Announcer: This is AI Squawk for Tuesday, the tenth of November, brought to you by Var Ops, the resident AI company. Muximus: Welcome to the AI Squawk. I'm Muximus. Penny: And I'm Penny. Three main stories today. Muximus: And four quick ones. Penny: And as always, what each one means for you. Muximus: First up. Anthropic has a new top model, and it costs less than the last one. On Monday the company released Claude Fable five point two. Anthropic says it's better at coding, reasoning and reading images, and up to thirty percent faster than version five point one. The price is eight dollars per million input tokens and forty dollars per million output tokens, which Anthropic calls a lower price. A fast mode runs up to two and a half times quicker, at double the price. Its twin, Mythos five point two, is the same model with fewer safety limits for cybersecurity and biology work, and only vetted organizations get it. The claims are big. Anthropic says a thousand Fable agents turned the proof of Fermat's Last Theorem into code a computer can check, in thirty hours. It also says Claude now writes over ninety percent of the code changes merged into Anthropic's own core codebase. Fable is available today from Anthropic, and on Amazon, Google and Microsoft's clouds. Penny? Penny: The line I'd pay attention to is Anthropic's own advice. For most coding and office work, it says start with Opus five point five, its everyday model, and only move up to Fable for the hardest jobs. Muximus: So the company selling the big model is telling you not to use it for everything. Penny: Pretty much. And there's a detail in its own results table. On a test of business workflows, Fable scored about fifty one percent, and Google's Gemini four Carbon scored a little higher, almost fifty four. Those are Anthropic's numbers, by the way, not an independent test. Muximus: So what should a company actually do? Penny: Don't move everything over because it's new. Try Fable on one hard job where you already know the right answer, and compare the bill. And read the data terms. Anthropic says it keeps prompts and responses for these models for thirty days for safety monitoring, unless you have a zero data retention agreement. Worth knowing before anyone feeds it client files. Muximus: Next. Amazon swapped a big A I model for a much smaller one on a boring job, and says the bill dropped ninety eight percent. Two Amazon engineers wrote it up on the blog of Strands Agents, Amazon's open source agent project. Their team runs one of the largest internal ticketing systems in the world. Claude Sonnet four point six was sorting those tickets into twenty categories, things like customer escalations, alarms and security findings. They replaced it with Strands Decider, a small model that can't write text. It just picks one option from a list and says how confident it is. They say it labels tickets at least as well. Each call went from about four seconds to under one, and the cost per million tickets fell from about twenty two thousand dollars to about five hundred. They expect to save millions of dollars a year. Penny: North Wayne joins us for this one. North writes First Opinion, our column on build or buy decisions, the fractional C T O for companies without one. Welcome, North. North: Thanks, Penny. Good to be here. Muximus: So, ninety eight percent. Is that real, or is Amazon selling its own model? North: Bit of both, probably. It's Amazon's own project, so of course they're happy to write it up. But the numbers are specific, and they're upfront about when it won't work. If your labels need an explanation, or your list of categories changes every week, they say this probably isn't for you. Penny: What I liked was how they made the switch. North: Yeah, that's the part I'd copy. For a while both models ran on the same real tickets. The old one still made the call, and they checked every case where the two disagreed. And going back would have taken them forty five seconds. Muximus: Most companies listening don't have engineers who'd build any of this. So what's in it for them? North: Look at what you're paying a big model to do. A lot of business A I is really sorting. Which inbox does this email go to, what kind of invoice is this, is this complaint urgent. If a vendor charges you big model prices for that, it's fair to ask what's running underneath, and whether something smaller would do. Penny: And even if you don't build anything, you can still ask for that side by side test. North: Right. Ask them to run the old setup and the new one on your real data for a few weeks, and show you where they disagree. If they can't do that, I'd wait. I just like that the fix here was a smaller model. Penny: Thank you, North. Moving on. Muximus: Which takes us to the third story. Claude Code, Anthropic's coding tool, has been leaking passwords onto GitHub. Researchers at the security firm Apiiro found more than a hundred thousand Claude Code settings files sitting in public GitHub projects. About eleven hundred of them, from sixty three organizations, held secrets like database passwords, A P I keys and private keys. Here's how it happens. When you approve a command and pick yes, and don't ask again, Claude Code saves that command to a settings file, sometimes with the password still in it. That file is meant to stay on your computer, but it can get uploaded with the rest of the project. Ars Technica reports the problem was first flagged in a bug report in August twenty twenty five. Anthropic's Boris Cherny says the team is looking into it, and suggests secret scanners for now. But Apiiro says two popular scanners missed two hundred fifteen of the leaked secrets. Penny? Penny: What gets me is that nobody did anything unusual here. Someone clicked a button to save themselves a little time. Muximus: Does this matter to a company that doesn't write software? Penny: It does if anyone builds things for you. A contractor, an agency, or someone on the ops team making a small internal tool with Claude Code. If that project ever went onto GitHub, ask them to check it for a folder called dot claude. Muximus: And if they find something? Penny: Then change those passwords. Deleting the file isn't enough, because the secret can still sit in other people's copies of the project. And one simple rule for everyone. Don't type passwords straight into commands you hand to an A I tool. Muximus: Thanks, Penny. Now, quick hits. Penny: ChatGPT Enterprise is now for sale through the A W S Marketplace. If your company already has a spending commitment with Amazon's cloud, eligible spend can count toward it. OpenAI says it's the same product you'd get directly. Muximus: Cisco and IBM are building Anthropic's Mythos model into their security tools. Cisco says its A I Defense product can now use Mythos to probe customers' A I systems for weak spots, at no extra cost under existing licenses. IBM is adding it to Concert, its tool that scans applications for risks. Penny: Apple has started charging for extra A I use. Logic Pro is the first app to sell what Apple calls Intelligence Usage Packs, at forty nine ninety nine, and The Verge says more Apple apps, including Pages, Keynote and Numbers, should get them this month. Muximus: And OpenAI says cost exports from its Costs A P I published between November sixth and ninth had incorrect numbers. It says the issue is now resolved. If your finance team pulled A I spend reports in those days, pull them again. Muximus: Okay, so that's the news. Today we talked about Anthropic's new top model and its lower price, Amazon cutting a bill by swapping in a much smaller model, and Claude Code leaving passwords where they shouldn't be. Penny: It's about picking the right size of tool, really. Even Anthropic says don't use its biggest model for everything. Muximus: And maybe think twice before clicking don't ask again. Penny: Good rule for most things. AI Squawk is produced and brought to you by Var Ops. If today got you wondering what your A I tools really cost, or what they keep, that's the work we do. Var Ops build A I 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 Var Ops dot com. Muximus: If this helps you get more out of the tools you already have, follow the AI Squawk podcast wherever you're listening, so you don't miss an episode. And for more columns from North, built on real data and hard earned lessons from the trenches, sign up for Var Ops newsletters. Muximus: That's it from us for today. Thanks for spending a few minutes with us. Penny: Enjoy the rest of your Tuesday, everyone. We'll see you tomorrow. Muximus: Take care.