A team folded into the model line
This account follows MadRobot's full-text relay of Fortune's exclusive report, corroborated by other secondary reports; ALFW did not access Fortune's original directly.
This is the second time OpenAI has broken up a robotics team. It wound down the first in 2020, with cofounder Wojciech Zaremba later pointing to a shortage of robot data. It rebuilt a standalone team in 2024, alongside a partnership with humanoid maker Figure. Figure ended that partnership in February 2025 to build its own robot AI.
The structure now is different from 2020 in one respect. The people have not been described as leaving robotics. OpenAI says frontier models can now be applied directly to robotics, while its researchers and applied staff continue the work inside different organizations: the ones that build the general models and the ones that ship products.
Read it alongside the context Fortune adds, and keep the two separate. Fortune notes cost and focus pressure at OpenAI, including a reported $10.9 billion adjusted operating loss in its latest quarter. OpenAI did not tie the robotics move to costs in what it told Fortune. Only the research belief is on the record here.
The number, with its conditions attached
The Nvidia result, as described by Fortune, is genuinely striking, and it needs all of its conditions kept attached to it.
Astra was given camera images, robot state and a control interface, and ran inside a Codex session. It succeeded on 74.8% of simulated pick-and-place tasks, against 51.0% for Lakeside, the best dedicated model on that study's reported leaderboard. No robot-specific training was involved. For a general model to beat specialists at all on their own benchmark is not a small thing.
Then the conditions. The cited study evaluates simulation; this reporting supplies no comparable physical-robot result. The best setup cost about $36.40 in tokens per attempt. The simplest version took 18.56 seconds per model call. Those figures describe a research system on a leaderboard, not a machine competing with a factory robot on speed or cost per task.
Guanya Shi, an assistant professor at Carnegie Mellon University's Robotics Institute, told Fortune the move suggests OpenAI sees robotics as no longer a special problem. He also made the point that survives the headline: robots still need small models that run on the machine itself, without a constant internet connection. Russ Tedrake, cofounder and chief executive of Walden Robotics, made the other surviving point in Fortune's report: robot makers hold data from real floors and real deployments, and better base models make that data more valuable, not less.
What to watch instead of the org chart
If OpenAI is right, the next evidence will not be another reorganization. It will be a general model running a physical robot, in public, at a speed and a cost per task that a buyer can compare with a dedicated system. Fortune reports that a humanoid demonstration remains targeted for 2027; the reporting does not specify that it will establish those performance comparisons.
Until then, the honest reading is narrow. A general model beat dedicated robot models in simulation, by a wide margin, under conditions the reporting does not extend to a warehouse floor. OpenAI has reorganized around the belief that the transfer is coming. That is a bet, stated in staffing rather than in a product anyone can buy.
ALFW did not access Fortune's original report directly; this account follows a full-text relay of it, corroborated against other secondary reports. The Nvidia study figures are as reported by Fortune.
Sources
- [1] Fortune, exclusive report, November 13, 2026 (OpenAI spokesperson confirming the reorganization; CoRL booth 208; expert comment from Guanya Shi and Russ Tedrake) — read via MadRobot's full-text relayRead source
- [2] Fortune exclusive, corroborated by ExplainX, November 13, 2026Read source
- [3] Nvidia Research study, October 14, 2026 (RoboLab-120 figures) — as reported by Fortune; the paper was cited, not linked, in the coverage, and ALFW has not read it directly.