US Pro users get ChatGPT Wallet for agent payments

OpenAI on November 13 began rolling out ChatGPT Wallet, which lets its dots agents pay at ordinary online checkouts with a one-time network token in place of the user’s card, so a store gets paid through the checkout it already runs, with no Wallet-specific integration. What can still stop the sale is the store’s own bot and fraud filtering, and the playbook OpenAI published alongside it names the fixes: signed agent traffic that four CDN and security providers already verify, and buyer signals passed along in a Forwarded header.
Per OpenAI’s announcement, Wallet starts in dots, the company’s always-on agents, for ChatGPT Pro subscribers aged 18 and older in the US with eligible US-issued cards. The user adds a card once. Before each purchase they review the items, the total and the card, then approve with a security check such as Face ID or a fingerprint scan, and each approval leaves a record of what the agent was authorized to buy. The payment network then issues the token, with its own card number, security code and expiration date, so neither the agent nor the business sees the real card. American Express, Mastercard and Visa are the partners, and card rewards and protections for eligible disputes stay in place.
OpenAI says it expects “over time” to let users set spending limits, purchase categories or time periods and pre-approve purchases inside them, with biometric checks and records of authorization still attached.
What a store needs to check
The first item on OpenAI’s list is product data: claim your business in ChatGPT and, once verified, upload a feed with current pricing and availability. That profile is the one in the Businesses in ChatGPT beta, where eligible US stores can also publish their own agent built from the same feeds. Paying through Wallet needs neither: browser access does not require a plugin or a product feed.
A dot shops through ChatGPT’s cloud browser, and those visits are signed with Web Bot Auth, a scheme that puts a cryptographic signature on each HTTP request so a site can confirm who sent it. Akamai, Cloudflare, HUMAN and Vercel verify ChatGPT’s signatures automatically. Cloudflare shows the traffic as “dots, OpenAI” under Personal Agents, with the bot tag chatgpt-agent; Akamai lists it as ChatGPT Agent in its AI Fetchers and Agents category; Vercel identifies it as chatgpt-operator.
On any other CDN or firewall the check is manual. The Signature-Agent header must exactly match "https://chatgpt.com", quotation marks included, the public key comes from https://chatgpt.com/.well-known/http-message-signatures-directory, and verification follows RFC 9421. That covers the cloud browser only; OpenAI’s crawlers, including OAI-SearchBot, are separate.
The playbook says risk systems tuned to catch bots lean on IP addresses and browser details, and the cloud browser can trip them by accident and make purchases fail. OpenAI is rolling out a relay of the user’s own signals in the Forwarded header, with seven parameters: asn, for, ip-country, ip-prefix, locale, tz and user-agent. Stores on a third-party risk provider are told to ask it to ingest them; stores with in-house systems should start watching the header, and OpenAI warns that the signals will vary while support reaches every site. On the provider side, OpenAI is working with Forter, Riskified, Sardine and Signifyd and with payment processors Adyen, Checkout.com and Stripe so their systems recognize legitimate agent purchases.
The token has its own rules. A store receives a network token and links it to the underlying account through the Payment Account Reference (PAR). It is good for the one approved purchase and must not be saved or reused, though subscriptions work if the existing checkout supports recurring billing. Orders, fulfillment, support and fraud decisions stay with the business, and OpenAI says support should treat the customer as if they had acted themselves.
Plugins and Wallet don’t work together yet. OpenAI says it is “still working on the interaction patterns” between them, so a plugin’s own checkout should assume the customer signs in or, as a guest, types everything from scratch. Plugins for dots have to be headless, since iframe-based MCP UI isn’t supported there, and Sign in with ChatGPT is open only to selected partners on request. Site owners also can’t set which actions users take on their site; OpenAI requires human review before any consequential action across all users.
Our analysis
What we checked
On November 14 we read the full Commerce Partner Playbook through its single-file Markdown export, fetched ChatGPT’s signing-key directory twice about 90 seconds apart and verified it with our own Node script. We also compared it with Cloudflare’s public Radar bots directory.
The announcement tells businesses to configure their CDN or firewall to recognize signed requests, while the playbook opens more softly: ChatGPT’s cloud browser can already use a site and needs no allowlist, so only stores whose security blocks those visits have work to do. The key directory behind that signing answered from chatgpt.com, its self-signature verified against the one key it holds, and Cloudflare’s directory points to the same key. The relayed buyer signals meant to cut fraud declines we know only from the documentation.
- The same directory path on openai.com returned 404. The single key is Ed25519, its ID equals its RFC 7638 thumbprint, and the response body matches its digest header. We checked the directory and its self-signature, not a signed request from a real dot, which we could not trigger.
- In both fetches the key’s expiry was exactly 7 days after the response’s signature timestamp, and that signature was valid for 86,400 seconds. Our inference is that the expiry is a rolling window recomputed on each request, not a fixed rotation date.
- Cloudflare Radar lists “ChatGPT agent” as a Web Bot Auth agent pointing to the same directory with one key. We did not check Akamai, HUMAN or Vercel; their support is documented by OpenAI, not measured.
Of the seven Forwarded parameters, only for is in IANA’s Forwarded parameters registry, which lists four: by, for, host and proto. The relay itself is documented, not measured: we did not see the header on a live request.
Up next
Dario Amodei and Jamie Dimon hold a live chat at 12 PM ET
At 12 PM ET, JPMorgan CEO Jamie Dimon sits down with Dario Amodei for a live chat, and the stream ships with an SEC disclaimer because Anthropic’s IPO filing isn’t effective yet. Before buying into an IPO, do you trust a CEO chat or the filing more?
Related stories
- OpenAI’s dots work 24/7 on their own cloud computer
- ChatGPT’s code points to a wallet for agent purchases
- Gemini reads your Google Wallet spending, US users only
- ChatGPT users in the EU report getting OpenAI’s dots
- GPT-6 Astra Pro splits one task among parallel agents
- ChatGPT Auto-review goes free for ChatGPT sign-ins
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

