By Sarah Xiao 
The champagne-coloured Optimus that Tesla has been showing since last year is not the next generation. Elon Musk said so himself: it is a 2.5, a transitional build, and “we haven’t shown Optimus V3 yet — it will be spectacular.”
That distinction matters more than it sounds. A transitional robot tells you what Tesla thinks it has already solved, and the footage Salesforce founder Marc Benioff shot in a Tesla office tells you what it has not. In the clip, the gold Optimus stands next to a phone running xAI’s Grok, answers questions out loud, and then takes a slow, uncertain walk down a corridor to look for a Coke.
The interesting part is not that a humanoid can talk. It is where the talking stops and the walking starts.
What 2.5 actually changed
Strip out the colour and the changes are external. Compared with V2, the shell edges are smoother, the joint covers fit tighter, and the seams read as deliberate parting lines rather than gaps. The metal finish is more even, the shoulders are rounded, and there is less exposed cabling and actuator hardware than before.
Tesla’s own framing of V2 was a robot doing repetitive work in controlled settings — a black V2 handing out popcorn at the Tesla diner, a unit moving parts inside a factory. The 2.5 body moves the silhouette from boxy to curved, which is a design decision with an engineering motive behind it: a machine that has to move through rooms built for people benefits from not having corners that catch on things, and a machine that has to be accepted by the people in those rooms benefits from not looking like laboratory equipment.
None of that is a capability claim. It is the body being prepared for one.
The test was an office, not a cell
Benioff’s video is worth reading closely, because the setting is unusual. Most humanoid demonstrations happen in a staged cell: a table, a bin, a taped floor, and a camera. This one happened in a working office, with a human walking beside the robot and asking it for a Coke.
The exchange went roughly like this. Benioff asked what it was doing; Optimus replied that it was not busy and asked what it could help with. Asked where to find a Coke, it said it did not have real-time information — then paused, and offered to walk him to the kitchen to check.
That pause is the whole story. The language layer answered fluently, and then the robot had to convert a sentence into a route through a building it had never mapped.
The walk itself was slow. Musk, standing nearby, intervened twice: first to explain that the robot needed to confirm more rooms and was “a bit paranoid about space right now,” then to insist, as the robot shuffled along the corridor, that it can walk faster.
Where the latency comes from
A voice assistant that answers in a second and a robot that moves in a second are not the same achievement, and the difference is not compute.
Language models are trained on text that already describes the world. A robot has to build the world it is about to walk through: where the doorframe is, how wide the corridor is, whether the floor in front of it is clear, and what “the kitchen” means in a building whose layout nobody has given it. Each of those is a perception and mapping problem with a physical failure mode. When the model is unsure, the safe behaviour is to stop and ask — which is exactly the behaviour Musk described as paranoia.
That is not a bug in the demo. It is the current state of embodied systems: the semantic layer is close to solved, and the spatial layer is not. What looks like hesitation is the robot declining to guess.
The gap a demo cannot close
Tesla has an advantage no competitor can copy quickly: its own factories, its own cells, and a manufacturing organisation that already ships complex hardware at volume. It also has a data problem that the factory does not solve. A robot that has learned one workstation has learned one workstation.
The competitive pressure is coming from the other end of the price curve. Chinese manufacturers are now selling humanoid robots in the 30,000 to 40,000 yuan range, roughly the price of a compact car, and shipping them into education, exhibition and light service work. Those machines are less capable per unit — and they are being deployed in numbers while Tesla’s units are still being walked down corridors.
Meanwhile reporting around the programme has pointed at familiar obstacles: production bottlenecks, yield, hand dexterity, and turnover among the engineers running the programme. Musk has announced ambitious volume targets more than once, and the schedule has slipped each time. That is not unusual for a first-generation hardware programme; it is unusual for a company that has made the programme central to its long-term story.
What is actually known about V3
Not much, and that is deliberate. Musk has said only that V3 exists and will be impressive. The most concrete public artefact is a photograph shared by Sawyer Merritt, a longtime Tesla watcher, showing what may be the clearest view yet of the V3 face — reportedly built around a full OLED display supplied by Samsung.
A face display is a design choice, not a capability. It signals that Tesla is thinking about the interaction surface — how the machine reads to a person standing in front of it — rather than only about the manipulator. Whether the hands have improved, whether the walking speed has improved, and whether the robot can be left alone in a room for an hour are the questions that matter, and none of them are answered by a face.
Three things to watch
One: whether V3 ships with hands that do useful work. Dexterity has been the recurring bottleneck in every humanoid programme, Tesla’s included. A generation that arrives with better hands and the same body is a bigger step than one with a new face and the same hands.
Two: whether the walking gets boring. The most reliable signal of progress is not a faster sprint on stage but a robot crossing a real building at a normal pace without a handler narrating its confidence. When that stops being noteworthy, the spatial layer is working.
Three: whether volume arrives before the price gap closes. A ¥35,000 humanoid in a showroom and a $20,000-plus Optimus in a factory are not competing for the same purchase order today. They will be within a couple of years, and Tesla’s argument has to be capability per unit of labour, not capability per demo.
The gold Optimus is a good-looking transitional build, and the Benioff footage is the most honest look at where the programme actually is: a machine that can hold a conversation about a Coke and then needs to be reassured about the corridor. The conversation is nearly done. The corridor is not.
Theodore Li