AI
“OpenAI disbanded its stand-alone robotics team and moved researchers into its main AI and applied research divisions, according to Fortune." Accompanying post text: "OpenAI just disbanded its stand-alone robotics team after finding GPT-6 Astra can directly control robots"; "In an Nvidia study, Astra achieved a 74.8% success rate on a…”
Plain restatementFortune reported that OpenAI dissolved its separate robotics team and reassigned its staff to the company's main research organization and its Applied organization. Secondary assertions: the change followed a finding that a general-purpose model can control robots without robot-specific models; an Nvidia study measured 74.8% success for that model on a simulated benchmark against 51% for the best specialized entry; and OpenAI plans a humanoid robot demonstration in 2027.
This one largely holds up. Fortune did publish an exclusive on 13 November 2026 reporting that OpenAI dissolved its separate robotics team, and OpenAI confirmed the change when Fortune asked, saying robotics remains central to its plans. Research staff moved to the main research group under Mark Chen and applied robotics and hardware staff moved to the Applied group under Greg Brockman, exactly as the post describes. The benchmark numbers are also real: an Nvidia study measured 74.8% success for GPT-6 Astra on a simulated 120-task benchmark against 51% for the best specialized entry on that leaderboard. What the post leaves out matters, though. That 74.8% came from one particular scaffolded setup, other setups of the same model scored as low as 50.2%, the simulation was paused while the model thought, Nvidia ran only five attempts per task and said so, and the approach was slow and expensive. The 2027 humanoid demonstration is a target Sam Altman named in an interview, not a scheduled event, and one OpenAI employee has publicly objected that the robotics work was merged rather than disbanded.
OpenAI [drifted from the evidence:] disbanded its [drifted from the evidence:] stand-alone robotics team and [drifted from the evidence:] moved researchers into its main [drifted from the evidence:] AI and applied research [drifted from the evidence:] divisions, according to Fortune." Accompanying post text: "OpenAI just disbanded its [drifted from the evidence:] stand-alone robotics team after finding [drifted from the evidence:] GPT-6 Astra can [drifted from the evidence:] directly control robots"; [drifted from the evidence:] "In an Nvidia study, [drifted from the evidence:] Astra achieved a 74.8% success [drifted from the evidence:] rate on a simulated [drifted from the evidence:] robotics benchmark, [drifted from the evidence:] compared with 51% for [drifted from the evidence:] leading specialized [drifted from the evidence:] models"; "OpenAI [drifted from the evidence:] is also developing its own humanoid robot, [drifted from the evidence:] with a public demonstration [drifted from the evidence:] planned for 2027.
[added by the neutral restatement:] Fortune reported that OpenAI [added by the neutral restatement:] dissolved its [added by the neutral restatement:] separate robotics team and [added by the neutral restatement:] reassigned its [added by the neutral restatement:] staff to the company's main research [added by the neutral restatement:] organization and its [added by the neutral restatement:] Applied organization. Secondary assertions: the change followed a finding [added by the neutral restatement:] that a general-purpose model can control robots [added by the neutral restatement:] without robot-specific models; an Nvidia study [added by the neutral restatement:] measured 74.8% success [added by the neutral restatement:] for that model on a simulated benchmark [added by the neutral restatement:] against 51% for [added by the neutral restatement:] the best specialized [added by the neutral restatement:] entry; and OpenAI [added by the neutral restatement:] plans a humanoid robot demonstration [added by the neutral restatement:] in 2027.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- The Fortune article exists, is correctly attributed, and is correctly dated. It is an exclusive by Wen Shao published 13 November 2026.
- Fortune reports that OpenAI dissolved the stand-alone robotics team and reassigned its researchers, and that the company confirmed the reorganization when asked about it by Fortune. This is an on-the-record company confirmation, not an anonymous leak.
- The destinations named in the caption match the reporting: research staff to the main research organization under Mark Chen, applied robotics and hardware staff to the Applied organization under Greg Brockman.
- The caption's statement that OpenAI is not abandoning robotics matches the company's position as reported: a spokesperson said robotics remains a core part of OpenAI's long-term vision and portrayed the reorganization as a commitment rather than a retreat.
- The 74.8% figure is accurately quoted from the Nvidia study, and the post correctly labels the benchmark as simulated.
- The 51% comparator is accurately quoted. It is the top learned policy on the RoboLab-120 leaderboard, Lakeside.
- OpenAI is developing a humanoid robot and Altman has publicly named 2027 for a first demonstration.
- The post presents 74.8% as what "Astra" scored. The study's own breakdown shows that figure comes from one specific scaffolded configuration, the Agentic Policy. Nvidia's table lists Direct Policy at 66.3 for end-effector actions and 50.2 for joint positions. The lowest Astra configuration is roughly level with the 51.0% specialist it is being contrasted against, so the headline gap depends on which harness is quoted.
- The image headline says Astra "can directly control robots." In the study, the configuration actually named Direct Policy is not the one that produced 74.8%. The top number came from an agentic setup with tool access and a persistent session, which is more scaffolding than the phrase "directly control" conveys.
- The two numbers were produced under different protocols. Nvidia ran 5 episodes per task with fixed seeds, which it states is insufficient to characterize per-task variability, while the comparator leaderboard entry is listed at 25 episodes per task across 3000 trials, and the Astra result was evaluated outside the leaderboard. The post presents the two as a like-for-like scoreboard.
- The post gives the success rate with no indication of what producing it required. Nvidia publishes token cost and latency tables for this evaluation and notes that increasing reasoning effort showed little performance change but incurred higher latency and token cost. Secondary coverage of the study puts the agentic configuration at roughly 36 dollars and many minutes per episode, which is a different picture from a bare percentage.
- The post does not carry Nvidia's own central caveat, that simulation is paused during inference, so the evaluation does not test whether the policies can respond effectively while the environment continues to change. That caveat bears directly on whether the number says anything about physical robots.
- "a public demonstration planned for 2027" is firmer than the source. Altman's wording was "I think we'll have a cool demo in 2027", and coverage notes the 2027 target refers to a prototype demonstration, not a product launch. It is a stated target, not a scheduled event.
- Definitional dispute, contested framing: the word "disbanded" is itself in dispute. An OpenAI employee, Michelle Fradin, disputes Fortune's headline, saying robotics research joined the main research team to improve future core models rather than being disbanded. The underlying checkable facts are not in dispute between the two accounts: the separate unit no longer exists and its people now sit in main research and Applied. The remaining disagreement is about what to call that, and the post's own caption happens to state both halves, which blunts the issue. It is worth noting that one write-up reports Fortune's own later social card used "reorganizes" in place of "disbands."
- Headcount is not established. No source located gives the size of the team that was dissolved or the number of people reassigned.
- Whether cost pressure contributed to the decision is reported as surrounding context by some coverage rather than as a stated reason. One write-up records that the spokesperson did not cite costs.
- Whether the Nvidia result is what actually drove OpenAI's decision rests on Fortune's characterization. OpenAI's quoted statement speaks to commitment and acceleration, not to a specific study.
- Whether the 2027 demonstration target survives the reorganization is unresolved beyond OpenAI's statement that it still stands.
- Whether the 74.8% figure transfers to physical robots is open, and the study's authors say their setup does not test it.
The Fortune article exists, is dated 13 November 2026, and says what the post says it says. OpenAI has disbanded its stand-alone robotics team and reassigned the researchers working there to other parts of the company. The company said the move was an effort to accelerate progress on robotics, not abandon it. The company confirmed the reorganization when asked about it by Fortune, and a spokesperson said that robotics remains a core part of OpenAI's long-term vision and portrayed the reorganization as a sign of the company's commitment to robotics rather than a retreat from it. Fortune also reported that this discovery, that general models can be applied directly to robots, is what motivated OpenAI to disband the stand-alone robotics team and move the affiliated researchers onto teams working on more general AI models. On where the people went, multiple independent write-ups of the Fortune piece agree: members of the robotics research team will now report to chief research officer Mark Chen, and the applied robotics and hardware team members were reassigned to the Applied group, which reports directly to OpenAI President Greg Brockman. On the benchmark, the Nvidia study is real and the numbers are correctly quoted. On RoboLab-120, Direct Policy achieves 66.3%, Agentic Policy achieves 74.8%, and Orchestrated Policy achieves 66.5%, and the Agentic Policy exceeds the strongest displayed leaderboard reference, Lakeside, by 23.8 percentage points with Lakeside at 51.0. The independent preprint confirms the comparator: at the time of writing, the strongest learned policies on the RoboLab-120 leaderboard, Lakeside, BiMind v0.5, and FLUX 3 Action, succeed in 51.0%, 49.2%, and 42.9% of episodes. On the 2027 humanoid, the source is a remark by Altman, not a scheduled event. Altman confirmed the company is developing its own humanoid robots with plans to demonstrate an early prototype as soon as 2027, saying "I think we'll have a cool demo in 2027," and the 2027 target refers to a prototype demonstration, not the launch of a product consumers will be able to buy.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Did OpenAI really shut down its robotics team?
Fortune reported, and OpenAI confirmed, that the company dissolved its stand-alone robotics team and reassigned its researchers to other parts of the company. OpenAI described this as an effort to accelerate robotics progress, not abandon it, and one OpenAI employee has publicly disputed the word 'disbanded,' saying the team joined the main research group instead.
Where did the robotics staff go?
Research staff now report to chief research officer Mark Chen in the main research organization, while applied robotics and hardware staff moved to the Applied group under President Greg Brockman.
Is the 74.8% success rate a fair comparison to the 51% specialist models?
The numbers are accurately quoted from a real Nvidia study, but they come from different conditions. The 74.8% figure reflects one specific scaffolded setup called Agentic Policy, other configurations of the same model scored as low as 50.2%, and the test ran only 5 episodes per task compared with 25 for the leaderboard comparator.
Does the benchmark show Astra can control robots in the real world?
The study does not establish this. Nvidia notes the simulation was paused while the model was thinking, meaning the test did not check whether the model can respond while the environment keeps changing, which is central to real robot control.
Is the 2027 humanoid robot demo a confirmed event?
No. It is based on a remark from Sam Altman saying 'I think we'll have a cool demo in 2027,' which refers to an early prototype demonstration, not a scheduled event or a product launch.