OpenAI was recruiting robotics researchers at up to $500,000 two months ago1, BigGo Finance reports. On November 13, it dissolved the team they were hired into. Fortune broke the news, and OpenAI confirmed it.
The people stayed at the company. Robotics researchers moved into core research under Chief Research Officer Mark Chen. Applied robotics and hardware staff moved into the Applied organization under President Greg Brockman, according to Fortune. The change came so fast that OpenAI's booth, #208 at the Conference on Robot Learning in Austin2, sat empty when Fortune visited.
The trigger was a score. Nvidia's Astranomy study reports that the Agentic Policy of GPT-6 Astra, OpenAI's most intelligent, aligned model for demanding work, hit 74.8%3 on Nvidia's RoboLab-120 simulated manipulation benchmark. Lakeside, the best specialized robotics model, scored 51%. That is a 23.8-point gap, won by a model that also writes code and runs computer tasks.
So read 74.8%3 as a signal with a wide error bar. If you sell a specialized robotics stack, your question is now simple: which piece of it ships free inside the next model?
OpenAI's org chart draws the line
OpenAI sorted its own staff along the absorption line. Above the line sits the work a general model can take over as it improves. Below it sits the work that physics and liability keep hard. Researchers who teach models to act went above the line, next to the team building the general model.
| GPT-6 Astra on RoboLab-120Nvidia Astranomy · simulated manipulation | 74.8%3 |
|---|---|
| Lakeside on RoboLab-120Nvidia Astranomy · best specialized model | 51% |
| Robotics researcher offers1BigGo Finance · top of range, two months before the fold | $500,000 |
| Embodied AI market by 2033Research and Markets · projection | $67.63B4 |
Motor and hardware engineers went below it, into Applied. The split is the clearest evidence of where OpenAI thinks the line sits today. One command owns the brain. The other owns the body.
Above the line, the facts are already stacking up. Natural-language instruction and object recognition are jobs a multimodal model already does for documents and screens. Astra's reported 74.8%3 on RoboLab-120 suggests simulated grasping may join that list. A startup whose main asset is a language interface for robots now competes with a feature on OpenAI's roadmap.
Below the line, the evidence is quieter but just as concrete. OpenAI's careers page still listed roles in control software and actuators on November 14.
The line moves, and every model release pulls a little more of the stack upward. For a builder, the only useful question is how fast.
Two disbandments and a different reason
OpenAI has closed a robotics team before. It wound one down in 2020, and cofounder Wojciech Zaremba cited "limited robot data," AI Weekly reports. It rebuilt a standalone team in 2024 and partnered with humanoid maker Figure. That partnership ended in February 2025, Fortune notes.
The 2026 logic runs the opposite way. In 2020 the bottleneck was data for a dedicated robot model. Now OpenAI says the general model can drive the robot itself. OpenAI told reporters its researchers began 2026 expecting robotics gains to come from video models such as Sora, then found Astra could be applied directly.
The economics explain the bet better than the org chart does. A general model spreads one research bill across coding agents and robots. A dedicated robotics model pays for its own data collection and safety validation, then sells into a single, narrower market. When one asset can serve many markets, a lab will route robotics through it.
Follow the cash. Benchmark scores are cheap to report, and cash flow is hard to fake. If robot makers rent intelligence from one model, the lab collects licensing and cloud fees on every unit shipped. The robot maker keeps the slow-cash parts: motors and service contracts. We think that split favors the lab on margin and the robot maker on switching costs, since a bolted-in warehouse arm rarely changes vendors.
Outside experts disagree on what it means. Guanya Shi of Carnegie Mellon's Robotics Institute told Fortune the move suggests OpenAI sees robotics as "no longer a special problem." Russ Tedrake of MIT, who runs Walden Robotics, argued that startups still hold the edge of real-world deployment experience, AI Weekly reports. Generalist cofounder Pete Florence told Fortune it was unclear that any existing model would be the one that dominates robotics in the long run.
Builders can decide one thing today. Assume generic perception and task planning get absorbed, and stop pitching them to investors as the moat. Two facts would change that call. The first is real-hardware performance at OpenAI's humanoid demonstration, which Sam Altman is targeting for 2027, according to Fortune.
The second is Nvidia's full RoboLab-120 protocol. Nobody yet knows whether a 23.8-point lead in simulation survives changes in friction and payload on a physical arm. If the lead shrinks on real robots, specialized policies regain room. If it holds, the absorption line drops faster than most robotics roadmaps assume.
Who owns robot intelligence by 2033?
By the market math, the prize is large. Research and Markets put embodied AI at $4.67 billion4 in 2025 and projects $67.63 billion by 2033, a 39.7% annual growth rate. Check it on a napkin: 1.4 multiplied by itself eight times is about 14.8, and $67.63 billion divided by $4.67 billion is about 14.5. Globe Market Research starts at $4.8 billion5 in 2026 and reaches $99.8 billion by 2035, at 40.2% a year.
The two firms likely count different things as embodied AI, so treat neither as a precise target. Growth near 40% a year also says nothing about who keeps the money. That depends on where the line sits when the money arrives.
OpenAI's own history gives a cadence. A standalone team closed in 2020, another was rebuilt in 2024, and it was folded into research in 2026. That is three structural changes in six years, each tied to a new view of where robot intelligence comes from. A builder's plan should survive two or three more moves before 2033.
The risk is lopsided. A builder who owns robot data or a certified safety system loses little if general models stall. If general models win, that same builder gets a better brain for its hardware at API prices. A builder whose product is a language wrapper on a robot faces the reverse, because the frontier lab's success case is that builder's failure case.
The likely 2033 shape is a barbell. Model labs hold reasoning and high-level control at one end. Hardware, data and deployment firms hold the physical bottlenecks at the other. Middle-layer robotics software gets squeezed unless it owns something the next release cannot copy.
Sort your robotics stack by absorption risk
The weekend job is a one-page audit plus a test you rerun on every model release.
Step one: open a spreadsheet with one row per layer of your stack. Typical rows are language interface, perception, task planning, manipulation policy, motion control, safety monitor, calibration, fleet software and field service. For each row, write down the asset you own there, such as proprietary data or a certification. Leave the cell empty if you own nothing, and be honest about it.
Step two: mark each row above or below the line with one question. If a frontier model API did this job at Astra's reported simulated success rate next quarter, would a customer still pay you for it? Rows that fail the question go above the line. Treat them as features that will soon be free.
Step three: put closed-loop numbers on every row below the line. Closed-loop means the robot acts, senses the result and corrects, over and over. Track latency per control cycle in milliseconds and the intervention rate. Intervention rate is how often a human has to step in to rescue the robot.
Step four: run a baseline against a general model. Pick your hardest tasks and hand them to a general model API inside your simulator. Expect it to break, and log every place it does, because those breaks map your moat. Save the prompts and the scores so the test reruns in an hour.
Step five: rerun the sheet whenever OpenAI or a rival ships a new model, and move any row that flips. Shift engineering hours from rows above the line to the rows that compound, like real-world data collection and safety logs. Each rerun gives you a dated record of where the line moved.
Our read is that the audit stings the first time. A row your team was proud of will land above the line. That row is the one to stop funding first.
Practise it step by step: Train the full skill in The Dojo
The bottom line
Fund the layers a model release cannot copy.
OpenAI sorted its own people by what a general model will absorb, and builders should run the same sort on their products. Perception, planning and language interfaces look exposed after a 74.8%3 simulated score. Robot data, certified safety and deployment experience still sit below the line. Rerun the audit with every release, and cut funding first for the row your team is proudest of when it lands above the line.