The checkout OpenAI demonstrated this week lasted about eleven seconds. A user asked for a specific model of coffee grinder, watched a short list of sellers appear with prices, chose one, pressed a finger to the sensor and was done. No card form, no address entry, no second app.
ChatGPT Wallet went live for Pro subscribers in the United States on Nov. 13. Users can link a Visa, Mastercard or American Express card, and from that point the company’s shopping agent, called dots, can compare prices and place orders on their behalf.
The mechanics are more interesting than the demo. Payment runs on a single-use token that is generated for one transaction. Neither the agent nor the merchant receives the card number, and the confirmation step requires Face ID or a fingerprint on the device. If the token leaks, it is worthless after the sale it was minted for.
That design is a concession to an obvious problem. An agent that can browse on your behalf can also be manipulated on your behalf, by a page that instructs it to buy the wrong item at the wrong price. Requiring a biometric confirmation before money moves puts a human in the loop at the only moment that matters, and it lets OpenAI argue that the model never holds payment credentials.
What the merchant sees is a tokenized card payment, which is to say something close to what they already accept. That was the point. Any design that required retailers to adopt a new rail would have died in procurement. The networks, for their part, get a new source of transaction volume that arrives without them having to build the assistant.
The unresolved questions are the ones that show up after a bad purchase. If the agent buys a counterfeit, who handles the dispute? If the price shown differs from the price charged, whose record is authoritative? There is also the smaller matter of a phone left unlocked on a table, which makes the biometric check less meaningful than it sounds.
OpenAI has published terms that assign liability to the user for authorized sessions. It has not said much about what counts as authorization when an agent acts on a prompt written by a stranger.
Retailers are watching that gap closely. Large merchants can absorb a disputed charge. Small ones cannot, and they have spent a decade learning that a new checkout flow usually arrives with new chargeback rules attached.
For the card networks, the calculation is simpler. Agentic commerce is coming whether or not they participate, and the version that runs on their tokens keeps them in the transaction. Two people familiar with the discussions said the networks pushed for a design in which the credential never leaves the network’s own tokenization system, which is roughly what OpenAI built.
The second piece of news this week pointed in a different direction. Fortune reported that OpenAI has disbanded its standalone robotics team. The group’s work will be absorbed into the model and product organizations, and the people will report into lines that build software rather than machines.
The company’s explanation, as described by people close to it, is that the general models have reached the point where they can drive hardware directly. A vision-language model that can look at a scene, plan a grasp and emit the commands is a robotics stack, and it does not need a robotics department to maintain it. That is a real argument, and it is also the argument every lab makes when it decides to spend its compute elsewhere.
The timing is what makes it a story. OpenAI has been signing compute agreements at a pace that has alarmed its own investors, and every dollar of training capacity committed to a robot arm is a dollar not committed to the next model. Retiring the team frees people and, more importantly, frees attention.
Analysts said the two announcements are best read as a single statement about where the company thinks value will be captured. “Payments is a place where a model can take a cut of a transaction without owning any inventory,” said an analyst who covers the company and asked not to be named because her firm restricts public comment. “Robotics is a place where it has to own the factory.”
That comparison is unkind to the robotics team and probably fair to the balance sheet. A wallet that takes a fraction of a purchase scales with usage and needs no capital. A robot needs a supply chain, a service network and a customer willing to let a machine move near people.
There is a counterargument inside the industry, and it is not trivial: the companies that own the physical interface own the data that the next generation of models needs. Walking away from hardware narrows that channel, and the decision will look different in three years if the models run out of text.
For now, the wallet is the product with a launch date and a support page. The robot is a memo. Users in the United States can link a card today, and the company has said nothing about when the feature reaches other markets or other subscription tiers.
The number to watch is not the volume of purchases. It is how often users decline the biometric prompt, which is the only clean measure of whether people trust an agent with their money.
