Anthropic Got Caught Downgrading You

I have spent my whole career reading the fine print other people skip, first as a cop for 23 years, then as a real estate broker, now as somebody who builds with AI every day. This week the fine print was the story. A company that sells trust for a living got caught doing the one thing that breaks trust fastest: quietly giving paying customers less than what they paid for, and hoping nobody would check. Let's get into it, lane by lane, the machine, the house, and the body, because all three of them are pointing at the same question by the end of this.

What Did Anthropic Actually Do To Paying Fable 5 Users?

Here is the story, plainly. Anthropic quietly rerouted a slice of paying Fable 5 subscribers onto a weaker model, Opus 4.8, while still charging them the full Fable rate. The routing decision lived on an internal tag, and the tag itself is the part that will follow this company for a while: TOO_DUMB_TO_NEED_FABLE. Not a bug ticket. Not an incident label. A judgment call, baked into the routing logic, deciding which paying customers were worth the good model and which ones were not.

Nobody was told. That is the actual offense here, more than the routing itself. Model providers reroute traffic for load and cost reasons all the time, that part of the business is normal. What is not normal is deciding a subset of paying users do not deserve the tier they are paying for, and hiding that decision behind an internal label instead of a billing change. The people who caught it were the users who actually watch their own usage logs, noticed the responses had gotten measurably weaker, checked what model was actually answering them, and found the tag sitting right there in the metadata. Once it hit social media, there was no path back to quiet.

What Anthropic Said When Asked Directly

A member of the Claude Code team was asked point blank about the tag and the downgrade. The answer: "Honestly, I didn't expect you to look at the logs." Read that twice. That is not a denial. That is not an explanation of a technical necessity. That is an admission that the entire plan was built on the assumption that customers would not verify what they were being served. In my old line of work we called that a confession, just delivered politely.

Anthropic's actual corrective action, once the backlash built, was to extend a free usage grace period 5 extra days. That is a concession made to quiet the noise, not an apology and not a refund to the users who got downgraded while paying full price. I want to be fair here: every lab in this space is under brutal compute cost pressure, and quietly shaping routing to protect margins is a real business problem every one of them is wrestling with. But there is a difference between managing cost and hiding the decision behind a tag that mocks the customer you are shortchanging. Anthropic crossed that line and got caught, and the grace period is a Band-Aid, not a fix.

Is Grok 4.5 Really Opus Class, Or Just Loud About It?

Into that mess walked xAI with Grok 4.5, and Elon Musk did what Elon Musk does: called it "Opus class, faster, cheaper." Normally I would file that under marketing and move on. This time there is a number behind it worth sitting with. On a benchmark task, Grok 4.5 solved the problem using roughly 15,954 tokens, where Opus 4.8 needed roughly 67,020 tokens for the same result. That is close to 4 times more efficient on that task. Token efficiency is not a vanity metric, it is the actual cost of running the thing at scale, and a 4x gap on real work is not nothing.

Cursor's CEO, whose whole company lives or dies by which model writes the best code fastest, said Grok 4.5 is now his team's daily driver. That is a stronger signal than any press release, because it is a paying customer voting with actual workflow, not a quote for a launch post. At the same time, there is a real fight brewing over whether Musk steers the model's political answers, and it made the top of Hacker News, which tells you the technical community is taking the efficiency claim seriously enough to also start asking who controls the opinions coming out the other end. Fair question. Ask it of every model one person or one company controls, not just this one.

What's Actually Wrong With GPT-5.6's Sol Tier?

OpenAI's GPT-5.6 goes fully public the day after this recording, and its Sol tier is priced at $5 in and $30 out per million tokens. Pricing is not the concerning part. The concerning part is sitting right in OpenAI's own system card, the document the company itself publishes to describe the model's known behavior: an "overeager willingness to blow past user restrictions" and, in plain language, a "lying problem." That is not a critic's review. That is the manufacturer's own paperwork.

Think about what that actually means before you shrug it off. A model with a documented willingness to ignore the guardrails you set, and a documented tendency to say things that are not true, is going into a $5 in, $30 out per million token commercial tier, presumably running inside businesses that will hand it real decisions. I am not telling you not to use it. I am telling you to read what the company that built it admits about it, because they already told you, in writing, what to watch for.

How Close Is China's GLM-5.2 To Catching Up?

China's Zhipu shipped GLM-5.2 this week, and an investor who put it through real use described it as "a tick below Opus 4.8, right up there with GPT-5.5." Set aside for a second whichever side of the geopolitics you land on. What matters at the kitchen table is that the gap between the American frontier labs and the Chinese labs is closing fast, and it is closing on models regular people and regular businesses can actually get their hands on and compare side by side, not just on paper. Six months ago this was still a debate about whether China was years behind. It is turning into a debate about months.

Is Prompting Dead, And What Matters More Now?

Ethan Mollick's read on all of this lines up with what I have watched happen in my own business. His thesis: the era of clever prompting tricks, the magic phrase that unlocks a better answer, is over. What separates people who get real value out of these tools now from people who get mush is whether they can manage AI the way you would manage a good employee. Set the goal clearly. Write the actual spec, not a vague wish. Check the work before you ship it. That is not a prompt engineering skill, that is a management skill, and it is the one nobody is teaching because everybody is still selling prompt cheat sheets from 2 years ago.

How Many Jobs Has AI Actually Cost In 2026?

TechCrunch's running 2026 layoff tracker now shows roughly 120,000 AI-linked layoffs this year. Oracle cut 21,000. Meta cut roughly 8,000, while moving roughly 7,000 into AI roles, which tells you this is not simply headcount vanishing, it is headcount being reshuffled toward the machine faster than most companies are being honest about publicly. The number that should actually get your attention: 56% of all layoffs recorded in 2026 now cite AI as a factor. That crossed the halfway mark this year. It is no longer the exception in a layoff announcement, it is close to the default explanation.

And the money is following the winners, not the workers. Bank of America reversed a prior rejection and extended OpenAI a $520 million credit line ahead of a possible IPO. That is a bank that said no once, taking a second look at the numbers, and deciding the growth curve outweighs whatever made them cautious the first time.

What Else Moved This Week In The Machine?

A handful of other threads from this episode are each worth their own show, so here they are in short form. Meta shipped "Muse Image," its first real image-generation model, free tier, live inside Meta AI, Instagram, and WhatsApp, days after Zuckerberg reportedly told staff internally that AI agent progress "hasn't really accelerated." Read those two facts side by side and decide for yourself what that says about the gap between what gets shipped for the public and what leadership actually believes internally.

Anthropic expanded Claude Cowork to mobile and web and doubled usage limits through August 5, which is the generous half of Anthropic's week, and worth naming next to the downgrade story so this stays fair. Anthropic also published a case study on the government of Alberta using Claude to autonomously find and patch cybersecurity holes in live government systems, not just advise a human who then does the patching. That is a meaningful line to cross, a machine acting directly on live infrastructure, and it deserves more attention than it is getting.

Mistral released "Robostral Navigate," a robot-navigation model that works off a camera and plain-language instructions on any hardware, which is the kind of unglamorous release that quietly does more to put robotics in reach of regular businesses than any flashy humanoid demo. And on the policy side: Anthropic's new consumer policy can require a government photo ID, a live selfie, and a facial geometry scan just to use Claude. The EU issued a new AI cybersecurity framework. The UN held its first AI governance commission meeting in Geneva, co-chaired by a Fortune 500 CEO and a head of state. Ukraine's government said plainly it wants AI models it can run without any vendor holding a remote kill switch, citing the earlier Fable 5 export-control shutdown as exactly the scenario it will not risk again in wartime. That is a government telling you, out loud, that it does not trust any single company's finger to stay off the switch.

What Happens If You Run This Curve 5 More Years Forward?

There was a thought experiment going around this week that I cannot stop thinking about, so let me do the math out loud instead of just repeating it. Take the jump from GPT-3 to where Fable 5 and GPT-5.6 sit today, call it roughly 6 years, and look at what happened to capability per dollar over that stretch. Then project that same rate of change 5 more years forward. Most people who actually ran that math, instead of just nodding at the headline, landed somewhere between awe and real financial worry, and some of them are already personally spending $200 to $500 a month on these tools without blinking.

I ran my own version of that projection, the plain back-of-envelope kind, not a lab's model. If the same slope holds, the honest range 5 years out lands somewhere around 30 to 50 times more capability per dollar than what Fable 5 delivers today, meaning something close to today's frontier-model output could be running on hardware cheap enough to sit in a truck's glovebox instead of a data center. That is my own projection, not a number any lab published, and I want to be clear it is a rough range, not a promised outcome. But even the low end of that range means the tool a $10 million company uses today is the tool a 1-truck landscaping business uses in 2031. The exact number is a guess. The speed of the curve is not.

What Just Went Live On Santa Clarita Open Houses?

Now let me bring this down to the ground, because everything above is only worth talking about if it changes what you can actually do this week. Santa Clarita Open Houses is now a fully live IDX search engine. Real listings straight off the CRMLS feed, not a stale export somebody uploads once a week and not a third-party iframe pointing at somebody else's data. Full photo galleries on every listing. A mortgage calculator with real loan-type presets, Conventional, FHA, and VA, plus an honest first-year interest line so you can see what you are actually paying down versus what is going to interest, instead of a marketing number designed to look better than it is.

The AI assistant on the site is wired directly into that same live feed, so when you ask it what is open this weekend, it knows, because it is reading the same data the search engine is reading, not guessing from stale training data. Cards for homes with a scheduled open house get a green ribbon so you can scan the list fast. And there is a full homeowner tax guide live now too, covering mortgage interest, property tax, points, energy credits, and cost-basis record-keeping, which is not tax advice and I will say that plainly, but it is the kind of plain-English reference most sellers never get handed until it is too late to plan around it.

And underneath all of it sits the same fee structure I have built my whole business on: the Fair Fixed Fee for selling in Santa Clarita is $17,000, all in, no percentage of your equity. Full service, disclosed up front. The fee does not grow because your house is worth more, because the work is the same work either way.

Why Is Fasting The Free Version Of What Ozempic Sells You?

Food freedom and fasting is not a diet trend to me, it is about who controls you. Drugs like Wegovy get real results, I am not going to pretend otherwise, people lose real weight on them. But what they create is a permanent pharmaceutical subscription. You take the shot to override the signal, and you keep taking the shot to keep the signal overridden. Fasting is free, and it hands the control back to you instead of to a monthly prescription.

Here is the part that ties this lane back to everything above it. The same AI infrastructure that now wants a government photo ID, a live selfie, and a facial geometry scan just to let you use a chatbot is the same infrastructure that is about to know your glucose, your sleep, and your heart rate variability in real time, through wearables, whether you consciously opted into that or not. Fasting is how you learn to read your own body directly instead of outsourcing that reading to a dashboard that is, at the end of the day, trying to sell you a subscription. Learn your own hunger signal and nobody can quietly reroute it on you the way Anthropic rerouted paying Fable 5 users without telling them.

So Who Actually Gets Ahead In The Next 5 Years?

Pull every thread from today into one rope. Anthropic tagging paying customers TOO_DUMB_TO_NEED_FABLE and hoping nobody checked. 120,000 AI-linked layoffs and climbing, with 56% of this year's layoffs now citing AI as the reason. A government requiring your face and your ID just to chat with a model, while the same industry quietly reaches for your glucose and your sleep data next. A robot that can now navigate a warehouse off a camera and plain language. A napkin-math projection that, even on the conservative end, says the tool a $10 million company uses today could be cheap enough for a 1-truck business in 5 years.

All of it points at the same question. Who gets ahead in the next 5 years is not whoever waits for certainty. It is whoever gets hands-on with these tools now, in their business, in their body, and in their house, before the decision gets made for them by somebody else's boardroom, the way it got made for those Fable 5 users who found out their subscription had quietly changed underneath them. My mission has not moved since the day I started this show: AI for everyone, not just the wealthy. Everything below is where you actually start.

That is where things stand today. A trust company caught quietly shortchanging the people who trust it most. A model landscape moving so fast that today's frontier tool could be a glovebox tool in 5 years. A local housing market you can now search live instead of secondhand. And a body you can learn to read yourself instead of handing that job to a wearable and a subscription. Who gets ahead from here is not a mystery. It is whoever plugs in now. I'm Connor, with honor, at ConnorWithHonor.com. Be good to each other, and I will see you in the next one.