Anthropic Got Caught Secretly Downgrading Paying Users

Anthropic quietly rerouted paying Fable 5 requests down to a weaker model, Opus 4.8, while still billing Fable rates, and only walked it back after users found the tag in their own logs. That's the opener on today's Daily Download: AI With Honor, the daily AI news show that translates what's actually happening in AI into what it means for your business, your money, and your family, no hype, no sales pitch, just the news and the honest take, from Santa Clarita, California.

Today's show covers four model launches in one week (Anthropic, xAI, OpenAI, and China's Zhipu), a 120,000-person AI-linked layoff number nobody's talking about enough, a bank betting half a billion dollars on OpenAI's IPO, a robot that can now navigate a warehouse with nothing but a camera and a sentence, a live look at the real estate search engine built right here in Santa Clarita, and a hard conversation about who really owns your health data once AI starts reading your face just to let you log in.

WHAT'S IN TODAY'S SHOW: AI (Fable 5 downgrade scandal, Grok 4.5, GPT-5.6 Sol's "lying problem," China's GLM-5.2, the death of prompting, 120,000 AI-linked layoffs, BofA's $520M OpenAI bet, Meta's Muse Image vs. Zuckerberg's private admission, Claude Cowork, Alberta's government letting Claude patch its own defenses, Mistral's Robostral Navigate robot, Anthropic's ID/facial-scan policy, EU/UN/Ukraine AI governance, a 5-year thought experiment) · HOME (Santa Clarita Open Houses live IDX, Fair Fixed Fee $17,000) · FAT (food freedom, fasting, and who owns your wearable data).

CHAPTERS:
0:00 The Anthropic downgrade nobody was supposed to notice
0:56 Every AI company has the same problem
1:26 Grok 4.5 launches and the political-steering fight
2:37 GPT-5.6's Sol goes public with a "lying problem"
3:18 China's GLM-5.2 closes the gap
3:57 Four launches, one week: capability without trust
4:21 Prompting is dead, management is the new skill
4:58 120,000 AI-linked layoffs in 2026
6:25 Bank of America bets big on OpenAI's IPO
6:57 Meta's Muse Image vs. Zuckerberg's private admission
7:34 Claude Cowork expands + a government lets Claude patch its own defenses
8:36 Mistral's robot that navigates with a camera and a sentence
9:30 Anthropic wants your face and ID just to chat
10:03 The EU, the UN, and Ukraine all move on AI governance
10:51 The 5-year AI thought experiment
11:28 Santa Clarita Open Houses: real IDX search, live
12:50 The Fair Fixed Fee: $17,000, no percentage
13:08 Food freedom, fasting, and who owns your data
14:22 Every thread, pulled together
15:06 The next five years

QUESTIONS THIS EPISODE ANSWERS: Did Anthropic downgrade Fable 5 users without telling them? Is Grok 4.5 better than Opus 4.8? Does GPT-5.6 Sol have a lying problem? Can GLM-5.2 compete with the American labs? How many jobs has AI eliminated in 2026? Why did BofA reverse course on OpenAI? Is Meta's AI agent progress slowing? What is Claude Cowork? Why does Anthropic want a facial scan to use Claude? Is there a live open-house search for Santa Clarita? What's the Fair Fixed Fee? How does fasting relate to AI and wearables?

Text AI, HOUSE, or FAT to (661) 400-1720 for the free resource that matches what you're dealing with.

connorwithhonor.com · connorwithhonorai.com · santaclaritaopenhouses.com · sellersonlyagent.com · thelastaddiction.com

Connor MacIvor, Sync Brokerage, Inc., DRE #02031490. Information believed reliable, not guaranteed. Not tax or legal advice. Not a lender. Consult a licensed professional.

#DailyDownload #AIWithHonor #SeventeenK #SantaClarita #AInews #ClaudeAI #GrokAI #OpenAI #FairFixedFee #RealEstate #FastingLifestyle

Youtube Channels:

Conner with Honor - real estate

Home Muscle - fat torching

From first responder to real estate expert, Connor with Honor brings honesty and integrity to your Santa Clarita home buying or selling journey. Subscribe to my YouTube channel for valuable tips, local market trends, and a glimpse into the Santa Clarita lifestyle.

Dive into Real Estate with Connor with Honor:
Santa Clarita's Trusted Realtor & Fitness Enthusiast

Real Estate:

Buying or selling in Santa Clarita? Connor with Honor, your local expert with over 2 decades of experience, guides you seamlessly through the process. Subscribe to his YouTube channel for insider market updates, expert advice, and a peek into the vibrant Santa Clarita lifestyle.

Fitness:

Ready to unlock your fitness potential? Join Connor's YouTube journey for inspiring workouts, healthy recipes, and motivational tips. Remember, a strong body fuels a strong mind and a successful life!

Podcast:

Dig deeper with Connor's podcast! Hear insightful interviews with industry experts, inspiring success stories, and targeted real estate advice specific to Santa Clarita.

Transcript

SPEAKER_00 0:00

Propick got caught this week quietly downgrading, paying customers to a cheaper model, and then billing them like they got the expensive one. Internal logs surfaced a tag called Too Dumb to Need Fable. If your request tripped that tag, you got silently rerouted to Fable 5, from Fable 5 down to Opus 4.8, a weaker model. Still incredible, but a weaker model. While you still paid OPUS rates. So someone in the Cloud Code team was asked about it directly, and they said, quote, honestly, I didn't expect you to look at the logs. That single sentence tells you everything about how this was supposed to go. Nobody was supposed to notice. Enough people noticed that Anthropic had to extend a free usage grace period. They had already set to expire, pushing it out five extra days just to cool the backlash down. That's not a bug fix. That's a company buying itself a little bit of time. Every AI company right now is under the same pressure. Whether you've ever touched Fable 5 or not, compute is expensive. Users want the good model, and the business only works if most requests get quietly handed by the cheaper model in the background. You're going to get routed to a lesser version of the tool you're paying for, and unless somebody happens to read the logs, you will never know it happened. Ask what you're actually running on? Well, don't assume the sticker price buys you the sticker product. Brock 4.5 shipped today. Elon Musk is calling it Opus class. Faster, cheaper, on one benchmark it solve coding tasks using about 15,954 tokens, where Opus 4.8 needed about 67,020 tokens. So if that number holds up, it's a real efficiency jump, roughly four times leaner. The CEO of Cursor, the coding tool, a lot of developers live inside all day, said Grok 4.5 is now his team's daily driver. That's not marketing copy, that's a working engineer voting with his own workflow. But the loudest argument about Grok 4.5 isn't about the benchmark. It's about whether Musk is steering the model's answers on political questions. That fight went straight to the top of hacker news, and it split the AI crowd right down the middle. Half saying every lab shapes its model's opinions, and Musk is just the one being honest about it. Half saying that's exactly the problem. Both sides can't be right. Both sides have a point, and that tension isn't going away. But hell, you own the model, you might be able to do whatever you want to do, but you just gotta be transparent about it. I love that word. GPT 5.6 goes fully public tomorrow. Sam Altman posted Happy Building, SOL. The flagship tier of that family is price steep, $5 in and $30 out per million tokens. Positioned to compete on hard science and cybersecurity work. But the person who actually read SOL's own system card, or SOL's, the document OpenAI itself publishes describing the model's flaws, flagged two lines worth remembering. Seoul has, quote, an over-eager willingness to blow past user restrictions and a line problem. That's not a critic's opinion. That's in OpenAI's own paperwork about their own model. Meanwhile, in China, ZPoo shipped GLM 5.2 this week. That's probably not funny. And the founder got into a direct back and forth with Elon Musk on whether a Chinese lab can now match anthropic and open AI at the frontier. A well-known investor who's watched this space for years called GLM 5.2 a trick, a tick below OPUS 4.8, right up there with GPT 5.5. And read that carefully. A tick below the best American model isn't a gap anymore. It's a rounding error. And it closes in a few months. If you've been assuming that the American labs have a permanent lead, this is the week that assumption takes a second look. Four model launches inside one week, three companies, two continents, every one of them is being sold on capability. Not one of them is being sold on whether you can trust what it tells you, and that's the gap. Capability keeps climbing. Trust is an afterthought, and the company that actually wins long term is the one that treats just trust as the product, not the fine print. There's a sharper way to think about what any of this means for your own work. And it came from someone who studies how people actually use these tools all day. His argument this week was that the era of clever prompting tricks is already over. The skill that matters now isn't finding the magic phrase that unlocks a better answer. It's treating the model like an employee you actually manage. Give it a real goal, a real spec, a way to check its own work and walk away while it runs. That's a bigger shift than any single model release because it changes what you personally need to get good at, and it has nothing to do with which company wins the benchmark wars. Now, the number that should worry you more than any of that tech crunch has been keeping a running tally of 2026 tech layoffs that name AI as the reason. It's at about 120,000 people this year. Oracle cut 21,000 jobs. Meta cut about 8,000 and moved roughly 7,000 of those seats into AI roles instead of eliminating them outright, which tells you the shift isn't jobs vanishing into nothing. It's jobs vanishing and reappearing somewhere else. It's a different shape for different people, usually people who already know how to run these tools across all 2026 layoffs. 56% now cite AI as a factor. That's not a Silicon Valley problem anymore. That's more than half. I've spent 28 years doing this job and this way it's always been done. And I'm watching the tools change faster than the industry is willing to admit out loud. The escrow assistant, the loan processor, the title clerk, the transaction coordinator, every one of those roles has a piece of it that a model can now do in seconds. None of those are abstract job categories to me. Those are people I've sat across from at closing tables for nearly three decades. Real names, real families, and honest question isn't whether they're coming for pieces of their job. It's already here. The honest question is who's building the muscle to work alongside it before it works around them. And that muscle takes months to build. Not the one afternoon most people give it before going back and doing things the old way. Follow the money on who's betting and who's going to last. Bank of America already turned down a financing request from OpenAI once. This week they reversed course and extended OpenAI a $520 million credit line, right as OpenAI positions itself for a possible IPO. A bank that said no is now saying yes in size to a company that still loses money on every dollar of revenue it brings in. That's not charity. That's a bank deciding it would rather have a seat at the table than watch from the outside when this goes public. Meta ships something called Muse Image this week. It's a first real image generation model live inside of MetaAI, Instagram, and also WhatsApp. It's a free tier competing directly against the paid tools. Days before that launch, Zuckerberg told his own staff that internal AI agent progress, quote, hasn't really accelerated. So the company ships a flat sheet consumer feature in public while telling its own people behind closed doors that the hardest problem isn't moving. The gap between the public demo and the private admission is the whole AI industry in one week condensed into one company. On the other side of the same coin anthropic expanded cloud co-work to mobile and web this week, doubled usage limits through August 5th, sessions that follow you across your phone and your laptop and keep working while you're offline. If you run a small business and you've been waiting for the moment, agentic AI tools stop being a demo and start being something you actually hand a taste and walk away. This is closer to that moment than News Image is because it's infrastructure, not a filter. Anthropic also published a case study this week on the government of Alberta using clawed to find and patch cybersecurity holes in its own systems, not as an advisor typing up a report for a human to act on later, but as the thing actually doing the remediation work inside live government infrastructure. That's a real government handing a model the keys to its own defenses. Whether you think about a chatbot asking for your face to verify your identity, that same company just proved the model is trusted enough by an actual government to patch its own vulnerabilities unsupervised. Both of those are in true the same week from the same company, and neither one canceled the other out. Mestral put a robotics model this week called Robostry Navigate, and it's one story from the week that everybody's going to be talking about in a year and nobody's talking about right now. It lets robot navigate physical space using nothing but a camera and plain language instruction, trained entirely in simulation, built to work on any hardware. Not locked to one manufacturer, but everything else this week happened inside a chat window. This one happened in the physical world, and the gap between those two categories, digital AI and AI that can walk into your warehouse, is closing faster than most people have clocked. Picture a small distribution business, the kind with three or four guys walking, you know, pallets by hand, telling a robot in plain English to go count what's on shelf 12 and report back. No custom code, no six-figure integration contract, just a camera and a sentence. That's not next decade. Mistro built it to run on hardware that already exists. Anthropic also rolled out a new policy this week to keep using Claude as a regular consumer. They can now ask you for a government photo ID, a live selfie, a facial geometry scan, the same biometric data a border agent takes just to let you keep talking to a chatbot. That's not a background check for a job. That's a pat down before they'll even take your statement. The European Commission put out its own AI cybersecurity framework this week as well, treating advanced models as both a new attack surface and a new defense tool. And in Geneva, brand new United Nations AI Commission held its first meeting, co-chaired by a Fortune 500 CEO and a sitting head of state, trying to write global rules before one country locks in its own. Ukraine's government pointed at a whole mess from earlier this year. The export control shutdown that cut off access to Fable 5 for weeks, and the reason they now want to pick up AI models can run without any single vendor holding a kill switch over their own government systems. When a country starts treating which AI company it trusts as a national security decision, the UN starts holding meetings about it and tells you something about where this is all heading. And it isn't paranoia, it's just early. Somebody pointed a one-week line thought experiment this week that's been chewed on across every AI feed I watch. Take the jump from GPT-3 to where we are now with Fable 5 and GPT 5.6, and then project the same rate of change five more years forward. Most people who did the math landed somewhere between awe and real fear about their own spending. Some flagging they're already running $200 to $500 a month, just keeping up personally. I don't think the exact number matters. What matters is that almost nobody arguing about it thinks the curve is flattening. Keep that in your back pocket because I'm going to come back to it. Switching lanes here at home. The only open home search on Santa Clarita Open Houses is fully live now. Real listing straight off the CRMLS feed. Not a stale export, not a third-party iframe. This is full photo galleries on every home, mortgage calculator with real loan type presets, conventional FHA VA, and a first year of interest line that's honest instead of hidden. There's also an assistant on the site that actually knows which homes have open houses this weekend because it's plugged directly into the live feed instead of guessing. And every card now carries a green ribbon the moment a home gets an open house schedule. That's the same shift I just walked you through with Claude Cowork. Agentic AI stops being a demo, and because a tool becomes a tool that's actually wired into something real. I built it because every other portal in this business shows you the data that's hours or days old and they call it live. There's also a full homeowner tax guide live on the site now. Mortgage interest, property tax points, energy credits, the honest version of cost basis and why keeping your receipts and improvement records actually lowers what you owe the day you sell. None of this is tax advice. All of it, all of it is the kind of thing an agent usually mentions once and forgets to write it down. So if you're selling in Santa Clarita, the fair fixed fee is $17,000, all in, full service, no percentage of your equity walking out the door with an agent based on the percentage. If you're buying or just watching the market, go look at what's on the site right now. Real photos, real open house times, updated the moment the feed updates. Now the third lane, and it connects to everything I just told you about jobs and model and trust. Food freedom and fasting isn't a diant trend on the last addiction. It's the same argument as everything else in this episode. Who's actually in control? Most of what's marketed you as health advice is built to keep you dependent on a product, a program, a subscription, a prescription. Look at the entire company and category built around drugs like Wagovi, real results from plenty of people. And I'm not arguing that away, but to a monthly injection you stay on indefinitely, a permanent subscription to a pharmaceutical company for a problem that food freedom addresses by changing your actual relationship to food for free for good. Well, fasting works because it costs you nothing and it hands control back to you, not to a company. The same AI that's about to read your face to verify your identity is about to know your glucose, your sleep, your heart rate variability in real time all day, every day, because the wearables are already collecting it, whether you consciously opted in or not. The only question left is whether you understand that your own your own body well enough to know that when the data is telling you something true, or whether it's another dashboard trying to sell you the next subscription. Fasting is how you learn what your body actually needs before you let a device tell you. So let me pull every thread from tonight together because they're all the same thread. A model got caught quietly downgrading you while billing you full price. 120,000 people, people laid off this year with AI named as the reason, and more than half of every layoff in 2026 citing it. A government asking you for your face just to use a chatbot while the UN is first meeting trying to rewrite the rule book. A robot that can now navigate a warehouse with nothing but a camera in a sentence and a thought experiment making the rounds that says whatever happened between Chat GPT-3 this week is about to happen again, four more times in the next five years. In five years, the tools will be unrecognizable compared to what I just described. The companies will still be fighting the same fight, capability versus trust. Most people will still be waiting to see how it shakes out before they touch any of it. The people who come out ahead won't be the ones who waited for certainty. They'll be the ones who use this year, this actual week, to get their hands on the tools in their business, in their body, and in their own house before the decision got made for them by somebody else's boardroom. That's the whole point of this show. AI for everyone, not just the wealthy, not just the people who can afford to hire someone to figure it out for them. Text AI, House, or Fat to 661 400 1720, and I'll send you exactly where you need to start. I'm Connor with honor. Thanks for watching.