OpenAI introduces ChatGPT Wallet for agent purchases
OpenAI's ChatGPT Wallet lets agents in ChatGPT Work, Codex and dots buy things with user approval and existing card protections. What changes at checkout.
Source: OpenAI NewsOpenAI has added a payment layer for its agents: ChatGPT Wallet, announced on 13 November 2026, is meant to let agents inside ChatGPT Work, Codex and dots finish a purchase rather than merely recommend one. According to OpenAI News, the user signs off on the purchase and the card protections the buyer already holds stay in force, which makes this as much a liability design as a checkout feature.
What is ChatGPT Wallet?
OpenAI describes ChatGPT Wallet as a secure way for agents to complete purchases. Three things are bundled into that description: the agents run in ChatGPT Work, Codex and dots; a person approves the purchase; and existing card protections are not replaced. The distinction worth holding onto is between an agent that drafts and an agent that transacts. A drafting agent's worst failure is an unhelpful answer. A transacting agent's worst failure is a real charge, to a real merchant, on a real card, with a refund process attached. OpenAI's announcement frames the wallet as the component that carries those obligations.
Which OpenAI products does it cover?
The announcement names three surfaces: ChatGPT Work, Codex and dots. They are not interchangeable spending contexts. A coding agent tends to buy developer resources: credits, compute, packages, services. An assistant surface tends to buy the things a person would buy: a subscription, a seat, a booking, a physical item. One wallet spanning both has to answer for very different baskets, and very different mistakes. The original post is the place to check what each surface is expected to purchase and how the approval step appears in each; the specifics of dots in particular are worth confirming at the source rather than assuming.
Does the user approve every purchase?
Approval is stated as a requirement, not a setting: according to OpenAI, agents complete purchases with user approval. What use will reveal is the granularity of that approval, whether it is asked per transaction, per merchant, per amount, or granted once as a standing permission and then reused. That distinction decides how much supervision the wallet really provides. A prompt before every charge is a confirmation step. A permission granted once is a mandate, and the engineering then moves into the limits around it: per-item ceilings, merchant allowlists, and hard stops when a basket crosses a line. Grafy has documented the same pattern for autonomous agents generally, an agent that runs freely but stops at a budget (read the walkthrough), and the design question transfers directly to payments.
Do existing card protections still apply?
Yes, according to OpenAI: existing card protections apply. Read plainly, the wallet is not presented as a new balance, a store of credit or an agent-specific line of finance; the purchase still runs through the payment credentials and the rules the user already has. For buyers, the practical consequence is continuity rather than novelty, since whatever dispute, fraud and issuer processes already govern that card continue to govern an order placed by an agent. For anyone building agent products, that is the more consequential half of the announcement. A checkout is straightforward to build; inheriting an existing protection regime instead of inventing one is what makes an agent's purchases survivable when they go wrong.
Why does agent checkout matter beyond OpenAI?
Because authorisation, not intelligence, has been the bottleneck. Agents have been able to search, compare and fill in forms for some time; what they have lacked is a legitimate way to be trusted with money. The friction runs both ways. Some merchants want no part of it, and Grafy News has reported that United will not let AI agents book flights on its site. Some products would rather flag uncertainty than transact on it, which is the approach in Grafy's write-up of an AI trip planner that says when a fare is a guess. A wallet with approval and inherited card protections is an attempt to move the argument from whether agents should buy things to under what controls they may, which is the more useful question and the one merchants can actually answer.
How can you explore this in Grafy?
Grafy has no wallet and no payment rail, and nothing in this announcement changes that. What it does have is the scaffolding around a spending decision. Operator is a browsing agent that carries out web tasks, which is the research half of any agentic purchase: gathering options, checking terms and collecting the pages a human then approves. Graph keeps the branches, sources and model history of that research in one place, so a decision can be retraced later. If you want to model the control side before connecting anything, copy the budget-ceiling pattern above: define the ceiling, the allowlist and the stop condition first, then let the agent run. And to read announcements like this one closely, save the page as a PDF with Files and question it in Reader.
Frequently asked questions
What is ChatGPT Wallet?
ChatGPT Wallet is a payment feature OpenAI announced on 13 November 2026. According to OpenAI News, it gives agents a secure way to complete purchases, so an agent can finish a transaction instead of only suggesting one. Two conditions are attached: the user approves the purchase, and the card protections the buyer already has continue to apply.
Does ChatGPT Wallet require user approval before a purchase?
Yes. OpenAI states that agents complete purchases with user approval, so approval is part of the mechanism rather than an optional extra. How often that approval is requested, whether per purchase, per merchant or once as a standing permission, is the detail to confirm in the original post before relying on it for anything expensive or hard to reverse.
Which OpenAI products can use ChatGPT Wallet?
Three are named in the announcement: ChatGPT Work, Codex and dots. The wallet is therefore not limited to a single chat window; it is described as shared across OpenAI's assistant work surface, its coding agent and dots. Check the original post for what each surface is expected to buy and how the approval step appears in each one.
Source: OpenAI News
More news
OpenAI board member warns AI takeover could arrive within months
Paul Cristiano, an OpenAI board member, said an AI accident may become the leading cause of death and warned a takeover could happen in the next six to twelve.
OpenAI dissolves its dedicated robotics team to fold researchers into general AI groups
OpenAI has disbanded its stand‑alone robotics unit, moving staff to core research and applied groups as its GPT‑6 Astra model begins controlling robots.
LangChain raises $300M Series C, pushes valuation to $3 billion
LangChain secured a $300 million Series C led by CapitalG, doubling its valuation to $3 billion and underscoring the growing demand for model‑agnostic AI agent.