OpenAI has disbanded its standalone robotics team and reassigned its staff, the company confirmed to Fortune on November 13, 2026. Robotics researchers are moving into the main research organization under Chief Research Officer Mark Chen. Applied robotics and hardware teams are joining the Applied organization led by President Greg Brockman.
OpenAI says the change is intended to accelerate robotics work. Its reasoning is that general-purpose frontier models, including GPT-6 Astra, have become capable enough at robotics tasks to justify bringing that research closer to its core model development.
That is a consequential bet about where robot intelligence should come from. It challenges the idea that robotics needs a separate model-development effort, but leaves open the practical question of how those models perform when machines must act reliably in the physical world.
What Changes Inside OpenAI
The reorganization separates foundational research from applied robotics and hardware work:
| Work | New organizational home | Leadership |
|---|---|---|
| Robotics research | Main research organization | Chief Research Officer Mark Chen |
| Applied robotics and hardware | Applied organization | President Greg Brockman |
OpenAI told Fortune that robotics remains part of its long-term vision. The reporting did not disclose headcount, layoffs, a new product timetable, or a change to the previously stated target for a public robot demonstration in 2027. The confirmation came through company statements to Fortune, without a first-party announcement detailing the reorganization.
Fortune also reported a visible sign of the change at the Conference on Robot Learning, or CoRL, in Austin. OpenAI had registered a booth, but when the publication visited booth 208, the space was empty and its logo was absent. That documents an unused conference presence, not the technical state of its robotics program.
OpenAI has changed direction in robotics before. According to Fortune’s account, it wound down an earlier effort in 2020, with cofounder Wojciech Zaremba later citing limited robot data and a more efficient route toward artificial general intelligence through other methods. It rebuilt a standalone team in 2024 and partnered with humanoid robotics startup Figure. That partnership ended in February 2025, when Figure said it would develop its own robot AI.
The latest move keeps robotics inside OpenAI while changing where the company expects the underlying intelligence to be developed.
Why OpenAI Is Betting on General-Purpose Models
OpenAI told Fortune that, at the start of 2026, its researchers expected robotics progress to build on work such as Sora. The reasoning was that generating realistic video requires some implicit understanding of the physical world.
The company now says its general-purpose frontier models, including GPT-6 Astra, have shown they can be applied directly to robotics. OpenAI identified that progress as the motivation for moving robotics researchers into its broader research organization.
Guanya Shi, an assistant professor at Carnegie Mellon University’s Robotics Institute, told Fortune he would interpret the move as a signal that OpenAI sees robotics as “no longer a special problem.” In that interpretation, robotics becomes another application of general intelligence, much as one underlying model can support coding, legal analysis, and financial work.
The attraction is clear: improvements in vision, reasoning, and instruction-following could benefit robots without requiring a separate team to recreate those capabilities.
But general-purpose describes a model’s scope, not its training history. The Decoder’s November 14 coverage notes that the amount of robotics data used to train these models remains unclear. Their performance cannot be treated as proof that robot-specific data is unnecessary.
My read is that OpenAI is moving the organizational boundary between foundational intelligence and robotics application work. Folding researchers into a central team may help them improve a shared model. It does not, by itself, eliminate the work needed to turn that model into a dependable robot system.
The Benchmarks Show Promise, With Important Conditions
Two reported experiments help explain the excitement around general-purpose robot controllers. They also show why a strong comparative result should not be confused with deployment readiness.
The Decoder reports the following results:
| Evaluation | GPT-6 Astra result | Specialist comparison | Important condition |
|---|---|---|---|
| Nvidia robotics simulation study | 74.8% task success for an Astra-powered agent | 51% for the best specialized models in the reported comparison | Used a simulated DROID robot, not physical hardware |
| StationeryBench | 7 fully completed tasks out of 100 trials | Ai2’s MolmoAct2 completed 0 out of 100 | Five desk-object tasks using dual-arm YAM robotic arms |
| StationeryBench progress score | Average score of 46 | Average score of 12 for MolmoAct2 | Progress score is a separate measure from full completion |
The Nvidia result is a vendor-reported benchmark, relayed through news coverage. The difference is 23.8 percentage points in that simulation setup. It supports the case that a general-purpose model can outperform specialized alternatives under some conditions. A simulated robot result cannot establish physical-world reliability.
StationeryBench offers a different signal because it involved physical robotic arms. Across 200 trials split between the two models, Astra fully completed seven tasks and MolmoAct2 completed none. Astra also scored higher on partial progress.
Seven completions out of 100 attempts is a 7% full-completion rate. The other 93 attempts did not satisfy the full-completion criterion, although some made progress. That distinction matters for buyers: moving objects partway toward a goal can be a useful research result without delivering a finished job.
I would give that absolute completion rate more weight than the relative lead when judging readiness for unattended work. Beating another model is evidence of comparative progress; consistently finishing tasks is the practical requirement.
The news coverage does not supply enough methodology to assess every difference in task setup, model integration, or scoring. These results should therefore be read separately. The 74.8% simulation result and 7% physical-task completion rate are not measurements of the same workload, so subtracting one from the other would not reveal a meaningful simulation-to-reality gap.
Robot Intelligence Still Has to Meet Timing Constraints
A robot controller can make several kinds of decisions. One system might interpret a request, recognize objects, and choose the next step. Another might handle movement, contact, and rapid responses to changing conditions.
OpenAI’s organizational announcement does not explain which of those responsibilities GPT-6 Astra would handle in a deployed robot. That missing detail matters because controlling a robot is not one uniform computing problem.
Several practical constraints shape the architecture:
- Latency: A model that takes time to reason may still be useful for planning. A movement system needs responses that arrive within the timing requirements of the task.
- Connectivity: If decisions depend on a remote service, network delays and outages become part of the robot’s operating conditions. The system needs a defined response when communication fails.
- Onboard computing: Running models locally reduces dependence on a connection, but introduces limits involving memory, power, heat, and hardware cost.
- Cost and consistency: A broadly capable model may be useful across many tasks, while a specialized model may offer a better operating tradeoff for one repeated job.
The Decoder identifies speed, energy consumption, cost, and reliability as areas where specialized robotics models may retain advantages.
Frequently Asked Questions
3 questions
1Is OpenAI Ending Its Robotics Work?
No. OpenAI told Fortune that robotics remains part of its long-term vision. It is moving robotics staff into its main research and Applied organizations.
