74.8% in sim, $36 a task: OpenAI killed its robot team

A general-purpose chatbot scored 74.8% on a robot manipulation benchmark, beating the best purpose-built robot model by 23.8 points — and OpenAI disbanded its robotics team the same week. The catch is in the fine print Nvidia published: the winning setup costs $36.40 in tokens per successful task and takes 18.56 seconds per inference call.

What Fortune reported

On November 13, Fortune reported in an exclusive — confirmed by OpenAI when asked — that the company's stand-alone robotics team no longer exists. The researchers moved into OpenAI's main research organization under Chief Research Officer Mark Chen. The applied robotics and hardware staff moved into the larger Applied organization under President Greg Brockman. OpenAI gave no headcount and published no statement of its own.

The reorganization happened so recently that OpenAI had already registered a booth at CoRL, the top academic robotics conference in Austin this week. Booth #208 was empty when Fortune's reporter visited.

The stated reason is the bet, not a retreat: general-purpose models such as GPT-6 Astra can now be applied to robots directly, so a separate team building robot-specific models is no longer the fastest way forward. A spokesperson said robotics remains core to the long-term vision, and CEO Sam Altman is still targeting a first public demonstration of OpenAI's own humanoid robot in 2027.

This is the second time OpenAI has done this. It wound down its first robotics team in 2020, rebuilt a stand-alone team in 2024 with a partnership with Figure to develop models for humanoid robots, and that deal collapsed after roughly 18 months.

The number that killed the team

The evidence Fortune cites is Nvidia's Astranomy study, published October 14, 2026 (research.nvidia.com). Nvidia's Seattle Robotics Lab asked a blunt question: can a general-purpose multimodal model, with no robot-policy fine-tuning at all, close the loop on real manipulation tasks if you specify the control interface in the prompt? They ran GPT-6 Astra on RoboLab-120, a 120-task simulated pick-and-place benchmark, using a DROID robot at 15 Hz control.

The headline result is real. I read the study itself, not just the coverage:

Policy (GPT-6 Astra, zero robot training) Success rate Agentic Policy (persistent Codex session) 74.8% Orchestrated Policy (Astra as monitor) 66.5% Direct Policy (stateless API calls) 66.3% Best specialized robot model (Lakeside) 51.0% 

The same model jumps from 66.3% to 74.8% purely by changing how it is called: a persistent Codex session with retained history instead of stateless API calls. The model didn't get better. The harness did.

A smaller real-robot data point backs the direction: on September 4, RoboCurve had Astra control two physical arms to pick up a block and place it into a bowl, succeeding 19 out of 20 trials. That is a toy task, but it ran on hardware, not in a simulator.

The fine print

Here is what the headline leaves out, straight from Nvidia's own report:

• The winning setup costs $36.40 per successful episode. Agentic Policy burns 30.9 million tokens per successful episode. Direct Policy costs 160,000 tokens and $1.96 — roughly 190x fewer tokens for 8.5 points less success. The cheap config is the interesting number, not the expensive one.
• 18.56 seconds per inference call. Direct Policy averages that per call, and inference consumes 94.9–97.8% of estimated episode wall time. A robot that waits 18 seconds between thoughts is not a robot. It is a slideshow.
• The simulation is paused during inference. The clock stops while the model thinks. On a real robot, the world does not pause. Nothing in the 74.8% tells you what happens when the object keeps sliding while Astra is still deciding.
• It is one benchmark. RoboLab-120 is pick-and-place in simulation. It does not cover contact-rich real-world tasks, and Nvidia says so.

The dissenters are on the record too. Guanya Shi at Carnegie Mellon's Robotics Institute, Russ Tedrake, and Pete Florence — cofounder of Generalist, a startup training general-purpose models for robots — all told Fortune that general models have not solved embodied intelligence. And the general-model lead is already being chased: a simulation-trained policy called ZAP, which never sees a real robot during training, matches Astra at 74.0% on RoboLab-120.

The team wasn't disbanded because robots are solved

It was disbanded because the org bet moved. If a general model with a prompt harness beats your specialist by 24 points in simulation, the specialist team becomes a cost center — even if the general model costs $36 a task and thinks in 18-second increments. OpenAI is betting that the token bill and the latency collapse faster than a dedicated team could close the capability gap.

That bet has a deadline and a test: the 2027 humanoid demo. A real robot, in real time, with no paused simulation and no $36-per-grasp budget. If Astra's successor can pick things up at human speed for cents, the disbandment was a good call. If the demo is another carefully staged simulation with a live audience, we will know the 74.8% was a benchmark artifact and the team was dissolved for a press release.

The honest scoreboard is not the leaderboard. It is dollars per successful grasp, seconds per decision, and whether the robot works when nobody pauses the world for it.

Sources: Fortune exclusive, "OpenAI Disbands Standalone Robotics Team," November 13, 2026, as reported by Fortune and secondary coverage (madrobot.blog, explainx.ai, aiweekly.co); Nvidia Research, "Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy," October 14, 2026 (research.nvidia.com/labs/srl/projects/astranomy/) — success rates, token costs, and inference latency are from the study itself; RoboCurve September 4, 2026 experiment via robonomics.substack.com; ZAP results via zero-real.github.io.