← Dailies

Daily · 2026-11-14 00:00 UTC

ChatGPT Wallet pays for you; JPMorgan blocks the agents

OpenAI launches ChatGPT Wallet while JPMorgan blocks agent payments, LangChain hits a $3B valuation, Anthropic puts Claude inside factory software, and researchers find export-banned Nvidia chips behind China's best open models.

Source ↗

Ten stories from the last 24 hours. The agent economy got its checkout lane and its first roadblock on the same day: OpenAI shipped a wallet, and the biggest US bank said no.

1. OpenAI launches ChatGPT Wallet so its agents can buy things — OpenAI's second attempt at AI shopping, after scrapping Instant Checkout as too rigid. Users add a card once and approve each purchase with Face ID or a fingerprint; American Express, Mastercard and Visa issue a payment token, so neither the agent nor the merchant sees the real number. It is US-only at launch, requires an adult account with a Pro subscription and an eligible US bank card, and runs through Dots. OpenAI says it plans to eventually let users hand off purchases entirely, within set spending limits, categories and timeframes, and published a Commerce Partner Playbook for merchants working with Stripe, Adyen and Riskified. Anthropic's Claude Duet, Meta's Muse and xAI's Grok Bot already pay for things. The Decoder · Nov 13
2. JPMorgan blocks AI agents from using customers' credit cards — The bank is refusing agent-initiated purchases made with a customer's card number, the Wall Street Journal reported Friday, while cards from Bank of America, Capital One and Citigroup are already being used by agents with their owners' permission. Stripe routes around the block with Link, which gives the agent a single-use virtual card instead of the customer's own; merchants told the Journal that card is classed as commercial and costs more to accept. PYMNTS Intelligence found 52% of consumers want to choose how they pay when an agent reaches checkout on a $300 purchase. PYMNTS · Nov 13
3. LangChain raises $300M at a $3B valuation — The agent-management startup's Series C was led by CapitalG, with Singapore's GIC, DFJ Growth, IVP and Sapphire Ventures participating — more than double the $1.25B valuation from its $125M round about a year earlier. LangChain began in 2022 as an open-source project for tasks early models struggled with, then pivoted to managing agents; it now claims 7,000 LangSmith customers, including Harvey, Rippling, Clay, Nvidia, Toyota and Salesforce. The pitch is model- and cloud-agnostic agent ownership, against the integrated stacks from the big clouds and labs. TechCrunch · Nov 13
4. Anthropic puts Claude inside factory software — Claude for Manufacturing is a plugin with 17 skills, connected through four universal interfaces to Autodesk Fusion, Oracle NetSuite, Microsoft 365 and Infor, with other systems reachable through file exports. It drafts quotes from actual inventory, compares suppliers, generates quality reports and design variants, and reads sensor data and manuals during machine downtime; an Infor plugin lets it order from backup suppliers and reschedule production after human approval. Output is linked back to source documents for review. It is available to Team and Enterprise customers but does not control physical machines — Anthropic is developing a Model Hardware Standard as a research preview for that. The Decoder · Nov 13
5. Export-banned Nvidia chips appear to sit behind China's leading open models — Will DePue, a former OpenAI researcher who leads the nonprofit Unemployment Labs, analyzed the published weights of models from DeepSeek, Moonshot AI, MiniMax and Ant Group and found fingerprints of training on Blackwell GPUs, including in Moonshot's Kimi K3 releases and DeepSeek's V4 models. That does not establish a violation: the chips could have been in an offshore data center, a route Epoch AI estimates brought 275,000 Nvidia accelerators into Malaysia in 2024 and 2025, out of roughly two million advanced chips that entered China through routing schemes. TechCrunch · Nov 13
6. Amazon reportedly plans a catalogue of every product on the internet — Reuters reported an internal project called World Product Graph: a vault of data on products sold anywhere online, including items Amazon does not sell, aimed at keeping shoppers from leaving for Reddit or Google — and at becoming the index AI agents lean on when they shop. The internal document reportedly says "Many Rufus answers are factually wrong or incomplete" and that running searches through the assistant is expensive, and carries an October 1, 2027 release date. Amazon told Reuters that draft, outdated documents often paint an incomplete picture and that a brainstorming document does not necessarily reflect current plans. The Verge · Nov 13
7. Mirendil raises $1B to automate model training — The startup, founded last year by former Google DeepMind and Anthropic researchers, raised a Series A jointly led by Kleiner Perkins and Clearlake Capital, with Google among more than half a dozen other investors. Its platform, Lor, splits training-run work across hundreds of AI agents that build variants, evaluate them and fix errors. Mirendil says one researcher used it to build a coding model for $10 million in about three and a half months, with agents writing more than 99% of the code; its Fern coding model, unveiled last month, claims a 10-million-token input window per prompt, ten times Claude Opus 5.5. SiliconANGLE · Nov 13
8. Onyx Security brings agent governance to Glean — The integration inventories every agent a company builds on Glean, records what access and capabilities each one holds, and assigns ownership, business purpose and a risk classification. Policies then apply across the estate, with continuous checks on permission, tool and data-access changes, intent-aware runtime controls that can block or steer an unsafe agent action, and centralized audit evidence. Onyx has raised $153 million to date, including a $113 million Series B led by Bessemer in July; Glean last raised $150 million at a $7.2 billion valuation in June 2025. SiliconANGLE · Nov 13
9. AI cracks open Kurt Gödel's shorthand notebooks — Gödel left 16 workbooks, 1,009 densely written pages on maths and logic, much of it in Gabelsberger shorthand that was obsolete when he used it. A Helsinki project led by logician Jan von Plato had worked through the material at about a page a week since 2017; in September he brought in Claude, which produced a comprehensive digest of the whole corpus within a couple of days. The material is said to contain hundreds of proved theorems nobody knew about, dozens of anticipations of results only established in the 1970s and 1980s, and 14 themes in logic that were never pursued. "I was flabbergasted," von Plato said. The Guardian · Nov 13
10. Singapore's healthcare breaches point at a wider exposure — Ten cybersecurity incidents involving compromised information were reported in the past two months, six of them involving healthcare-related organisations. The Cyber Security Agency of Singapore said Friday the string of incidents reflects "a broader global trend of opportunistic attacks" rather than coordinated campaigns targeting Singapore, while experts pointed to an increasingly interconnected healthcare ecosystem holding valuable personal data and dependent on uninterrupted digital service. CNA · Nov 13

Try This Next

• 🔧 Map your own agent's payment path before you build checkout: token, virtual card or raw card number — the JPMorgan block shows the difference decides whether the purchase completes.
• 🎥 Read the Gödel notebooks coverage and note the ratio: a page a week for seven years, versus a two-day digest. That is what domain-specific reading help actually buys.
• 📝 Save the Amazon World Product Graph detail for later — if agents end up shopping on a single index, product data becomes distribution.

Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.