OpenAI's ChatGPT Wallet Lets Its Agents Pay Without Seeing Your Card. Eight Days Earlier, Visa's Benchmark Had Shopping Agents Failing 42.8 Percent of Runs.
On Friday, November 13, OpenAI started rolling out ChatGPT Wallet, a way for its dots agents to pay for things on your behalf. It was built with American Express, Mastercard and Visa. You add a card once, a dot finds what you asked for, and you approve the purchase with a check such as Face ID or a fingerprint.
The clever part is what the agent never holds. After you approve, the payment network issues a token with its own card number, security code and expiration date, and OpenAI says neither the agent nor the merchant sees your real card details.
Eight days earlier, on Thursday, November 5, Visa published a benchmark that put 22 AI models through shopping tasks. They failed 42.8 percent of runs. I think those two announcements belong in the same sentence, because the wallet fixes one problem very well and leaves the other one with you.
What Shipped, and to Whom
The rollout is narrow on purpose. CNBC reports it is open to a limited number of US ChatGPT Pro subscribers, aged 18 and up, with US-issued American Express, Mastercard or Visa cards. Dots, OpenAI's personal agents, were introduced at its developer conference in late September, according to CNBC.
| ChatGPT Wallet at launch | Detail |
|---|---|
| Who can use it | A limited number of US ChatGPT Pro subscribers, 18 and up |
| Cards | Eligible US-issued American Express, Mastercard and Visa |
| What the agent and merchant get | A network payment token, not the card number |
| Approval | Every purchase, with a check such as Face ID or a fingerprint |
| Rewards and disputes | Card rewards kept; protections for eligible disputes unchanged |
| OpenAI's cut | None at launch, per CNBC |
| Merchant integration | “No Wallet-specific integration is required.” |
That last line comes from OpenAI's developer guide for merchants, and it is the most important design decision in the launch. The merchant does not get a new checkout to build. It receives a network token, can use the Payment Account Reference to tie it to the underlying account, and keeps “the order, fulfillment, customer support, and fraud decisions.” The guide also says the credential covers one sale: “The customer consents only to the specific transaction.”
Dots can shop through a cloud browser on the merchant's normal website. To keep firewalls from blocking it, OpenAI signs those browser requests with Web Bot Auth, which uses the HTTP Message Signatures standard, RFC 9421. Its agent access guide says “Akamai, Cloudflare, HUMAN, and Vercel verify ChatGPT request signatures automatically.”
This is also OpenAI's second attempt at checkout. It launched Instant Checkout in fall 2025 and moved away from it, CNBC reports. This time, merchants need no special integration, which looks to me like the lesson learned.
What Visa Measured
Visa's Agentic Commerce Benchmark tested 22 models from 12 labs on 15 shopping tasks, such as buying a TV, a laptop and a personal care appliance. Each model got three tries at each scenario, The Next Web reports.
| Visa benchmark, November 5 | Result |
|---|---|
| Runs failed | 42.8% |
| Best and worst model pass rate | 82.2% and 28.9% |
| Scenarios passed on all three tries | 34.2% |
| Failed runs that still ended in a purchase | 54% |
| Five most vs five least expensive models | 67.1% vs 50.2% pass rate |
| Models that asked the most vs fewest questions (top and bottom five) | 70.2% vs 44.4% pass rate |
The 54 percent figure is the one I keep coming back to. In more than half of the runs where the agent got it wrong, it bought something anyway: a product that did not meet the need, or a purchase made before it had found all the requirements, FourWeekMBA and The Next Web report.
Visa did not name the models. Fortune reported that agents from Google and Meta tied for first at 82 percent, a Claude agent came third at 76 percent and an OpenAI agent was sixth at 64 percent. Fortune does not say that agent is what powers dots, so I would not read the 64 percent as a score for the wallet.
Visa's global head of growth, Rubail Birwadker, put it plainly to Fortune: “We're not there yet from a trust or technology perspective.”
Two Days Earlier, an Airline Said No
On Wednesday, November 11, United Airlines said it is restricting unauthorized third-party AI agents, on its own website and app and through travel agencies and aggregators, Skift reports. United cited the risk of exposing personal information, unintended purchases, reservation changes or MileagePlus redemptions, and conflicts with FAA and DOT rules, according to AirData News. CEO Scott Kirby, as quoted by The Wall Street Journal via Skift: “I don't want to create another middleman.”
Amazon also blocks agents from buying on its platform, according to CNBC and TechCrunch. A wallet only works where the merchant lets the agent in, which is why OpenAI shipped a Commerce Partner Playbook with it.
Our Take
The wallet is good engineering aimed at the wrong half of the problem, for now. Card theft was never the main risk of agent shopping. Payment networks have tokenized cards for years, and a limited-use token with a biometric check is a sound, boring answer. The risk Visa measured is judgment: an agent that buys the wrong laptop with a perfectly secure token has still bought the wrong laptop.
The per-purchase approval is what covers that gap today. You see the items, the total and the card before anything is charged. In my reading, that screen is doing far more safety work than the token, because it puts a person in front of the 54 percent of failed runs that would otherwise have ended in a purchase. It also makes the product less magical. If I have to check every cart, the agent saved me the search, not the decision.
That is why the roadmap matters more than the launch. OpenAI says that over time it expects users to be able to delegate purchases to their agent by setting limits on spending, categories or time periods, Android Authority reports. The day the approval screen goes away is the day Visa's failure rate becomes your problem, up to the limit you set. Chaya Nayak, who leads product for monetization and payments at OpenAI, told CNBC that it will take time for the ecosystem to build the right infrastructure. I agree, and I would add that the infrastructure that is missing is reliability, not rails.
Taking no cut is the other tell. OpenAI is not earning on these purchases yet, so the point at launch is to learn and to get merchants comfortable. American Express's Luke Gebb told CNBC, “It just makes sense to watch it, start it in a contained environment, and then keep rolling from there.” Contained is the right word. United's move shows the other side: some sellers will treat any agent as a middleman to keep out, however well it pays.
For builders running a store, the practical step is small. Check that your CDN or firewall is not rejecting signed ChatGPT browser requests, and tell your fraud vendor to expect agent purchases. For builders of agents, Visa's gap between the models that asked the most questions and those that asked the fewest, 70.2 percent against 44.4, is the clearest product lesson in the whole week: ask before you buy.
Three Signposts for the Next 60 Days
First, scope: whether OpenAI opens the wallet beyond a limited group of Pro subscribers, or to cards and countries beyond the US. Second, delegation: whether pre-approved spending limits arrive, and whether OpenAI publishes any measure of how often dots buy the wrong thing. Third, merchants: whether another large airline or retailer follows United and Amazon in shutting agents out, or whether one signs up as a named partner instead.
Sources: OpenAI announcement, OpenAI Developer Community, OpenAI payments guide, OpenAI agent access guide, OpenAI Commerce Partner Playbook, CNBC, TechCrunch, The Next Web (Wallet), Android Authority, Visa release, The Next Web (Visa), Fortune, Small Business Trends, FourWeekMBA, Skift, AirData News.