OpenAI broke up its robotics team on the strength of a 74.8% score. The test was a simulation.
OpenAI says GPT-6 Astra can run robots directly, so a stand-alone team isn't needed. The number behind that claim comes from a simulator that stops the clock while the model thinks.
Fortune reported on November 13 that OpenAI has disbanded its stand-alone robotics team. The company confirmed it when Fortune asked, and a spokesperson called the change a sign of commitment to robotics, not a retreat from it.
The reasoning is the interesting part. According to Fortune’s account of OpenAI’s statement, researchers started the year assuming robotics would build on video models like Sora. Since then, OpenAI says, its general frontier models, GPT-6 Astra included, have shown they can be applied to robots directly. So there is less reason to keep a separate group building robot “brains”.
That’s a big claim to hang a reorganization on, so I went looking for the evidence Fortune cites.
Where the people went
Per Fortune, the robotics research team now reports to Chief Research Officer Mark Chen. The applied robotics and hardware teams moved into the Applied organization under President Greg Brockman. OpenAI hasn’t said how many people were affected, and nothing in Fortune’s report mentions layoffs.
The timing was messy. OpenAI had registered booth #208 at CoRL, a robotics conference in Austin, this week. Fortune says it found the booth empty, with no OpenAI logo, though a few former team members were standing around and still drew a crowd.
Robotics at OpenAI has been on and off. Fortune notes the company wound down its first team in 2020, rebuilt one in 2024 with a Figure partnership that ended in February 2025, and now says it’s building its own humanoid, with Sam Altman targeting a first public demonstration in 2027.
The number: 74.8%
Fortune points to a study from Nvidia’s Seattle Robotics Lab, in which an Astra-powered agent succeeded on 74.8% of tasks in a simulated manipulation benchmark. The best specialized robot model on that leaderboard got 51.0%. Nvidia’s own project page for the study has the fine print, and it changes how the headline reads.
| Setup | Success rate | Cost per successful episode | Latency per model call |
|---|---|---|---|
| Astra, agentic policy | 74.8% | $36.40 (30.9M tokens) | 13.37 s |
| Astra, orchestrated policy | 66.5% | $17.72 (11.0M tokens) | 11.50 s and 2.74 s |
| Astra, direct policy | 66.3% | $1.96 (160k tokens) | 18.56 s |
| Lakeside (best specialized entry shown) | 51.0% | not listed | not listed |
| π0.5 | 28.0% | not listed | 46.96 ms |
Figures are from Nvidia’s study page, and the 51.0% and 28.0% rows come from the same RoboLab-120 leaderboard it shows. Nvidia estimated the dollar costs from OpenAI’s API prices published on October 1, 2026.
My read: three things stand out. The top score needs a persistent Codex session that burns roughly 30.9 million tokens per successful run. The cheap version costs $1.96 but scores 8.5 points lower. And a call that takes 18.56 seconds is a long wait for a machine that’s supposed to catch a falling cup.
Nvidia says as much itself. Its page states that general-purpose models “suffer from long latency for practical real-world deployment” and that the simulation is paused during inference, so the test doesn’t show whether the policies can react while the world keeps moving. It also says the benchmark doesn’t cover contact-rich work like insertion or assembly.
A real-robot check
There’s a separate preprint on arXiv, LLM4Robot-Bench, that ran Astra zero-shot on 16 household tasks in both simulation and on a real robot, three trials per task. It’s a small test, three trials per task. The authors report success of 80.0% on the easiest tier of tasks in the real world, 47.6% on the middle tier and 0.0% on the hardest, where the robot has to change contacts and control force together. Across all 16 tasks the average was 45.8% in the real world and 41.7% in simulation.
Those two papers use different benchmarks, so the 74.8% and the 45.8% shouldn’t be set side by side as if one corrects the other. My read: they point the same way. Astra can do some pick-and-place work with no robot training, and falls off fast when the job needs a feel for force.
My read
My read: the reorganization is a reasonable bet and a thin proof. OpenAI is under cost pressure. Fortune reports the company is trying to close a new funding round, reportedly lost $10.9 billion in its latest quarter, and has cut side projects like Sora. A separate robotics group is easy to fold into a “general models do everything” story when money is tight, and I can’t tell from the outside how much each reason weighed.
The outside voices in Fortune’s piece are split in a useful way. Carnegie Mellon’s Guanya Shi reads the move as robotics becoming “no longer a special problem”. MIT’s Russ Tedrake, who runs Walden Robotics, argues robot makers still hold the deployment data that labs lack. Shi also still sees a need for smaller models that run on the robot itself, without a constant network connection.
If you want to check the claim, watch for two things: whether Astra-driven robots show up outside a simulator before 2027, and whether OpenAI’s humanoid demo runs on a general model or a specialized one. For how much “best model” claims can slip between a launch post and a leaderboard, see our look at Astra’s computer-use numbers.
Sources
- Fortune: Exclusive, OpenAI disbands its stand-alone robotics team as its general models start running robots
- Nvidia Research: Astranomy, an empirical study of GPT-6 Astra as a native robot policy
- arXiv: LLM4Robot-Bench, how far can GPT-6 Astra go in robotic manipulation
Note on sources: OpenAI published no statement of its own, so the company’s reasoning comes from what its spokesperson told Fortune. The Nvidia numbers are from its own study page and have not been independently reproduced.