AI
Teknosfera Newsroom
AI

OpenAI's robotics bet: scale wins, but who explains the failures?

Folding robotics into general models may be efficient, but it also dissolves the specialists who would know why a robot dropped the bottle, and that matters more than the benchmark.

Published
November 14, 2026
Reading time
3 min
Categories
AI

Photo: Cwoo4792 / Wikimedia Commons (CC BY-SA 3.0)

What do you call a robotics program with no robotics team? Either the most confident strategy in AI, or a very elegant way to cut costs. Probably both.

According to an exclusive from Fortune, OpenAI has disbanded its stand-alone robotics team. The researchers now report to Chief Research Officer Mark Chen, while the applied robotics and hardware staff moved under President Greg Brockman. The company says robotics remains a core part of its long-term vision, and Sam Altman is reportedly targeting a first public demo of its own humanoid in 2027.

The stated reason is that general frontier models, GPT-6 Astra among them, can now be applied directly to robots. Fortune cites an Nvidia study in which an Astra-powered agent hit 74.8% on a simulated manipulation benchmark, against 51.0% for the best specialized robotic models on that leaderboard.

The scale bet is more credible than it sounds

It's easy to scoff, but we think the logic is coherent. Carnegie Mellon's Guanya Shi told Fortune he reads the move as a sign that OpenAI sees robotics as "no longer a special problem," just another downstream application of general intelligence, like coding or legal tools. If the same model that writes code can also handle a bottle in an unexpected place or odd lighting, a separate robot-brain team starts to look like duplicated effort.

That is also the story of the last few years in AI: specialized systems keep losing to bigger general ones. A gap of nearly 24 points on a benchmark is hard to wave away. Our caveat is that it's a simulated benchmark, not a factory floor, and benchmarks are exactly where general models tend to look their best.

The money story is hiding in plain sight

Don't read this as pure science. Fortune notes OpenAI is under pressure to rein in costs, has dropped "side quests" (Sora included), has reportedly lost $10.9 billion in operating losses in its latest quarter, and is eyeing a possible IPO in early 2027. A stand-alone robotics team looks to us like a textbook side quest. "General models do it better" is a wonderful finding to have exactly when you need to simplify the org chart.

We're not claiming the finding is fake. But when a technical conclusion and a budget need point the same way, readers should hold both in mind. It's also the second time OpenAI has closed a robotics team, after the 2020 wind-down, which Fortune says cofounder Wojciech Zaremba tied to limited robot data.

Who will know why the robot failed?

Here is our real objection. A language model that gets something wrong produces a bad paragraph. A robot that gets something wrong can break a plate, or something worse. Debugging that probably takes people who live with actuators, latency, calibration and contact physics every day. Spread those people across general research and applied teams and you may not lose their skills overnight, but you lose the place where failure analysis was the whole job.

The counterweight comes from Russ Tedrake, an MIT professor and cofounder and CEO of Walden Robotics. He argues robotics companies hold something labs don't: deployment experience in real workplaces, and site-specific data that gets more valuable as foundation models improve. That reads to us as the real moat. Whoever has robots working in factories learns why they fail, and a general model alone can't tell you that.

So the startups that raised billions on the premise that robots need brains of their own aren't dead, but their pitch probably has to change. Their value is data and operations, less the model.

Our verdict

We think OpenAI is probably right that general models will do most of the "brain" work, and that dedicated robot-model teams are the part most at risk. But the 2027 humanoid will be judged in the physical world, not on a leaderboard. If OpenAI wants that demo to hold up, it needs to keep embodied failure expertise somewhere with real authority, not just a reporting line. Watch for that: not the benchmark numbers, but whether the company can explain, in public, what goes wrong when its robots do.

Debate topics

No topics yet: start the first one.

More stories