Overview
When an agent buys things for you, the first question was never whether it clicks the wrong button. It is who pays when the money goes to the wrong place. OpenAI's answer, published Nov 13, is to keep the card number away from the agent entirely.
The mechanism splits one payment into two halves. A user links a card once. For each purchase, dots puts the item, the total and the card to be charged in front of them, the user confirms with face or fingerprint, and the card network issues a payment token carrying the number, security code and expiry. The money leaves the card the user picked. The merchant receives a one-time credential, and the agent never holds the real number at any point. The card's own rewards, returns and dispute handling keep working, which matters more than the token itself, because it means a problem is resolved under rules the user already knows.
All three networks joined on the same day: American Express, Mastercard and Visa. That combination suggests this is not one network's experiment but three clearing paths opening together.
Availability is narrower than expected. US only, ChatGPT Pro, 18 or older, a US-issued card, and spending limited to sites in the US that support dots. No fee is disclosed and no spending cap is published; spending limits, category restrictions and time windows appear only in the roadmap. In other words this is supervised shopping with per-purchase approval, not delegation.
The merchant side has four tasks, all in the Commerce Partner Playbook: claim the business and upload a product feed; configure the CDN or firewall to recognise Web Bot Auth signed requests; work with payment and fraud vendors, with Forter, Riskified, Sardine, Signifyd and Adyen, Checkout.com, Stripe named; then build a plugin or host WebMCP tools. OpenAI itself admits two failures will happen in the wild: a site may block a legitimate request as a crawler, and a fraud system may flag a real purchase as fraud.
So the real bottleneck is not payment, it is identity. Whether a merchant's edge can verify the Web Bot Auth signature decides if the agent completes checkout or hits a 403. The wallet answers whether I let it spend. The signature answers whether the site believes it is who it says.
Key Features
- Payment tokens instead of card numbers: After approval, the card network issues a payment token holding the number, security code and expiry. Merchants and agents both see only the credential, and the original number never reaches either side.
- Per-purchase review with biometrics: Every purchase shows the item, total and payment method first, then clears with face or fingerprint. One approval covers one purchase, with no blanket consent granted in advance.
- Three card networks from day one: American Express, Mastercard and Visa all went live at once, which means the product rides existing clearing rails rather than standing up a separate balance.
- Existing card protections survive: Rewards, returns, chargebacks and dispute handling follow the issuer's existing rules, so having an agent buy something does not cost the cardholder those protections.
- A merchant onboarding checklist: Claim the business and upload a product feed, teach the CDN or firewall to recognise Web Bot Auth signed requests, connect payment and fraud vendors, then ship a plugin or WebMCP tools, all four spelled out in the Commerce Partner Playbook.
- A second risk layer beyond the wallet: OpenAI names two failure modes: sites blocking legitimate requests, and fraud systems flagging genuine purchases. Both currently depend on the merchant's own rules and third-party fraud vendors.
Use Cases
- Buying everyday goods from US stores that support dots, with each purchase confirmed before payment
- Letting an agent compare prices across sites, then checking out through the same approval flow
- Turning a repeat order into one sentence, gated by a single biometric check at payment
- Sourcing something that needs repeated price checks, reviewing the total before choosing which card to use
Pros
- Card numbers never reach the agent or the merchant, narrowing risk and capability at the same time
- Approval is granular to a single purchase, closer to how most people actually think about money
- Riding existing card rails means rewards, returns and dispute rights do not need rebuilding
- Merchants do not have to rebuild checkout, with the hard part concentrated in verifying signed requests
Pricing
The wallet adds no fee and OpenAI publishes none. Access requires a ChatGPT Pro subscription and is limited to the US, users 18 and over, and US-issued cards. No spending cap is published; spending limits, categories and time windows sit on the roadmap. Merchant-side cost depends on the payment and fraud vendors they already use, with no extra charge from the wallet.
Summary
It splits agent shopping into two problems: payment through a one-time token, authorisation through per-purchase approval, so the card number never has to leave the user. What decides whether it works is not the wallet but whether sites verify Web Bot Auth signatures and whether fraud systems treat legitimate purchases as fraud. Coverage today is US Pro users only, which makes it a supervised shopping pilot rather than delegation.
Version History
- Release (2026-11-13): OpenAI launched ChatGPT Wallet, where the card network issues a payment token for dots purchases so neither merchant nor agent sees the real card number. A user links a card once, reviews the item, total and payment method before each purchase, then approves with face or fingerprint. Built with American Express, Mastercard and Visa, limited to US ChatGPT Pro subscribers aged 18 and over spending at US sites that support dots. No fee or spending cap was disclosed, with spending limits, categories and time windows left to the roadmap. A merchant-side Commerce Partner Playbook shipped the same day covering business claims and product feeds, Web Bot Auth signed requests, payment and fraud vendor integration, and plugins or WebMCP tools.