It's Monday, November 16th: OpenAI began rolling out ChatGPT Wallet so U.S. Pro subscribers can approve a dot's purchase with a face scan, and Anthropic published a guide arguing that a safety claim is only as sound as the evaluation behind it.
🌟 Community Spotlight: Microsoft AI

What becomes scarce when intelligence is abundant? Mustafa Suleyman (CEO, Microsoft AI), Sarah Guo (founder, Conviction; co-host of No Priors) and Aaron Bastani (cofounder, Novara Media) will be in conversation, moderated by Stephen Morris, the Financial Times' San Francisco Bureau Chief.
The evening, hosted in part by The AI Collective, is Wednesday, November 18, from 8:00 to 9:00 PM PT at The Commonwealth Club, 110 The Embarcadero, San Francisco.
Get in the room with Mustafa Suleyman. The CEO of Microsoft AI sits down for an intimate, ticketed conversation about what becomes scarce when intelligence is everywhere. The night also launches the second issue of The Humanist Review of AI, Microsoft AI's quarterly journal, and every in-person guest leaves with a print copy. If you're in SF, this is one to grab a seat for.

Covering what’s happening on the ground in AI, every Monday.
1️⃣ AGENT CHECKOUT: ChatGPT Wallet Lets Dots Pay After A Face Scan

Image from OpenAI.
OpenAI began rolling out ChatGPT Wallet to U.S. Pro subscribers 18 and older, so a dot can finish a purchase after the user reviews the cart and approves it with Face ID or a fingerprint.
ChatGPT Wallet is rolling out in dots to ChatGPT Pro subscribers aged 18 and older in the U.S. with eligible U.S.-issued cards. OpenAI built it with American Express, Mastercard and Visa, and says users keep existing card rewards and protections for eligible disputes.
After the user reviews the items, total and selected card, they approve with a security check such as Face ID or a fingerprint. The payment network then issues a payment token, so the agent and the merchant never receive the actual card number.
Luke Gebb, EVP and head of global innovation at American Express, said in OpenAI's public announcement that the launch is Amex's first Agentic Commerce Experiences integration, with eligible card members reviewing and approving each purchase.
OpenAI also published a Commerce Partner Playbook for merchants, covering product feeds and Web Bot Auth so agents are not blocked. TechCrunch and The Verge both confirmed the Face ID or fingerprint check on every purchase.
U.S. Pro users who let a dot shop should watch how often a site accepts the payment and keep the biometric check turned on for every cart. Merchants should review the Commerce Partner Playbook and their fraud rules before those checkouts start failing or sailing through unnoticed.
2️⃣ EVAL PLAYBOOK: Anthropic Shows How To Build Trustworthy Safety Tests

Image from Anthropic.
Anthropic published recommendations for building safety evaluations, arguing that a claim a model is safer is only as sound as the test behind it, and reporting a 34-point gap between raw API results and claude.ai.
Researchers including lead author Ruth Appel write that many safety scores fail to measure the intended construct, and that single-turn tests have already saturated, with current API harmless-response rates of 94% to 98% on those suites.
The research post says cyber evaluations should run in a sandbox with no internet access and real-time monitoring, and that it temporarily stopped running many evaluations that give models internet access after its own incidents.
On Claude Fable 5.1, the appropriate-response rate on multi-turn safety evals differed by as much as 34 percentage points between the raw API and claude.ai. After system-prompt changes, one suicide and self-harm eval rose from 60% on the API to 94% on claude.ai.
Anthropic recommends preregistration, sample-size floors of 100 multi-turn or 1,000 single-turn items, and reporting confidence intervals, which can exceed plus or minus 10 points. It points builders to the UK AI Security Institute's Inspect framework and Anthropic's Petri toolkit.
Teams that publish or rely on safety scores should treat a single headline number as incomplete, especially when confidence intervals are wide or the eval is already saturated. Read Anthropic's recommendations before you compare two models, and check whether the eval used the same configuration, grader and construct you care about.
📰 Other Headlines
HYDROGEN IN SEARCH: Google released Gemini 4 Hydrogen at $1.50 input and $7.50 output per million tokens and is rolling it out as a Search upgrade for users worldwide.
COPILOT ADDS HYDROGEN: GitHub made Gemini 4 Hydrogen generally available in Copilot for Pro through Enterprise users across VS Code, the CLI, and other Copilot surfaces.
STUDIO CLOUD CREDITS: Google let Pro and Ultra subscribers claim monthly Cloud credits for the Gemini API inside AI Studio without leaving the workspace.
ROBOTICS TEAM FOLDS: OpenAI disbanded its stand-alone robotics team, moving researchers under Mark Chen and Greg Brockman after finding general models can control robots.
KITESURF OPEN SOURCE: Cloudflare open-sourced Kitesurf under Apache 2.0, letting developers run its Workers-based browser engine for AI agents on their own accounts.
LLAMA.CPP 1.0: The llama.cpp project released libraries v1.0.0, committing to semantic versioning so breaking public C API changes get major-version bumps.
LANGCHAIN SERIES C: LangChain raised $300 million at a $3 billion valuation, led by CapitalG, to expand LangSmith, which it says more than 7,000 organizations already use.
XY PROBLEM BENCH: Allen AI released XYBench, an 8,000-plus-question test of whether models correct a user's mistaken approach instead of answering it literally.
OCEAN DATA CENTERS: Drogue told TechCrunch it raised $20 million to build floating data centers powered by ocean temperature differences, with a first ship-based prototype still unbuilt.
The AI Collective is built by volunteers across 180+ chapters in 40+ countries.
Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.
🧑💻 About the Editors

About Joy Dong
Joy is a news editor, writer, and entrepreneur at the intersection of AI and blockchain. Whether she is demystifying complex systems in her newsletter, TEA, or building streamlined solutions through her automation agency, Ownly, Joy’s mission is to make emerging tech accessible and actionable for everyone.

About Lindsay Gross
Lindsay is an AI engineer, researcher, and writer focused on how AI systems behave in practice and what it takes to make them safe. Her work sits at the intersection of AI safety, governance, and product design. At AIC and in her newsletter, Hidden Layer, she writes about the questions that matter most as these systems scale.
