What OpenAI actually changed
As The Decoder reported, the standalone robotics division is gone for the second time. Research staff now sit inside the core research group led by Chief Research Officer Mark Chen. The people working on applied robotics and hardware moved to the Applied organisation under president Greg Brockman. OpenAI presents this as consolidation rather than retreat. Robotics stays part of the long-term plan, and the argument is that a general frontier model should eventually drive a robot the same way it drives a chat window.
The empty booth at CoRL in Austin is the detail that sticks. A booked booth nobody staffs is a public signal, and it landed in the same weeks as the internal move.
The benchmark that made the argument
There is one hard number behind the general-purpose case. In the Nvidia simulation of manipulation tasks described by The Decoder in November 2026, pick-and-place work, the general model GPT-6 Astra completed 74.8% of attempts. The best specialised robotics model in the same test reached 51.0%. That is a gap of 23.8 percentage points, and it is the sort of result that makes a separate robotics stack look like a detour. The Nvidia test used GPT-6 Astra, not the current flagship GPT-6.1 Terra.
What one attempt costs
Then the invoice arrives. The Decoder's November 2026 report puts the inference cost for GPT-6 Astra on that benchmark at $36.40 per manipulation attempt, about €33.40 at $1.09 to the euro. Success rate matters more than raw cost, so divide one by the other: at 74.8%, the cost per successful pick works out to roughly $48.70, or €44.70.
The Decoder's report does not publish the token cost of the specialised model at 51.0%, so the other side of that comparison cannot be computed. That is a gap worth naming, because a cheaper model at a lower success rate can still lose on cost per completed task.
Scale it up. A thousand attempts cost $36,400 in tokens alone. At the stated success rate, 252 of those attempts fail and need a retry, and every retry is billed like any other call.
Latency: 18.56 seconds per call
The second figure is harder to design around. The same Decoder report measured each model call at 18.56 seconds. If a controller waits for the model before every action, one arm cannot exceed about 194 decisions an hour, and that assumes zero time for the robot to move, grasp or settle. A line that needs one model decision per pick is therefore capped near 194 picks an hour.
Run that continuously and the token bill reaches roughly $7,060 an hour, about €6,480. Those numbers do not describe a production cell. They describe a research loop that still has to solve latency and cost.
The European angle: open weights and a 2028 deadline
European robot builders do not have to wait for OpenAI either way. Mistral Large 4 shipped with open weights on October 29, 2026. According to Mistral's pricing page, it costs $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07. According to OpenAI's pricing page, the current flagship GPT-6.1 Terra lists $1.00 input and $5.00 output per million tokens. For a team testing whether a general model can drive a cell, that output price is the difference between a metered API bill and a fixed hardware cost. The Mistral weights can be downloaded and pinned to a fixed version.
Regulation sets the clock. Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, the compliance deadline for Annex I embedded high-risk systems (machinery with safety-relevant AI) moved to August 2, 2028. Annex III standalone systems moved to December 2, 2027. Any European company putting a learned policy into a robot in the next two years is building toward those dates. A model whose behaviour shifts with every provider-side update is a documentation problem as much as an engineering one, and open weights with a pinned version remove part of it. Article 50 transparency duties have applied since August 2, 2026, which means synthetic output and chatbot interaction now carry labelling obligations.
What OpenAI has and has not confirmed
OpenAI frames the reorganisation as consolidation, not retreat. Robotics remains part of the long-term plan, and the company says the move should speed progress toward AGI. But nothing in the public reporting says OpenAI has built a robot, bought a robotics company, or set a demonstration date. What exists is a simulation benchmark, a cost per attempt, a latency figure, and an org chart with no robotics box on it.
If you want to know what a general model actually costs to run on your own hardware rather than through an API, our AI Arena rig publishes VRAM use and tokens per second for local models on a 16 GB card. That is the number that decides whether a general-purpose policy is affordable at the edge of a factory floor.
Is OpenAI leaving robotics?
The company says no. The teams were reassigned rather than the strategy retired: research staff joined the core model group, applied robotics and hardware staff joined the Applied organisation. What disappeared is the separate division, not the stated long-term interest in robots.
Do the 74.8% and 51.0% results apply to real robots?
No. They come from an Nvidia simulated manipulation benchmark with pick-and-place tasks. Simulation-to-real transfer is not measured by those figures, and the 18.56 second call latency is a further obstacle outside a simulation loop.
Could a European team self-host a general model for robot control instead?
Partly. Open-weight options such as Mistral Large 4 can be downloaded and pinned to a fixed version, which helps with AI Act documentation. Whether the model fits your hardware depends on size and quantisation, and VRAM is usually the binding limit on a single workstation GPU.
