaries of it: OpenAI's robotics researchers move into the main research organization under Chief Research Officer Mark Chen, and the applied robotics and hardware teams move into the Applied organization under President Greg Brockman. No headcount was given. An OpenAI spokesperson told Fortune that "robotics remains a core part of OpenAI's long-term vision," and that the change is meant to speed progress on general-purpose robots rather than end the work. The reasoning reported is that OpenAI's general frontier models, GPT-6 Astra among them, now apply directly to robot control, where the company had previously expected progress to come from video models such as Sora. The benchmark cited is an Nvidia study of a simulated pick-and-place task: an Astra-powered agent succeeded on 74.8% of tasks against 51.0% for the best dedicated robot model on the same leaderboard. The same account puts the cost at about $36.40 in tokens per attempt and the latency at 18.56 seconds per model call, which it calls too slow for most factory floors. History: OpenAI wound down an earlier robotics team in 2020, with cofounder Wojciech Zaremba citing a shortage of robot data; it rebuilt a team in 2024 alongside a partnership with the humanoid maker Figure; that partnership ended in February 2025 when Figure said it would build its own robot models. OpenAI is developing its own humanoid, and Sam Altman has targeted a first public demonstration in 2027. Fortune reports OpenAI had registered booth 208 at CoRL, a robotics conference in Austin, and that the booth was empty and unbranded when visited. A Chinese hourly digest on November 15 carried the same reorganization and called it OpenAI's second robotics disbanding. Not established: how many people were affected, whether anyone left the company, the date the decision was taken, the title of the Nvidia study, and any result on physical hardware.

How it got here

2020
OpenAI winds down its first robotics team. Cofounder Wojciech Zaremba cites a shortage of robot data as the reason.
2024 TO Feb 2025
OpenAI rebuilds a robotics team and partners with the humanoid maker Figure. Figure ends the partnership in February 2025, saying it will build its own robot models.
Nov 10
Google DeepMind's robot model sorts parts on live Mercedes-Benz production lines, succeeding 96% of the time. Edition 086's card. A specialist model, on real hardware.
Nov 12
SoftBank and Nvidia put $500 million into Agile Robots, a German humanoid maker. Edition 089's card. Capital is still backing robot specialists.
Nov 13
Fortune reports the disbanding. OpenAI confirms the reorganization through a spokesperson and publishes nothing of its own. Its CoRL booth in Austin is found empty and unbranded.

Who has to react

Figure, Agile Robots and every company whose pitch is a specialist robot model; Nvidia, whose study supplied the comparison; OpenAI's own humanoid programme, now reporting into 2 groups; CoRL's organisers; and the researchers deciding whether robotics is still a field.

What to watch

Would strengthen itA general frontier model from any laboratory beats specialist robot models on a physical-hardware benchmark, not simulation, by September 30, 2027.
Would weaken itOpenAI rebuilds a named robotics team or hires a robotics lead, or its humanoid demonstration slips past December 31, 2027.
Would weaken itA specialist robot model beats GPT-6 Astra by more than 10 points on the same leaderboard, or the cost per attempt stays above $10, by June 30, 2027.

Why it weighs this much

Signal strength

CostLow

CommitmentMedium

ReversibilityHigh

PersistenceMonths

Cost is low. Moving people between internal groups costs nothing to announce, and OpenAI did not even announce it. Commitment is medium: a spokesperson confirmed it on the record, which is harder to walk back than a leak, but there is no post, no filing and no headcount. Reversibility is high, and the company has proved it, dissolving a robotics team in 2020 and rebuilding one in 2024. Persistence is months rather than years. A thesis that general models subsume specialists gets tested at the next benchmark round, and $36.40 and 18.56 seconds per attempt say it is not yet true on a factory floor.

The words

Pick-and-place benchmark: A test in which a robot must grasp an object and put it somewhere. Easy for people, hard for machines, and the standard way to score robot control.
Specialist robot model: A model trained mostly on robot sensor and motion data, rather than a general model trained on text, images and video and then pointed at a robot.
Latency: The delay between asking a model and getting an answer back. At 18.56 seconds a call, a production line stands still waiting.
Applied organization: OpenAI's product and deployment side, led by Greg Brockman, as distinct from the research organization that trains the models.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.