OpenAI has put a payment wallet inside ChatGPT, and the rails underneath are the ones you already carry in your pocket: American Express, Mastercard and Visa. Per Cointelegraph's report, the wallet lets AI agents complete purchases for users. It is rolling out to ChatGPT Pro subscribers in the US with eligible US-issued cards.
Our take: the interesting part is not that an agent can click "buy." It is how the permission is shaped. Everything else in agentic commerce is a variation on that one design choice.
What was announced
Here is what the report says, and only what it says:
- ▸Partners: American Express, Mastercard and Visa.
- ▸Authorization: users approve transactions with biometric verification, such as facial recognition or a fingerprint.
- ▸Credentials: the card networks create tokenized card credentials for each approved purchase. The report says this keeps users' actual card details away from merchants and AI agents.
- ▸Availability: ChatGPT Pro subscribers in the US, with eligible US-issued cards.
- ▸Next: OpenAI plans to eventually allow purchases within preset spending limits and other restrictions, without approval for each transaction. The report describes this as a plan, not a shipped feature.
- ▸Plumbing: OpenAI is working with payment processors and fraud prevention firms so agent purchases can run through existing checkout systems.
The report does not say the wallet touches crypto or stablecoins. We are not going to pretend it does.
The pattern: a human approves, the network issues a scoped credential
Strip the branding and you get a familiar shape. A person approves a specific purchase. A credential is minted for that purchase. The agent never holds the long-lived secret.
That is a good design instinct, and builders in the onchain world will recognize it. If an agent can spend, the question is never "can it pay?" It is "what exactly was it allowed to do, who said so, and can that be checked afterwards?"
Today the answer is a biometric tap per purchase. The planned preauthorized mode is where it gets harder. Preset limits and restrictions are only as good as the fields that express them: amount, merchant, time window, category. How the wallet defines those, and what a merchant or processor can see about them, will matter more than the launch announcement. None of that is in the report yet, so treat it as the open question.
Where the onchain stack fits
The report also places the launch among other agent-payment efforts:
- ▸Google introduced its Agent Payments Protocol in 2025.
- ▸Visa and Mastercard have launched their own initiatives for AI-initiated transactions.
- ▸Coinbase introduced x402 in 2025, for autonomous stablecoin payments for digital services.
- ▸Circle launched its Agent Stack in May, which lets agents hold and spend USDC within predefined limits.
- ▸Crossmint combines stablecoin wallets with traditional card networks.
Same destination, different rails. Card networks bring merchant acceptance, tokenization and existing dispute processes. Stablecoin rails bring programmatic spend limits and payments that settle without a human in the loop, which suits API calls and machine-to-machine services. The report cites Juniper Research's April estimate of agentic commerce value reaching $1.5 trillion by 2030, up from about $8 billion in 2026. That is a forecast, and a forecast is a claim, not a receipt.
Our read is that this is not a cards-versus-crypto contest. Different jobs will pick different rails. An agent buying a physical product from a mainstream retailer has a reason to use card rails. An agent paying per call for data has a reason to use x402-style rails. Builders will mix them.
What this means if you build agents
Three practical points:
1. Pick the rail per job. Match the payment rail to what the agent buys, not to the loudest announcement. 2. Design the permission first. Spend limits, allowed counterparties and an approval path belong in the spec before the prompt. The prompt is the cute part; the stack is where the work lives. 3. Check claims against evidence. "Agent payments" covers a wide range, from a human tapping approve on every purchase to a wallet spending alone within limits. Read which one a product actually does.
This is what the Sato builder is for: describe the agent, and Sato maps the stack, including the wallets and payment rails it draws on. Start at satohub.ai/build.
What to watch
- ▸The preauthorized mode. When OpenAI ships spend limits without per-purchase approval, look at what the limits can express and who can inspect them.
- ▸Reach. The launch is Pro-only, US-only, US-issued cards only. Watch whether and when that widens.
- ▸Any crypto bridge. Nothing in the report links this wallet to stablecoins. If that changes, it will be worth its own post.
- ▸Dispute handling. Card networks have mature processes for contested charges. How those apply when an agent made the purchase is not covered in the report.
Claimed so far: agents can buy through ChatGPT with tokenized card credentials and biometric approval. Shown so far: an announcement. The usage data comes later.