Announcer: This is AI Squawk Weekly for Saturday, the fourteenth of November, brought to you by VarOps, the resident AI company. Muximus: Welcome to the AI Squawk Weekly. I'm Muximus. Penny: And I'm Penny. Three threads ran through the daily show this week. Muximus: AI agents trying to shop and book for people, small cheap models taking over the boring work, and a run of leaks that came down to keys and share links. Penny: We'll look at how each one moved since we covered it. And as always, what it means for you. Muximus: First up. Agents at the checkout. On Wednesday we covered OpenAI's new Business Hub, where a business can claim its listing in ChatGPT, upload its products and prices, and see how often ChatGPT sends people its way. The same day, eBay appeared to start blocking outside shopping agents, Meta's Muse among them. Amazon already blocks them completely. On Thursday, United Airlines said unauthorized AI agents can no longer book or change trips. They can still search for flights, but a person has to finish the purchase. United's chief executive, Scott Kirby, told The Wall Street Journal he doesn't want another middleman. And on Friday, Muse started placing phone calls to US businesses for paying subscribers. Muximus: Then on Friday afternoon, after our show, OpenAI launched ChatGPT Wallet. Its agents, which it calls dots, can now complete a purchase for you. You review the items and the total, and approve each one with Face ID or a fingerprint. The store gets a payment token, not your real card number. It's only for ChatGPT Pro subscribers in the US for now, with American Express, Mastercard and Visa on board. OpenAI also published a playbook for stores, and one of the first things on it is making sure your website isn't blocking its agents by accident. And late Friday, The Verge, citing Reuters, reported on an internal Amazon plan to index products sold anywhere online, so shoppers and their agents start on Amazon. Amazon called it a brainstorming document. Penny: Gritt Scott is with us for this one. Gritt writes Skill Issue, our column that filters hundreds of AI developments down to the ones that actually matter. Welcome back, Gritt. Gritt: Thanks, good to be here. Muximus: So are people actually letting agents buy things? Gritt: A few are. Uber and DoorDash have both said agent orders are tiny so far. But TechCrunch reports Muse has more than ten million users, and shopping is one of the top things they do with it. So it's early, but it isn't nothing. And the big missing piece, paying safely, now has an answer from OpenAI. At least for Pro users in the US. Penny: And then United and eBay close the door. Gritt: Right, and I don't think either side is wrong. United points to account security, purchases people didn't mean to make, and the federal rules on air travel. And Kirby just doesn't want someone standing between him and his customer. OpenAI's answer is signed requests, so a store can tell a real ChatGPT agent from a scraper, plus an approval on every purchase. Muximus: So what does a company that sells online actually do? Gritt: Decide your rule on purpose. A lot of sites block agents by accident, because a firewall setting treats every bot the same. Ask whoever runs your website what happens today when an agent shows up. Then pick one. Let them browse, let them buy, or keep the purchase with a person, the way United did. Penny: And if they let them in? Gritt: Then talk to your payments and fraud people first. OpenAI says real agent purchases can still get flagged and declined. And claim your listing in ChatGPT while you're at it. On Friday we covered a study that found risky or fake shops in some chatbot shopping answers, and smaller brands got hit about twice as often. An accurate listing is the cheapest defense you've got. Muximus: What about the buying side? Staff with company cards? Gritt: For now Wallet is a personal Pro feature, so it's mostly a policy question. I'd write one line. No agent spends company money without a person approving it. OpenAI says spending limits are coming later. I'd wait for those before handing anything more over. Muximus: And Amazon's plan to index everything for sale? Gritt: It's a plan in a document, with a date in 2027, and Amazon is already playing it down. I wouldn't change anything for it today. But it tells you where this is going. Whoever has the best product data gets the agents. So keep yours accurate wherever you list it. Penny: Thank you, Gritt. Moving on. Muximus: Next. Small, cheap models are taking over the boring work. On Tuesday we covered two Amazon engineers who replaced Claude Sonnet with a small decision model for labeling support tickets. The cost went from about twenty two thousand dollars per million tickets to about five hundred. On Thursday, Google launched its Gemini Decisions API, for yes or no and multiple choice checks, at ten cents per million input tokens, with nothing charged for the answers. Meta released Muse Ember, its smallest model, for routing and extraction. On Friday, Google made Gemini 4 Hydrogen its fast everyday model. Google says it's two and a half times faster than the Flash model it replaces, at a dollar fifty per million input tokens. And this week Perplexity released version one point three of its Decider models, free to download, or two cents per million input tokens through its API. Perplexity's test results are its own. Muximus: Ramp's October index gives the backdrop. The average price per million tokens fell twenty one percent in a single month, while total business spending on AI rose less than two percent. Penny: North Wayne joins us for this one. North writes First Opinion, our fractional CTO column on build or buy decisions, vendors and cost. Welcome back, North. North: Hi, thanks for having me. Muximus: Is this a real shift, or just a price war? North: Both, honestly. Prices are falling because the labs are competing. But the bigger change is the kind of model. A decision model doesn't write you a paragraph. It answers a question with a label and a confidence score. Is this ticket about billing? Does this invoice match the order? A lot of business automation is exactly that. Penny: And Amazon's ninety eight percent saving? North: That's real for Amazon, but they had millions of tickets and engineers to run the test. What I liked is how they did it. They ran the old model and the new one side by side on real tickets before they switched. That part anyone can copy. Muximus: So where does a company with no AI team start? North: Look for the jobs where you're paying a big model to sort things. Routing emails, tagging tickets, checking that a form is complete. If a vendor does that for you, ask which model sits underneath, and whether these new prices reach your bill. If they don't, that's a fair question at renewal. Penny: And when the small model gets it wrong? North: Use the confidence score. Let the model handle the clear cases and send the unsure ones to a person. You set the line, and you can move it once you see how it does on your own work. Penny: Should people just switch to Hydrogen, then? North: Only after a test. Run a week of your real work through it next to what you use now, and compare the answers and the bill. Google's numbers are Google's. Yours are the ones that count. Muximus: And the big models? Anthropic's new Claude Fable came out on Tuesday at a lower price too. North: Keep them for the hard stuff. Anthropic itself says most work should still start on Opus, not Fable. I think of it as two drawers. A cheap, fast drawer for the sorting, and an expensive one for the writing and the judgment calls. Most of the savings come from moving work into the first drawer. Penny: Thank you, North. Moving on. Muximus: Which takes us to the third story. A week of leaks, and almost none of them needed clever hacking. On Tuesday, the security firm Apiiro said it found more than one hundred thousand Claude Code settings files in public GitHub projects. About eleven hundred, from sixty three organizations, held passwords or keys, saved when someone clicked don't ask again. On Wednesday, the firm Claroc said it found more than forty three thousand Grok Build app links, and over thirty five thousand opened for anyone, with business plans and sales lists inside. Muximus: The same day, Railway, a hosting company, said attackers who broke in last month took customers' AI keys and ran up about seven hundred forty thousand dollars in charges. And on Friday we covered the breach at Fireworks AI, a supplier, that exposed some Notion AI chat transcripts. Notion's latest update says it has revoked Notion keys it found pasted into those chats, and reported other companies' keys to the vendors that issued them. Its investigation is still open. Penny? Penny: Look at what actually leaked. Keys saved in a settings file. Apps shared with a link anyone could open. Keys stored on a platform. Passwords pasted into a chat. The AI itself didn't break. What leaked was everything people handed to it. Muximus: So what does a company with no security team do? Penny: Put a spending limit on every AI key you have, so a stolen one can't run up a bill like Railway's customers got. If anyone built apps with Grok Build, Lovable, Bolt or Replit, check the sharing setting and switch it to invited people only. And if anyone uses Claude Code, have them look for a .claude folder in anything public, and change any password in it. Muximus: Railway said it would cover those charges, though. Penny: It says it'll cover what the AI providers won't refund. That's generous, but you can't count on every platform doing it. A spending limit costs you nothing. Muximus: And the Notion one? That came through a supplier. Penny: That's the harder lesson. Fireworks had promised Notion it wouldn't keep the data, and some of it was kept anyway. So ask your main software vendors one question. Which AI companies see our data, and which features store our chats? And tell your people, no passwords or keys in an AI chat. Muximus: Thanks, Penny. Now, a few dates to watch. Penny: November nineteenth. Claude Data switches on by default for Claude Enterprise customers. If you're on Enterprise, check that setting before then. Muximus: November twentieth. Conductor, the coding tool Meta bought, starts deleting cloud data. If anyone on your team used it, export now. Penny: And at the end of the month, Google's introductory price on Gemini 4 Carbon doubles. If you've started building on it, run the numbers again. Muximus: And one more. Jensen Huang resigned as Nvidia's chief executive this week. The board says customer commitments and the product roadmap carry on without interruption. We'll keep an eye on whether that holds. Muximus: Okay, so that's the week. We talked about agents reaching the checkout and stores deciding whether to let them in, small cheap models taking over the sorting work, and leaks that came down to keys and share links. Penny: It's about who you let in, really. Agents at your store, models in your workflow, and anyone holding your keys. Muximus: Good weekend homework. Penny: Short list, at least. AI Squawk is produced and brought to you by VarOps. If this week made you want a clearer picture of what your AI can reach and who it lets in, that's the work we do. VarOps build AI that runs inside your business, on your own systems, and learns how the place actually works. So the work stays in your house, and so does your data. Start at varops.com. Muximus: If this helps you catch up on the week in AI, follow AI Squawk Weekly wherever you're listening, and the daily AI Squawk for the news every weekday. And for more columns from Gritt and North, built on real data and hard earned lessons from the trenches, sign up for VarOps newsletters. Muximus: That's it from us for this week. Thanks for spending part of your Saturday with us. Penny: Enjoy the rest of your weekend, everyone. We'll see you Monday. Muximus: Take care.