News · · 7 min read

ChatGPT Wallet launches: what merchants should check even if their customers can't use it yet

In this post
  1. Who can use it
  2. What the merchant receives, and what it doesn't
  3. Bot blocking and signature verification
  4. What the key directory actually returns
  5. Community reaction
  6. What it changes for merchants outside the US
  7. Limitations

OpenAI launched ChatGPT Wallet on November 13, 2026. It lets a dot, ChatGPT's agent, pick products and pay for them, with the user approving each purchase through Face ID or a fingerprint. According to launch-day reporting, it is open to a limited number of US ChatGPT Pro subscribers aged 18 and over, using a US-issued Visa, Mastercard, or American Express card. Users outside the US, including in Korea, cannot use it yet.

The more immediate change is on the merchant side. OpenAI's developer documentation says Wallet works with existing checkout pages and needs no Wallet-specific integration. Put the other way around, an agent can already reach your checkout page whether or not you have done anything. This post covers who can use Wallet, what a merchant receives and the new header that comes with it, CDN bot blocking and signature verification, and what OpenAI's public signing key directory returned when we requested it ourselves.

Who can use it

OpenAI's announcement on openai.com returned HTTP 403 from the environment this post was written in. Eligibility therefore comes from two launch-day reports, and the merchant-side details come from OpenAI's developer documentation, which did load.

ItemDetailSource
Eligible usersA limited number of US ChatGPT Pro subscribers, 18 and overTechCrunch, CNBC
CardsUS-issued Visa, Mastercard, American ExpressTechCrunch
ApprovalReview items, total, and card, then approve each purchase with Face ID or a fingerprintTechCrunch
Card benefitsCard rewards and protections such as disputing charges still applyTechCrunch
FeesOpenAI is not taking a cut of transactions to startCNBC
PlannedSpending limits and pre-approval by purchase category or time periodTechCrunch

None of these sources gives a timeline for other countries, and merchant readiness is still a work in progress. Talking about merchant sites in the CNBC piece, Chaya Nayak, who leads product for monetization and payments at OpenAI, said "It's going to take time for the ecosystem to build the right infrastructure to be able to make this experience really amazing and seamless across all of their sites."

What the merchant receives, and what it doesn't

OpenAI's Accept payments guide describes the merchant side. The split of responsibilities is the core of it: the customer reviews and approves the purchase inside ChatGPT, and the merchant processes the payment and handles the order, fulfillment, customer support, and fraud decisions.

The merchant never sees the full card number. It receives a network token instead, and the guide says to use the Payment Account Reference (PAR) to link that token to the underlying account. The token is scoped to the one transaction the customer approved; the guide says "Don't reuse or save that payment credential for a separate purchase." A checkout that saves payment methods by default needs to skip saving this one. Subscriptions are allowed if your checkout already supports recurring billing and you collect the customer's consent.

What is new is the Forwarded HTTP header. ChatGPT's cloud browser connects from cloud infrastructure, so fraud systems that rely on IP addresses and browser signals can block legitimate purchases. OpenAI has started relaying the real user's signals in this header. The guide lists these parameters:

ParameterMeaningExample from the guide
asnAS number on the route to the user's IP64496
forThe user's IP address198.51.100.23
ip-countryCountry code estimated from the IPUS
ip-prefixMasked network prefix (CIDR)192.0.2.0/24
localeThe user's configured localeen-US
tzThe user's time zoneAmerica/Los_Angeles
user-agentThe user's browser or appExampleApp (iOS)

The guide warns that the signals "may vary as we roll out support for every site," and tells merchants using a third-party risk provider to confirm that provider ingests them.

Bot blocking and signature verification

An agent has to get into the site before it can reach checkout. OpenAI's Agent Access page says you don't need to allowlist ChatGPT's cloud browser, but your security settings may block it unintentionally. We compiled how each provider labels this traffic from that page.

CDN / security vendorHow the traffic appearsVerifies signatures automatically
AkamaiChatGPT Agent (AI Fetchers and Agents)Yes
Cloudflaredots, OpenAI (Personal Agents) / bot tag chatgpt-agentYes
HUMANChatGPT AgentYes
VercelChatGPT Agent (chatgpt-operator)Yes

Anyone else has to verify requests themselves. ChatGPT uses Web Bot Auth to identify certain automated browser requests; which uses the HTTP Message Signatures standard (RFC 9421). A signed request carries Signature and Signature-Input headers plus a Signature-Agent header set to "https://chatgpt.com". Public keys live at https://chatgpt.com/.well-known/http-message-signatures-directory. These signatures are separate from OpenAI's crawlers such as OAI-SearchBot; we covered crawler-side controls in our post on the shared robots.txt AI-training token.

What the key directory actually returns

The documentation doesn't say how long a key can be cached, so we requested the directory twice, at 08:51 and 08:52 UTC on November 16, 2026. The response looked like this (compiled by us):

{"keys":[{"crv":"Ed25519","kty":"OKP",
  "x":"P_1c4SX6fCavDcq6FRegWcVvaOKsp_aXHPEN1FcKDIA",
  "kid":"ArwjqcCqtA5oxEWkzXGC-To_0A9whHmajCHtfe8I9aY",
  "use":"sig","nbf":1735689600,"exp":1795423860}],
 "signature_agent":"https://chatgpt.com","purpose":"ai"}
FieldFirst request (08:51:00 UTC)Second request (08:52:11 UTC)
Keys / algorithm1, Ed255191, Ed25519
kidSame in bothSame in both
nbf1735689600 (2025-01-01 00:00 UTC)Same
exp1795423860 (2026-11-23 08:51:00 UTC)1795423931 (2026-11-23 08:52:11 UTC)
Cache-Controlno-store, no-transformno-store, no-transform

In both responses, exp was exactly 604,800 seconds (seven days) after the response's Date header. The key itself stayed the same; only its expiry moved forward with each request. We also computed the key's JWK thumbprint (RFC 7638, SHA-256) and it matched the kid character for character. The no-store header means intermediate caches won't keep the response.

Two observations aren't enough to call this a rule. If it holds, though, verifiers would be safer re-fetching the directory before exp rather than hard-coding the key. If the key rotates, a pinned key would fail every signature check and block legitimate agent purchases.

Community reaction

Three days after launch, we found little developer discussion. On Hacker News, a submission of the Reuters story drew a single comment, a question: "Will it be a time where we can actually trust models with our wallets?" (ahmedketata, Hacker News).

Slashdot commenters landed on both sides. On the positive side, Iamthecheese wrote "if it were safe I would prefer the frictionless ai concierge," and lukewarmfusion said they already give agents "credit cards that I generate at privacy.com and give budget caps to." On the other side, Whateverthisis asked "Does anyone want a bot spending your money for you?" (Slashdot).

In the press, Gizmodo's AJ Dellinger noted that "Wallet will collect detailed records of purchases, including merchants, items, prices, and shipping information," and that "OpenAI will still be collecting plenty of information about your spending habits" (Gizmodo). From the merchant angle, Rajesh Beri pointed out that OpenAI's payments guide doesn't address chargebacks and advised merchants to agree with their acquirer on what evidence wins a dispute over an agent-placed order (beri.net).

What it changes for merchants outside the US

If your customers are all outside the US, very little changes for now. Wallet requires a US card, and no wider rollout has been dated.

Stores that sell to US customers are in a different position. An English-language storefront that accepts foreign cards can already receive a US Pro user's dot at checkout. Three checks look worthwhile: whether your CDN or firewall is blocking ChatGPT's cloud browser; whether your checkout auto-saves the network token for later reuse; and whether your fraud rules reject purchases from cloud IPs outright, and if so whether they can read the Forwarded header.

You can't pick and choose what the agent may do. Agent Access says OpenAI doesn't "yet have features to allow specific site owners to control what actions users can take," and that an exception applies site-wide, so you should keep your existing page-level access controls in place. OpenAI says it instead requires human review before any consequential action for all users. For now, the practical decisions for a merchant appear to be whether to allow agent access, and if so, how much the existing page-level controls should leave open.

Limitations

ItemDetail
Primary announcementThe openai.com post returned HTTP 403, so eligibility relies on TechCrunch and CNBC
PerformanceThis is a payments feature with no public benchmarks; OpenAI hasn't published success rates or timings
Our own checkTwo responses from the key directory only; we did not receive or verify a real signed request (that needs a US Pro account)
CommunityThree days in, the only Hacker News discussion we found has one comment, and we could not confirm any Reddit thread; quotes come mostly from a few Slashdot comments and the press
Local payment gatewaysWe did not test whether a non-US payment gateway's foreign-card checkout completes in the cloud browser
1 / 1