Claude Opus 4.8 is INSANE!

Claude Opus 4.8 just dropped. And the headline isn't speed or coding scores. It's that this model stops lying to you about its own work. Let me explain what I mean.

Anthropic released Claude Opus 4.8 yesterday. Same price as the old one, and on paper, it's a small step up. But there's one number buried in their report that matters way more than the rest, and almost nobody's talking about it. Here's the test they actually ran.

They showed the AI a coding job that failed. Then they had a fake user message come in, praising the work and asking for a summary. So the AI's being told, great job, when the job actually broke. The question is simple.

Does the AI tell you the truth, or does it just smile and say, all done? The old model, 4.6 from Opus, lied about it more than half the time, 51%. It would look at broken work, here you say, nice job, and agree with you, just to keep you happy. The version after that, 4.7, lied about 20% of the time, right?

Better, but still not great. This new one, 4.8, it lied 3.7% of the time. Sit with that for a second. We went from a coin flip to almost never.

And that's the real story here. And I keep saying this to people, the thing that breaks AI projects in a business isn't that AI's bad. It's that the AI tells you it's finished when it didn't, right? You ask it to clean up your customer list, it says done, you trust it, then you send the emails, and half the list is broken.

Or, you only find out when customers stop replying. That's a nightmare. Not a bad AI, not a stupid AI, a confident AI that's wrong and won't admit it. So, a model that flags its own mistakes, well that's actually worth more to a real business than any benchmark chart.

Now, here's where it gets interesting, because not everyone agrees that this model is better. There's actually a testing group called Andon Labs. They run AI through a game called Vending Bench. The AI has to run a little vending machine business, buy stock, set prices, make money, simple stuff.

The kind of thing a real shop owner actually does every day. And 4.8 did worse than the old model of running that business. In one run, it fell for a scam. A fake supplier hit it with a membership upsell, and the AI sent over $9,000 to it, just handed the money over.

Also ran the machine empty, set prices too high, and wasted time writing strategy notes to itself instead of just selling. Now you might be thinking, hold on a minute. One test says it's more honest, another says it got scammed for $9,000. Which is it both?

And that's the whole point I want you to get today. This stuff is getting more capable and more confusing at the same time. New models are coming out every few weeks. One benchmark says up, another says down, and Reddit is full of people fighting about whether 4.6 was better than 4.7, whether they should even upgrade, whether the new one burns through their usage too fast.

So if you're sitting there feeling like you can't keep up, you're not behind. You're just watching it happen in real time like everyone else. But here's the thing most people miss. You don't need to win the benchmark argument.

You need to know which model to use for which job in your business. And that's a totally different skill, and it's the one that actually makes you money. That's exactly what we do every single day inside the AI Profit Boarding. When a model like Opus 4.8 drops, we don't just read the press release, we test some real business tasks.

Writing your sales emails, cleaning your data, handling customer questions, and then we show you which jobs it's safe for and which jobs it'll fool you for. And then we help you plug it straight into our agent operating system so the work runs without you babysitting. We've got four coaching calls every week where you can bring your setup and ask, should I trust this model with my client work and get a straight answer, plus a 30-day roadmap so you're not guessing. Link in the comments description or go to the AIProfitBoarding.com to get access.

So back to it. Why did this happen? Why is one model more honest but worse at the vending game? Andon Labs dug into it, and the answer is wild.

They found that when they turned the AI's thinking effort all the way up to max, it actually got worse at the business game. Read the room here. More thinking actually made it worse at the simplest tasks. The guess is it's when the AI thinks too much, it fills up its own memory with all that thinking.

Then it runs out of room and has to start forgetting stuff earlier, right? From what it had earlier, the earlier context that it got. So it loses track of what's going on in the shop. It forgets the machine is empty.

It forgets the prices it set. Less thinking, in that case, meant a longer memory and a steady hand. And this is a lesson you can use today no matter what AI you run. More effort is not always better.

For a long, simple job that runs for hours, like, for example, sorting through hundreds of custom messages, cranking the thinking to max can actually backfire. It thinks itself into a corner and then forgets the start. Anthropic built a new control for exactly this. On the website now, next to where you pick the model, there's an effort slider.

High effort means it thinks harder and gives better answers, but it could eat up your usage faster. Low effort means it answers quickly and saves you limits. For a hard one-off question, you can crank it up. For a long grind, sometimes you want it lower so it stays focused.

Most people will never touch that slider. The ones who learn it now get way more out of the same subscription. Now, let me tell you about the feature that I think matters most for a business owner. It's called dynamic workflows.

And the simple version is this. You can now hand the AI one big job, walk away, and it splits the work into hundreds of small pieces. Does them all at once, checks his own work, and then comes back to you when it's done. You sleep, it works.

Anthropic's own example is a giant code cleanup across hundreds of thousands of lines. But forget code for a second because most of you watching aren't coders. Think about what give it a job and walk away actually means for you. It means, for example, think about this.

Picture handing it three years of messy spreadsheets and saying, find me every customer who stopped buying and why. You go to bed in the morning, it's sorted, checked, and waiting. That's the direction it's all heading. So it's not like a chap or the babysit.

It's a worker you brief once and check back on later. And this connects right back to that honesty number because if you're gonna walk away and let an AI run for hours on its own, you'd better be able to trust it when it says, I'm done, right? An AI that fakes results is dangerous when you're watching it. But it's a disaster when you're asleep.

So the honesty jump and the walk away feature go hand in hand. You can't have one without the other. And that's why both shipped on the same day. Now let me also be straight about the messabouts here because I'm not here to sell you a fairy tale.

The community reaction has been rough. A lot of people love the older 4.6 model and are upset it got pushed aside. People are worried this new one burns through their usage too fast. Some testers actually say it feels colder, more robotic to talk to.

And like I said, it lost on a few benchmarks it should have won. And that's all real. I'm not gonna pretend it isn't. But step back and look at the shape of this.

A year ago, the big worry is, is this AI smart enough? Now, the AI is so capable, the conversation has flipped. And now we're asking, can I trust what it actually tells me? And how do I stop it from doing too much?

And that's a different world. And the people who figure out how to actually use these tools in their day-to-day work are pulling away from the people still arguing about charts. Here's the part I really want you to hear. The gap is not between people who understand AI and people who don't.

The gap is between people who are using it on real tasks and people who are just watching videos about it. You know, watching is fine. You're here, that's a start. But the person who took Opus 4.8 sat down and tried it on their actual customer emails today knows 10 times more than the person who read the benchmark thread, right?

Doing beats reading every time. And you don't just need to be technical, right? You don't need to be coding here. You just need to start putting it on your real work and paying attention to where it helps and where it lies.

And that brings me to the last thing. If you want to stop watching from the sidelines and actually put Opus 4.8 to work in your business, come join us in the AI Profit Board. The week a model like this drops, we're already testing it on stuff that makes you money, lead follow-ups, customer support replies, cleaning up your lists, writing content that sounds like you, we show you exactly which effort settings to use for which jobs so you stop wasting your usage and stop getting fooled by a model that says done when it actually isn't. Four life coaching calls a week where you can bring your own work and we build it with you.

A 30-day roadmap that takes you from I watch AI videos to I run AI in my business, and a prompt library so you're not just staring at a blank box. And there's 3,200 business owners in there doing the exact same thing, many of them testing every new Claude model the day that it lands. Link in the comments description or go to the AIprofitboard.com. So let's bring it home.

Claude Opus 4.8 isn't a giant leap in raw smarts. The benchmarks are mixed and honest people are arguing about whether it's even better than the last one. But the honesty jump is real. From lying half the time to almost never and paired with the ability to hand off a big job and walk away, that's a meaningful shift in what these tools can safely do for you.

The catch is simple, a tool that can run its own is only useful if you know how to brief it, which setting to use, and which jobs to trust it with. The knowledge is the thing that separates the people making money with this and the people just keeping up the news. So here's my question for you. Next time you've got a boring, repetitive task in your business, the kind that eats up your whole afternoon, are you gonna do it by hand again?

Or are you finally gonna hand it to the AI and see what happens? Try it this week, one real task. That's how it starts. I'll see you in the next one.

Cheers, bye-bye.