OpenAI disbands its robotics team to bet on GPT-6 Astra. The best evidence: 74.8% in a simulator that waits 18.6 seconds per move

OpenAI has disbanded its stand-alone robotics team. Fortune's Wen Shao reported on November 13 that the company confirmed the reorganization: robotics researchers now sit in the main research organization under Chief Research Officer Mark Chen, and the applied robotics and hardware people moved to the Applied organization under President Greg Brockman. A spokesperson framed it as acceleration, not retreat, and gave a technical reason. At the start of the year OpenAI's researchers assumed robotics would build on video models like Sora; since then, the spokesperson said, general-purpose frontier models, GPT-6 Astra among them, have shown they "can be directly applied to robotics."

The timing was tight enough that OpenAI's booth, #208 at CoRL in Austin this week, was still on the floor plan and sat empty.

So the reorganization is, in effect, a bet on a claim we can check: that a general model with no robot training is already a credible robot brain. The best public evidence for it is a study Fortune itself cites, and it is worth reading past the headline number.

What Astranomy actually measured

"Astranomy: An empirical study of GPT-6 Astra as a native robot policy," by Jenai Xuning Yang, Ashna Khetan and Shun Zhang of NVIDIA's Seattle Robotics Lab (a CoRL 2026 late-breaking report dated October 14), puts GPT-6 Astra in charge of a simulated DROID robot, a Franka Panda arm with a Robotiq gripper, on all 120 pick-and-place tasks of NVIDIA's RoboLab-120 benchmark, five episodes each. Astra gets three 384×216 camera views and the arm's measured state, and answers with end-effector targets. It runs in three roles: called directly with no harness (Direct Policy), inside a persistent Codex session that keeps the whole episode in context (Agentic Policy), or as a supervisor over a robot foundation model (Orchestrated Policy).

The results are genuinely strong:

We computed the intervals ourselves (Clopper-Pearson, treating the 600 episodes as independent, which flatters them slightly since episodes of the same task are correlated). Even so, the 23.8-point lead over Lakeside is far outside sampling noise. That is the number behind Fortune's "74.8% versus 51.0%," and it holds.

Two other rows in the same table matter just as much. Change nothing but the action format, from end-effector poses to joint positions, and the same model on the same tasks drops from 66.3% to 50.2%, level with Lakeside. Ask it for B-spline trajectories and it falls to 38.8%. The authors' own takeaway is that "the prompt structure and content matter." Astra's robot skill is real, but a large part of it is decided by the interface someone wrote around it, which is the same conclusion the JustRUA study we covered earlier today reached on a bimanual benchmark, where harness changes alone took GPT-6 Astra from 29.7 to 50.3 on RoboDojo.

The clock that was switched off

The detail that does not make the headline is in the study design: "Simulation is paused during inference," and "no walltime limit or retry limit is imposed." The world waits for the model.

It has to. Direct Policy averages 18.56 seconds per model call; inference is 97.0% of the estimated episode time, and 94.9–97.8% across all Astra configurations. The specialized models in the comparison answer in milliseconds: the report cites 46.96 ms for π0.5 and 713.76 ms for Cosmos 3 Nano. By our arithmetic, Astra's direct calls are about 400 times slower than π0.5's. The authors' own yogurt demo makes it concrete: about 5.2 seconds of arm motion, an estimated 3 minutes 57 seconds once the thinking is added back in.

The authors are candid about what this means. Their limitations section says the setup "does not test whether the policies can respond effectively while the environment continues to change," and that real-time responsiveness "remain[s] an open question." Nothing in RoboLab-120 moves while Astra thinks. A real kitchen does.

Then there is money. The best policy is also the most expensive: the Agentic Policy averages 30.9 million tokens, about $36.40, per successful episode (the summary says "per episode"; the results table and conclusion specify successful ones), against 160,000 tokens and $1.96 for Direct Policy. As our own rough arithmetic, the 449 or so successful Agentic episodes in the main run alone come to about $16,000 at that rate, before counting the failures.

What OpenAI's bet does and doesn't rest on

None of this makes the reorganization strange. Astranomy shows that an unmodified general model can outscore every purpose-built robot policy shown on the RoboLab-120 leaderboard, across 120 pick-and-place tasks, which would have sounded absurd two years ago, and Fortune quotes CMU's Guanya Shi reading OpenAI's move as treating robotics as "no longer a special problem." If capability is the bottleneck, folding robotics into the team that builds the general models is a reasonable bet.

But the evidence so far is capability in a paused, simulated world, on pick-and-place rather than contact-rich work, at tens of dollars and minutes per task. The gap the public studies leave open is latency and cost, and both Shi and MIT's Russ Tedrake point at it in the Fortune piece: Shi still sees a need for small on-board models, and Tedrake argues that robot companies' deployment data becomes more valuable, not less, as foundation models improve. OpenAI's statement claims Astra "can be directly applied to robotics." The published numbers support "can," not yet "directly."