Anthropic Built an AI So Dangerous They Won’t Release It (Claude Mythos)

Oh boy, it's happening. Everybody stay calm. What's the procedure, everyone? What's the procedure? Stay fucking calm. Another one of these big cataclysmic AI events, leaks, releases. Well, not even releases, more like previews. Claude Mythos. It's another one of those significant moments in AI history and AI development where we just need to talk about this. Claude Mythos was previewed by Anthropic, their new flagship model. It's not even released. But the numbers on this, the benchmarks on this are unbelievable. I've spent the past few hours looking at this. I've had some pretty good conversations here. And I think I have 2 to 3, probably like 3 insights that to me seem non-obvious, yet significant for you to want to get the most out of this. But to really just wrap your head around what's happening in this space and in this world, because this model simply puts Stan's head and shoulders above anything else we've seen about all the OpenAI models, about all the Anthropic models. It's yet again, about 6 months ahead of what we see in the open source space. I want to talk about all of that. But before I even switch to my screen and start, you know, showing you, you know, benchmark numbers that will, frankly, like benchmarks don't impress people these days anymore. These are unbelievable. How did they make this jump? I don't know. One thing that comes to mind is like, no wonder that they've been shipping a new feature every day for the past month. If they're using this internally, one. Two, they haven't released this yet. So if you didn't hear about the release, it's basically new Anthropic model, their new flagship model, and they're not releasing it. They created what they call Project Glasswing, which is a partnership with, well, all the large companies, all the large tech companies to use Claude Mythos to harden their security before they put it out into the world. Because one obvious thing that this does, it can hack and it can exploit security vulnerabilities. Now, here's my first non-obvious insight. Okay, I think this is a really big one. I've been talking to a lot of people recently. As you might know, I've been spending a bit too much time with OpenClaw. And, you know, that's a whole separate video that's worth exploring. By now I have like 12 workflows that run regularly, some of them weekly, some of them daily, that I wouldn't want to give up anymore. I think content creation, you know, I'll preface that by saying I think content creation is one of the areas where you can benefit the most from OpenClaw. You know, probably plumbing is one of the ones where you can use it the least, you know what I'm saying? But here's the non-obvious insight. What made OpenClaw work, right? I'm not saying it works for everybody. I realize a lot of people have trouble with it. I think it's usually in the context layer, but what made this shift happen, right? Because it's not the code. It's not the fact that somebody put together, you know, all these tools, you know, shout out Peter. He did an incredible job with that, with building this. But we've seen versions of this over the past years. For anybody who's been around like in, what was it, like early to mid-2023, watching the development of all this, we had projects like, what would they call it? It was like Baby AGI was one of them. Another one was AgentGPT. And they were basically attempting the same thing, like an autonomous system that just runs and gets things done. And back then it ran a GPT-3.5. And I guess GPT-4 when that came out, it just wasn't good enough. The whole thing wasn't good enough. And I've been giving these some thought recently. Why was that? Why wasn't it good enough? It was the underlying models. It wasn't the code. It wasn't the fact that the Telegram connection didn't work. Heck, I had a YouTube tutorial back then on how to build Telegram bots with GPT-4 and voice input, all that, right? It was a thing you could do in 15 minutes as a non-coder. It was that simple, right? So it was certainly not in that connection. And Igor, maybe it was like in the setup or the heartbeat or the soul. Those are all nice additions. The whole thing is really, you know, incredible. Here's my insight. The difference maker was the model. It's really the Opus 4.5 release that was kind of the tipping point. Now we have Opus 4.6, everything works even better. But it was really these model releases were the thing that made agentic tools work, right? If you've played around with them, the browser view still is so far from perfect. It's so inconsistent. But its ability to hold context and its ability to just work for 20 minutes and actually not have amnesia about what it worked on 15 minutes ago, incredible. Its ability to hold a large amount of context is what I'm saying, one, but also its ability to adhere to context. Even with GPT-4, you told it, hey, do these 14 things, and it did like 9. And then you were like, what happened to these other 5? It's like, ah, true, you're right. Now the prompt adherence is so good that that's not the case anymore. That was the tipping point that made this entire agentic revolution that we're seeing right now. We have Claude Code kind of catching up with OpenClaw, with all the madness that OpenClaw unleashed, with Claude Cowork maturing up as this like new category of consumer-friendly agents. It's incredible stuff. We covered this weekly on the channel, right? Every Friday, Newsyki News, we cover this stuff. But it's really the model that made it work. So now when we look at this, well, this model is so far ahead of Opus 4.6. Like from a consumer perspective, I just asked myself like, If right now you see massive unlocks for content creators when it comes to using OpenClaw, and maybe it's a spectrum where on the other end of the business, you know, I said it jokingly with like the plumbers, but at the other end of the spectrum, you have all these businesses that don't get affected so radically, right? Because they're not software-based, they're not trend-based, they're not digital. You can't scan the internet for new stories and have that repurposed into scripts and create visuals and all these things, right? Heck, you're like plumbing, like you're not gonna be able to use software to help with that. But if that's a spectrum, which I think it is, you know, with, you know, finance is probably somewhere in the middle, maybe more towards the software side. Like as these models get better, it's gonna start moving along this spectrum further and further where it's just a no-brainer to use these agentic tools for the different areas. And like now I would want you to think about like, okay, what industry do I work in? And how is that gonna affect me? Right? If you're a coder, heck, you're all the way out here. You're even beyond content creators. Like I think coding is the one first big use case where it just like, you know, cataclysmic shift in how it's done, right? Over the past 3 years, for sure. Whereas all the physical businesses, they're on the other end of the spectrum. So it's not going to be clear. It's not going to be black and white. But think about where you are on that spectrum. And then think about that if this used to be the spectrum and agentic tools, well, they maybe only worked for like coding and a few other niche applications over here. Now it's kind of like moved over here with Opus 4.5 and 4.6 now. But then we What does this move look like? If this is the spectrum, like it just moved here, is it gonna move here? Or is it gonna move like here? I don't know. We're yet to find out. So that's the first thing that I wanted to bring up. Secondly, I mean, just look at these numbers for a second. Software engineering bench, it jumped from 80.8 on Opus, right? Versus Gemini 3.1 Pro, which was 80.6, to 93.9. Excusez-moi. What? What the— okay, that's one. What about MMLU? Okay, improvement, not that dramatic. All these software benchmarks are insane. OS World increase, usually it's MMLU, that's like general world understanding, and software engineering that you're kind of like first looking at. But just this jump in software engineering is unbelievable. Multimodal, right, where it's using various inputs, not just text, 27.1 to 50. 59%. You could imagine what this can do if you just give it a screenshot. The understanding of it is going to be so much deeper. Okay, this is me just kind of like talking about these numbers. You know, you can look at this, they're impressive. But furthermore, as I told you, like, they're not putting this out. They're basically giving it to the biggest companies to harden their security. And there's a few interesting kind of findings here. One of them was that it found security vulnerabilities in a 17-year-old operating system, a 27-year-old vulnerability in OpenBSD, which is a reputation as one of the most security-hardened operating systems in the world. Okay, I wasn't familiar with this one, but basically ran that. And somewhere else I saw that it just cost $50 in token to find that. I think that is not verified, but unbelievable. Also, in another source, I saw that they used it to expose some vulnerabilities in Firefox. The previous Opus model found 2. This model found 108. 81. So, you know, a real change. So what they're doing is they're giving it to all these companies. And here's my second insight that I wanted to tell you about. The second insight is this was, and I don't think it's a coincidence. You know how OpenAI kept holding their models over the past years? Like, you know, no announcement for a few weeks and then Google comes out with something and haha, all of a sudden, who would have thought OpenAI is announcing their next big thing a few hours before the Google keynote, right? They kept doing that throughout the past few years. Well, Anthropic's timing is a bit odd here, no? Look at that. We have a new open source model, which was basically the new king in open source coding. Okay, GLM 5.1. This came out roughly, what is that, like 9 hours ago? Well, 2 hours later, we get this Mythos announcement. Okay, now this open source model scores 54.9 on SWE Bench Pro. Opus 4.6 scored 57.5. So I guess on this graph, Let me have a look on this graph. This was SWE Pro here. So it would be right up to par with the current best-in-class models, the paid models, right? But the difference is this open source model basically is available under a fully open Apache 2.0 license, meaning you can like do everything with it. And it's just yours. And they just released it out there. So this came out and basically what that meant, you know, hey, Anthropic, your best model, not that unique anymore. People are paying hundreds of dollars for it daily. How about here's an open source version of that that people can run for free on their beefed up Macs? I told you I have this new MacBook Pro, which can run this 128 gigabytes of RAM. This model might be slow, but I would be able to run this on here. And we're talking Claude Opus 4.6 level performance when it comes to development, right? So they came out with this and immediately Claude had to defend their business model, so to say, with this press release, right? They sent out, I got this release with different partner quotes of, I'm not going to read all of them, but you can see all the big players, Microsoft, Linux Foundation, Google, all of them in here, all of them claiming how helpful it's to have a model like this in their development. And you know, not to yap on forever about how big of a deal this is. But with everything moving forward so fast, I would kind of like end on this point and wrap it up this way. I would really say that it's not even the fact that this model will be available. It's rumored to be available, I think on PolyMarkets, over 50% of the bets are being available about before the end of April. It's not even the fact that we'll have access to this. It's going to be incredible. I think everything that we have, you know, Claude Code, OpenClaw, even like Claude the web app, even more people are going to be switching now. All of those are going to become better and some of them significant. On that spectrum of different jobs that are affected, like it's going to move, sure. What really makes me think is the second and third order consequences. What does that mean, Igor? It means that all the competitors are going to have to follow now. Just like I think Anthropic had to follow this open source release here, OpenAI and Google are going to have to follow this Claude Mythos announcement with their own products, right? So everything is accelerating. And you know, that's nothing new. It's been accelerating over the past few years. But this jump is so significant that it's going to force the competition if they want to stay relevant, if they want to stay in the race, especially OpenAI, quite ruthless with, you know, how they're releasing things. They're just going to have to keep up and they're going to ship their own version of OpenClaw, you know, like seriously, a lot of people over the past few days, as if you're not familiar, in OpenClaw, you could use Claude and not pay for it, right? Otherwise, you were paying like hundreds $200 a day for the API, if you like paid for the API credits, or you could just pay $200 per month and use the authentication with the subscription, the OAuth. Well, they banned that now on Sunday, if you haven't heard that. And now people are kind of like fleeing it because it's so expensive right now. They have to pay like $2,000, $3,000 a month for their OpenCloth running what used to cost $200. And people are trying alternative models. And everybody that I know that tried like the OpenAI models is like, if you have a different opinion, please leave a comment. I mean, this is like friends talking about this stuff. That's what, that's the way I kind of see this conversation. And everybody that I know, all like other people that I know, like, you know, they don't like the OpenAI models in OpenClaw. It's like it has no personality. Coding is not as robust, like it's robust, but it just feels off. And it doesn't work as well. Doesn't work as reliably. Tried it briefly. Not a big fan. I love Opus. And the big part of it is maybe the psychology for me. I just love the way it feels, the way it just gets things done, the way it like adheres to the prompts and retains the memory. It's hard to put into words. It's not even about the terminology of like the context window, it's more about how it handles the context window. Okay, so second order consequences, all the competition is gonna like now flood in. Third order consequences is really because the competition is gonna flood in and because the coding is gonna become so damn good and because there's gonna be so many new vulnerabilities exposed and everything. Well, these OpenClaws running a model like this, you're gonna unlock a whole new world of madness because people, you know, with unlimited budgets all of a sudden already, I know some people who are spending over $10K a month on their OpenClaw bill just on Opus API credits now because they just found so many ways to use it and and they just run it every day. This is gonna like 10x that. You're gonna have people running up $100,000 per month bills, people who can afford it, right? These are the further order consequences. The people running $10,000 per month are laughing their asses off because they're like, because it's so valuable to them, right? They would, in terms of human resources, they would be paying $50K, $100K, maybe more for what OpenClaw does for $10K. That's why they're paying that. They're not stupid. They're not just like, oh, I'm spending money, whatever. It's so valuable. I don't know. I personally spent like, what, like $2.5K over the last month on it. And it's for me like the value that it creates is so immense. This is going to be even larger, right? And the people spending $100K a month all of a sudden on AI, like this is the tipping point. This is really that future turning into reality where the people spending $100K, they're going to be getting way more value. They're going to be getting hundreds of thousands of dollars of value, meaning their AI system is going to create hundreds of thousands, maybe millions of dollars of value. And I'm talking about individuals. Heck, we saw the first story of the first billion-dollar company powered by AI. That's different. You know, it was— okay, let's not go there. But like, there was a story this week. Point is this, people using it and the people not using it, that gap is going to open so much more with this release. It's already happening. You see it in the world. You know, people buying Mac Minis going crazy, doing some crazy Blackmagic type of stuff. This is only going to widen that gap. And I think it's now, it's more important than ever to keep up with this stuff. You don't have to be using it, but just try to wrap your head around what's possible. Try to listen to people who are actually using it. I'll do more content on this type of stuff. I have to be honest, as I wrap this video up, I had a hard time creating content around OpenClaw because my goal is to like serve this, a non-technical audience of this like 80% of the people, show you what's possible, make your life better, get you time back with the use of AI tools so you can do more of what you care about, which, you know, we're humans. So we're good at, you know, connecting and building relationships and being creative and all that. That's the whole idea of the channel. That's the whole idea of the AI Advantage. And I kind of struggled fitting OpenClaw in there. But I think with this happening, happening. And with a new framework that I built, it's basically this like AI learning roadmap. Everybody starts in a certain place. There's a middle level, the context layer that most people miss. And then there's this agentic level. And you need to go through the context level to get to the agentic level. Otherwise, it doesn't work. It makes no sense. There's way more to that. That's kind of the cliff notes of it. But basically, since I put that together over the past few weeks, I feel way more confident in bringing you content like this because we can just contextualize this and we can say, okay, this is you know, level 3, agentic level stuff. And if you're not there yet, that's fine. But we can still talk about it in a way that is helpful. And now that that is in place, I feel like we can create more content around that. And also finally, and you know, I didn't create this video to point you there. But if you're interested in this 3-level framework, we're putting on a summit end of this month. It's completely free, not kind of free, completely free, April 22nd through April 24th, AI Advantage Summit, together with Tony Robbins, Dean Graziosi, myself as the head of education and a partner, and many more speakers. It's going to be great. In my parts, I'm going to be speaking about this progression, showing use cases, also teaching. A lot of this is going to be high-level stuff of like understanding where this space is at. Anyway, you can sign up to that. I'll put a link in the description. It's completely free. But that's, by the way, that's if you're interested in this progression. If you're interested in what's happening with these tools, if you're interested in my opinion once I get my hands on this, if you're interested on, you know, benchmarks, but also real-world use cases of this for non-technical users, then subscribe to the channel. That's what I do here. And I hope that you feel the urgency around this and that you start using these tools in a way where you have rich context. And where you're not just using a variety of applications, but you also have a depth of usage in one application. Because that's what allows these types of systems to thrive. That's what makes agents work. That's what's going to empower Mythos once it's available. And yeah, it's just unbelievable that we're seeing this type of acceleration in not just our lifetime, but in like a sequence of months and years, heck, weeks. It's crazy. Great to be on this journey with you. My name is Igor Pogany, and I hope you have a wonderful day.