OpenAI breaks up its robotics team: not an exit, but a bet that general models will run robots
On 13 November OpenAI confirmed to Fortune that it has disbanded its stand-alone robotics team: robotics research joins Mark Chen's main research line, and hardware and applied engineering join Greg Brockman's Applied organisation [1][2]. OpenAI is not leaving robotics; it has changed how it does it.
Reads · OpenAI disbands its stand-alone robotics team, moving its researchers into the main research team and its applied robotics and hardware staff into Applied 2026-11-13
What it signals
A change of route: no separate foundation model for robots
OpenAI's spokesperson said its general-purpose models, GPT-6 Astra among them, have shown they can be applied to robots directly; at the start of the year its researchers expected to build on video models such as Sora [1]. In other words, OpenAI now treats physical intelligence as a downstream capability of general models, not a branch of its own. Guanya Shi of Carnegie Mellon reads it the same way: robotics is "no longer a special problem" [1].
Cutting cost and pooling resources
Fortune notes the move comes as OpenAI cuts costs, has dropped projects such as Sora, and prepares a new funding round and a possible IPO in early 2027 [1]. Cutting a stand-alone team that overlaps the main model pools compute, data and people into it and avoids building twice. OpenAI's own stated reason is the model alone.
A new division of labour: robots as test bed and data source for the general model
Robots will serve more as places to test the general model and to collect data for it; hardware will lean on outside partners, and building its own drops down OpenAI's list.
OpenAI says it is still building its own humanoid, with Sam Altman targeting a first public demonstration in 2027 [1]. On hardware, the company's own word does not match this for now.
What it means for the industry
General models may gradually displace robot-specific models
There are two routes today: general models made to control robots directly, as at OpenAI, and the specialist VLA (vision-language-action) or robot foundation models built by Figure, 1X and most Chinese humanoid makers. General models are likely to reach both the S2 layer (slow thinking: understanding and planning) and the S1 layer (fast reaction: real-time motor control), and so replace specialist embodied models outright. Figure ending its OpenAI partnership in February 2025 to build its own robot AI [1] is the other route.
The only evidence so far is a simulation: in an Nvidia study an Astra-powered agent scored 74.8% on a simulated manipulation benchmark, against 51.0% for the best specialised models, not on a real robot [1]. The people Fortune asked do not expect a sweep: MIT's Russ Tedrake, CEO of Walden Robotics, says companies whose robots already work on factory floors have their site data as an edge; Guanya Shi says large models are slow and need a connection, so robots still need small models that run on board; Generalist's Pete Florence says it is unclear any model will dominate robotics [1].
Data collection changes its target
OpenAI's data collection and real-robot training will continue, but the aim shifts from training a separate robot model to filling a general model's long tail of physical-interaction data, so multimodal datasets matter more. No source says OpenAI's data collection continues; Tedrake says better foundation models make site data more valuable [1], which fits "data matters more" but is about robot companies' own data.
What would change my mind
- OpenAI's 2027 humanoid demonstration: whether a general model runs the robot, and whether OpenAI built the hardware.
- Success rates for general models on real robots, not only in simulation, in factories or warehouses, against specialist models on the same tasks.
- Whether specialist-model companies such as Figure and 1X see their next rounds or customer deployments slow.
Sources
- Wen Shao, "Exclusive: OpenAI disbands its stand-alone robotics team as its general models start running robots", Fortune, 13 November 2026 fortune.com ↗
- Timeline: openai-disbands-standalone-robotics-team

