ChatGPT Wallet Requires Purchase Approval
7 stories · ~7 min read

Listen
If You Only Read One Thing
Buying the wrong thing can be perfectly authorized. ChatGPT Wallet keeps approval tied to the purchase. Meanwhile, Every's Hydrogen editing test shows how easily fluent execution can hide a missed decision: the model polishes a claim without checking whether the document supports it. Both developments make the size of the delegated decision more consequential than the smoothness of the answer.
ChatGPT Wallet Keeps Purchase Approval With the Customer
ChatGPT Wallet moves agents from finding products to completing checkout, while keeping approval tied to the purchase. That changes what an agent can finish. It doesn't make a broad shopping request sufficient permission to spend.
OpenAI's merchant payment documentation explains the division: the customer approves the purchase in ChatGPT, and the merchant processes payment using a network token instead of the full card number. Consent covers that transaction. The credential can't be reused for a different purchase.
That separation matters because shopping contains two different problems. An agent has to select a suitable item, then obtain authority to charge for it. Hiding card details reduces exposure during the second step. It can't establish that the first step found the right size, delivery date or cancellation terms. A valid payment can still buy the wrong thing.
OpenAI's commerce playbook gives businesses two execution paths. Browser use follows the existing website; a plugin exposes structured actions that the business validates through its backend. Wallet works with existing checkout forms, so a merchant doesn't need a Wallet-specific integration just to accept it. That lowers the adoption barrier while leaving existing checkout failures in the path.
Fraud detection is one such failure. OpenAI says cloud browsers can trigger bot defenses, and is rolling out forwarded user signals to help merchants assess visits. Merchants still own fraud decisions and fulfillment. Reaching the payment form doesn't establish that an order will be accepted.
There's a reasonable objection: ordinary online shopping already combines card tokens, risk checks and human confirmation. The advance is putting those pieces around a persistent agent's work. The agent can investigate and prepare the transaction while the customer retains the final decision. Reliability now depends on preserving the selected purchase across that handoff.
A merchant acceptance test can make the distinction concrete: prepare an order, change its amount or recipient, and check whether the eventual charge still matches what the customer approved. That's a proposed test, not a reported Wallet result. Successful checkout with the approved merchant and amount preserved is the evidence that would make this delegation dependable.
Hydrogen Misses Problems Across a Whole Draft
Gemini 4 Hydrogen looks more useful as a paragraph editor than as the final reviewer of a long document. Every's editorial experiment found a gap that yesterday's automation comparison couldn't reveal: the model can improve a sentence while preserving the mistake inside it.
Yesterday's briefing examined Hydrogen's automation accuracy and estimated cost. Every asks a different question. Its Frontier team ran Hydrogen twice across 12 editorial tasks, including opening revisions, interview-based transitions and critique of a long feature. Responses were graded with task-specific rubrics alongside other frontier models. Hydrogen averaged 51–57% across its two attempts; Carbon scored 69%. Those are editorial rubric scores, not percentages of articles ready to publish.
One example explains the gap better than the averages. An opening described somebody as Every's biggest AI user. A caption farther down identified the underlying measurement as an OpenAI token leaderboard, although other people used Claude. Carbon and the tested OpenAI and Anthropic models caught the overclaim. Hydrogen polished the opening and retained it in both attempts.
The error required no obscure outside knowledge. The qualification was already in the document. Hydrogen had to connect the sentence being edited to the evidence elsewhere, then decide that the sentence needed a narrower claim. Better phrasing alone couldn't repair it.
That suggests a different way to divide editorial work. A short rewrite with its facts supplied is a bounded transformation. Checking an opening against a whole draft requires preserving relationships between claims, evidence and attribution. Sending both jobs to the same model because both look like writing can conceal a capability gap.
The experiment is small and reflects one publisher's standards. It doesn't establish a universal model ranking, and Every reports that Hydrogen's stronger short-copy attempt could match or beat Carbon. The result supports a task boundary more convincingly than a blanket rejection of the cheaper model.
For repeated work, the transferable practice is to retain examples where the correction depends on information outside the selected paragraph. A model comparison then measures whether the editor preserves the document's meaning, alongside whether its sentences sound better. Those examples add something a style guide can't supply: an observable failure the guide is supposed to prevent.
Hydrogen becomes a credible whole-draft reviewer when it consistently catches these cross-document contradictions on unseen examples. A smoother rewrite of the same erroneous claim wouldn't clear that bar.
The Contrarian Take
Everyone says: Business Today describes Wallet as “autonomous shopping”, and says agents can buy once given instructions.
Here's why that's wrong: Wallet's documented checkout still requires the customer to approve the specific purchase. OpenAI has automated more of the preparation and payment handling, while retaining a human decision at the point of commitment. That distinction changes the product's economics: an agent can save browsing time without removing the interruption required to buy. The relevant comparison is how much work happens before that interruption, and whether the approval presents an accurate purchase. Card secrecy alone can't answer either question.
Under the Radar
-
A model can pass ordinary tests and still carry an executable trigger. Microsoft's November 10 Lazarus disclosure describes backdoors made by changing fewer than ten dormant neurons without retraining. The lab demonstration requires altered weights; it isn't evidence of a deployed compromise. Comparing a checkpoint with its trusted original addresses a different risk from testing its everyday answers. Microsoft says dependable detection needs that original reference.
-
Claude Code repairs where commands run after a directory disappears. November 13's v2.1.325 stops shell commands from falling back to the launch directory after the working directory is deleted. It also repairs missing recent turns after an unclean shutdown. Long-running agents need both properties: a resumed transcript must preserve the work, and the next command must still target the intended place.
Quick Takes
-
IBM trains small models to distinguish instructions from retrieved material. November 12's LLMON previews ship Granite 4.1 models and tooling that bind instructions to named input blocks. IBM reports strong resistance to static indirect prompt injection, where malicious orders arrive inside documents or tool results. The authors explicitly describe learned behavior, without an enforced guarantee, and label the release unsuitable for production. Static tests can’t tell us how well the defense handles an attacker who changes tactics. (Source)
-
Deepnote tests whether agents know when to stay quiet. Its November 13 benchmark puts 12 models through 50 workplace situations with identical tools, instructions and time budgets. GPT-6.1 Sol scored 8.65/10 at $0.112 per task; Astra scored 8.97 at $0.693. Their uncertainty intervals overlap. Both stayed silent in eight of ten cases requiring silence. The runnable simulated workplace tests interruption judgment alongside answer quality; it doesn't measure benefits to real teams. (Source)
-
Copilot's December retirement crosses every interface. GitHub will retire six models across Copilot experiences on December 14, including chat, agent modes and completions. Haiku 4.5 remains available to individual annual subscribers; Kimi K3's replacement is still unannounced. Business and Enterprise access to alternatives depends on model-enablement policy. Migrating one chat selection won't establish that every integration has an allowed replacement. (Source)
The Thread
The shape of a task is becoming part of model selection. “Shop for this” contains research, preference judgments and a payment commitment. “Edit this” contains sentence repair, factual checking and decisions about the whole argument. A single chat request hides those different jobs. A fast model can polish a supplied paragraph while a stronger model checks the full argument. An agent can prepare a purchase while a person decides whether to commit. It also changes what a product needs to expose. A writing interface needs to preserve which facts an edit relied on; a shopping interface needs to preserve which transaction a person approved. Faster answers won't reveal either relationship by themselves. The next useful interface may show the decision being handed over as clearly as it shows the finished output.
Predictions
- I predict: By December 14, ChatGPT Wallet will still require a customer approval step for each purchase, rather than offering generally available unattended purchasing under a standing spending allowance. OpenAI's documented transaction-specific consent makes that a more substantial change than adding another checkout integration. (Confidence: medium; Check by: 2026-12-14)
2026-11-14 · 03:42 AM ET
Tomorrow morning in your inbox.
Subscribe for free. 10-minute read, every weekday.