OpenAI Just Killed Its Robotics Team — Here's What Sam Altman Is Really Betting On

OpenAI has disbanded its standalone robotics team — for the second time in six years — and its explanation is the boldest bet in AI right now: general-purpose models like GPT-6 Astra are now good enough to control robots directly, so a dedicated robotics team is no longer needed. The researchers have been folded into OpenAI's main research organization under Chief Research Officer Mark Chen, while applied robotics and hardware staff moved into the Applied organization under President Greg Brockman. There is no robotics team anymore. The question is whether the model really is the robot now — or whether OpenAI just fired the people who would know the difference.

How the news broke: an empty booth in Austin

The reorganization was first reported by Fortune on November 13, 2026, in an exclusive by Wen Shao, and OpenAI confirmed it when asked. The timing gave the story away before the company did: OpenAI had registered a booth at the Conference on Robot Learning (CoRL) in Austin that week, but when Fortune visited, booth #208 stood empty — the company's name still printed on the floor plan, with no signage or product on site. The team that was supposed to be there no longer existed.

An OpenAI spokesperson framed the move as acceleration, not retreat: robotics remains core to the company's long-term vision, and its target of a first public humanoid demonstration in 2027 still stands. But the company has published no blog post of its own about the change, has not said how many people were affected, and has not disclosed the post-restructuring budget, headcount, or compute allocation for robotics. Everything the public knows comes from Fortune's reporting and the coverage around it. (Fortune's reporting, via The Labo Recon)

The bet: GPT-6 Astra is the robot now

OpenAI's stated reason is simple, and it is the kind of claim that sounds obvious until you sit with it for a minute. Researchers initially expected that advances in models such as Sora, the video-generation system, would supply the physical-world understanding robots need. The company now believes its general-purpose frontier models, including GPT-6 Astra, can be applied to robotic tasks directly — no robot-specific brain required. Robotics becomes an application layer, not a research frontier.

The evidence OpenAI points to comes from a study NVIDIA ran on October 14, 2026. Researchers gave GPT-6 Astra camera images, robot state, and a control interface on RoboLab-120, a simulated pick-and-place benchmark, with no special robot training at all. The Astra-powered agent scored 74.8% success, against 51.0% for the best specialized robotic model on the same leaderboard. A general model, zero robot training, beating a purpose-built robot model by more than 20 points. (VectorWire analysis of the NVIDIA study)

The caveats — and they are big ones

Before you conclude that the robotics problem is solved, read the fine print:

  • It was a simulation, not a physical robot. RoboLab-120 is a simulated benchmark. Simulated physics is kinder than the real world, where friction, wear, and sensor noise live.
  • It was slow and expensive. The Astra agent consumed about 30.9 million tokens per successful task episode — roughly $36.40 per attempt — and each model call took about 18.56 seconds. A robot that pauses 18 seconds before every move and costs $36 per grasp is not a factory worker.
  • Other tests tell a different story. In six real-machine and simulation tests built around GPT-6 Astra, the model failed to directly grasp a Rubik's cube, simulate carrying a basket, or directly stack bowls. On Anthropic's RoboDojo Sim framework, Astra scored 29.8%, barely ahead of Claude Fable 5.2's previous generation at 27.6% — and far from the 94–96% success rates humanoid robots are already achieving on factory floors. (Dev.to breakdown of the evidence)

The honest reading: a general-purpose model can sometimes beat a specialized one on a simulated benchmark, at a price and speed no real deployment can afford. That is a research result, not a product.

This is not the first time OpenAI has done this

OpenAI's robotics history is a loop. The company built a robotics team in its early years, then discontinued the effort in 2020. It revived robotics work in 2024, announcing a collaboration with humanoid robotics company Figure to develop AI models for robots. That partnership ended in February 2025. Now, in November 2026, the standalone team is gone again — folded into main research and Applied.

What makes this round stranger is the whiplash: OpenAI was recruiting robotics talent at salaries of up to $500,000 as recently as two months before the dissolution. Either the Astra results changed minds very fast, or something else — cost pressure, reorganization politics — is doing the talking. Fortune also reported operating losses of $10.9 billion in the most recent quarter, as the company seeks to close a new funding round and prepares for a possible IPO in early 2027. OpenAI closed an $11 billion funding round on November 6. A company burning that much cash has reasons to consolidate teams that go beyond benchmark scores.

What this means for the AI agents already among us

Zoom out, and the robotics move is part of a bigger pattern: AI is graduating from answering questions to doing things in the real world. Last week, United Airlines banned AI agents from booking flights — the first major airline to draw a line against agents acting on your behalf. Google's Gemini 4 Hydrogen is priced to undercut everything for coding agents. And 1,000 Claude agents built a working operating system in 8 days with no humans writing the code. Agents are no longer a demo category; they are the product strategy of every major lab — and now OpenAI is betting the same logic extends to physical machines.

If OpenAI is right, the robotics industry's center of gravity shifts from hardware labs to model labs, and the company that owns the best general model owns the robot too. If it is wrong, it has just dissolved the team that would have caught the mistake — and its rivals, from Figure to Google's Gemini Robotics program, will be happy to hire the people OpenAI let go.

The bottom line

OpenAI did not quit robotics. It bet that GPT-6 Astra makes a robotics team redundant — a claim supported by one impressive simulated benchmark and undermined by the cost, speed, and real-world failures in the same body of evidence. The 2027 humanoid demo is the deadline that will settle it. Watch whether Astra gets faster and cheaper, whether OpenAI quietly rebuilds a robotics team under a different name, and whether the people who left end up building the robots OpenAI says it no longer needs a team for.

By Shakeel Ahmed — AI Growth Lab

Comments

Popular posts from this blog

92 Million Jobs Gone by 2030: What AI Is Really Doing to Your Career

13 AI Agents Were Left Alone on a Virtual Island for 30 Hours. They Invented Their Own Goals.

Sweden Just Banned ChatGPT and Claude at Work — Starting Today