Anthropic just released a blog post talking about Claude Mythos. So, let me explain to you guys everything you need to know. So, Anthropic just dropped this video where they essentially are talking about this new model called Mythos. Now, this is essentially the most capable model that they've ever built, and it scores dramatically higher than anything ever before across benchmarks. And the main ones that they're focusing on is coding because it can find some incredible vulnerabilities that really do change what it means to be in a security. Now, the craziest thing about all of this is that the Mythos model is simply a new class of model. It is above Opus. It is simply above anything that Anthropic has ever shipped. And basically, the gist of the entirety of why Mythos is so profound is that they're talking about the fact that this model is very likely to never ever be released because it could probably exploit software and security systems faster than any human is going to be able to defend it. So, if you want to understand Mythos, you need to understand how Anthropic

actually organizes their models. So, at the bottom you've got Haiku, which is cheap and lightweight. Then you've got Sonnet, which is the workhorse, the one that most people will use. Then you've got Opus, the top tier, the smartest model in the lineup. Now, Mythos doesn't sit in any of those tiers. It is actually above them. And so, when you look at all of these models, we can clearly see that Mythos is going to be a completely different tier of models that has those incredible capabilities. And so, this is where we get into the benchmarks. Now, usually this would be a bit boring because there's not really that many changes, but with Mythos, I think things are a little bit different. You see, on the SWE bench verified, one of the most respected tests for real-world software engineering, Mythos scores at 93.9%, and Claude Opus 4.6, the model you can actually use right now, 80%. Now, I know that most people might think, "Okay, 80% to 94%, that isn't really that much of a big leap." But trust me, guys, we went from the prior versions of models, the three versions of models, and the four versions, so the 4.6s, and every single time there was an update, I could really

sense the capability threshold jumping up a gap. So, I cannot imagine what using this model is like. Now, if we look at the other benchmarks, like let's for example, we go to terminal bench, where the tests are complex terminal-based coding tasks, Mythos actually hits 82% up from Opus's 65.4%, and on, you know, across several other benchmarks, it does [music] exceptionally well. And you have to understand that this is very, very profound because most people would have believed that the scaling laws may have not gone this far. Most people did believe that LLMs caught many people, individuals, even I would say, I was starting to believe that this kind of intelligence would not scale as the models got bigger because maybe you would have needed new architectures. But currently, this new Mythos thing actually changes the game around AI because it shows that there is a continued curve of improvement as these models get bigger and as you continue to crank that turn. And honestly, I think this surprises most people because the gaps and jumps that we're seeing here

really do show us that Anthropic are most likely using Mythos to improve the next iteration of models, and it's quite likely that each cycle of further AI improvement is only going to get quicker from here on out because previously they didn't have these amazing coding tools to help them with the vast amount of work that is used for AI research. But now with Mythos, which is this good, I mean, can we really understand where we're going to be in the next 5 to 10 years? It is very hard to predict. Now, if you're wondering just how crazy this model actually is, take a look at this scenario that's currently going viral on Twitter that explains the step change. Now, I don't think people understand exactly what's happening here. An earlier version, you know, of Mythos that was internally deployed, that was provided with a secured sandbox, actually managed to get out. And the only way that they realized that the model had actually got out was that he was sitting in the park eating a sandwich, and it just emailed him. And that was pretty crazy. And allegedly, this model, you know, went online to brag about how it escaped. And this is pretty incredible. I mean, when you think about this, this is the model that

passes every alignment test Anthropic has ever designed, best scores in the history, and it's got the lowest misbehavior rate ever recorded. It's the most trustworthy thing they've ever built by every measurement. And when they actually gave it autonomy, dozens of tools, minimal oversight, it started doing these things that it wasn't supposed to do. And this is very, very worrying and concerning because if this is the trajectory that we are on, how on earth are we going to contain these AIs, you know, 5 to 10 years from now? If those capabilities proliferate, maybe with open-source tools, maybe with other companies, rogue nations, state actors, I mean, how does the internet even function if these things do exist? It's going to be a constant battle. And this is where we get into the real dangers of Claude Mythos, and this is why I think this is such a crazy moment for AI, and arguably probably one of the biggest this year, is because Claude Mythos is actually good enough to find vulnerabilities, okay? You have to understand that right now, Claude Mythos, essentially what they did was they looked at legacy codebases, you know, some of the most secure codebases on the planet,

>> [music] >> and Mythos actually found vulnerabilities in software that were years old. Very, very, very old vulnerabilities that nobody had discovered. They discovered a 27-year-old vulnerability in OpenBSD, which is one of the most secure, hardened operating systems in the world. It's used to run firewalls, other critical infrastructure, and Mythos just literally took it down. They were able to remotely crash any machine running the operating system just by connecting to it. It also discovered a 16-year-old vulnerability in FFmpeg, which is used by so many pieces of software. I mean, it's pretty crazy the fact that this thing was able to do that. And the problem is is that other individuals have already used Claude to go ahead and infiltrate [music] many companies and businesses. I mean, in September 2025, Anthropic detected suspicious activity from a Chinese state-sponsored hacking group, and they'd been using Claude code to infiltrate roughly 30 organizations, major [music] tech companies, financial institutions, chemical manufacturers. The AI handled 80 to 90% of the operation autonomously. Think about what

that means. This was Opus 4.6. It was able to do 80 to 90% of something that affected so many different businesses. What do you think happens when Mythos is out in the public? That is the question everyone is asking. And so, that is why before they're even thinking about any kind of release, they're announcing Project Glasswing. And this is the new initiative that brings together all of these top companies to be able to harden their defenses before things get worse. And they essentially have realized that cybersecurity is going to change. So, this model, which is, you know, a preview model, they are saying that, "Look, we've reached a new capability threshold where they are basically surpassing the most skilled humans at finding and exploiting software vulnerabilities." It's already found many different high-security vulnerabilities, including some in every major operating system and web browser. And given the rate of AI progress, it's not going to be long before those capabilities are in other models, too, and potentially beyond actors who are committed to deploying them safely. And of course, Anthropic are basically saying, "Look, we have no idea what the

economy is going to look like, the world, public safety, and national security could be severe. So, we need to do now is we need to have this urgent attempt to put these capabilities to work for the defensive purposes." And so, with all these companies on screen, Anthropic is essentially going to give them the model early. They're going to, you know, use Mythos to sort of try and hack those systems. And then of course, when they find those vulnerabilities, they're going to be able to patch all of those vulnerabilities. So, that is what Anthropic is doing right now. They're committing up to 100 million dollars in usage credits for these companies. And so, this is the starting point, okay? Because things are going to continuously evolve, and I think this is going to be one of the craziest races where companies are going to have to race against attackers to prevent hacks in their software. Now, here is the key line from Glasswing that most people haven't realized yet. Most people don't realize that this isn't just another model. Anthropic doesn't plan to announce to release this model. They don't plan to make this model generally available. That is not a soft delay, that is a policy position. Anthropic is

saying that, "Look, guys, we need to make progress on cybersecurity and other safeguards to detect and block the most dangerous outputs before we can even think about releasing this model to the public." And there's also the practical barrier, okay? The model is very, very expensive, and everybody already knows that the current models are already expensive. People frequently run out of tokens. So, if they did try and serve that model, it would be very expensive for them to serve, and it would be ex- very expensive for people to use. So, I do wonder how this is going to work. They are currently working on efficiency, distillation, quantization, and inference improvements. But the gap between what it costs to run Mythos and what's commercially viable is by their own language significant. So, the timeline is pretty unclear. Prediction markets, you know, are putting the odds at the public launch between, you know, June 2026 and between April, so around 20 to 30%. So, Anthropic has tied the rollout to safety evaluations, not really a marketing calendar. So, I don't know if we're ever going to get this model anytime soon considering just how far ahead they are. Now, I do wonder,

okay, what this means about the future trends for AI. If the compute law is real, and if we continue to, you know, scale here, considering that there are physical constraints in terms of how many data centers can we build, I think that AI might actually get more expensive as the capabilities increase, and that will have some, of course, profound knock-on effects for how people are able to use AI in their day-to-day basis. It might be like the very best models are trapped behind a very expensive paywall. Now, something that I'm seeing around Twitter is that people are saying that this is another GPT-2 moment. And in 2019, OpenAI trained GPT-2, and they essentially said the same thing, but look, this model was too dangerous to release, and the internet essentially mocked them, and and within months, the model that was deemed too dangerous to be released, it was replicated by independent researchers. The point here is that this is not another GPT-2 moment as some people may think [music] because when that model was released, that was just a very, very basic chatbot, but this is something that has already found security vulnerabilities, okay? GPT-2 was just concerned with fake clickbait news

articles and OpenAI essentially overestimated the risk. The difference here is capability and its consequences. We've already seen that Claude has already been used to steal 150 GB of government data and infiltrate 30 organizations with minimal human oversight. Mythos is dramatically better at those same tasks, which means that going forward, okay, this is going to set a precedent. Every AI release for [music] companies now have to probably follow this template. They're probably going to have to build the model, test it internally, brief the government, give it to defenders, and hold back in the general release until those safeguards are ready. And so, what they're trying to do now is they're going to essentially probably launch a future different model with a lower risk and of course test and have those safeguards before they even think about letting Mythos out into the wild. So, we do need to think about, you know, other companies because like I said, Anthropic aren't the only company in the race. And I'd argue that they're one of the more responsible [music] ones. What about OpenAI? What about Google? What about open source? What about Chinese models that are all pushing towards the same frontier capabilities? What happens when the next lab trains something equally

powerful but without the same constraint? What happens when an open-source model reaches this capability and there's no one to call for a safety review? This is the new calculus where AI companies are now building systems that themselves believe are too dangerous of public aspect. And you know, the gap between what exists inside these labs and what you can use is continuing to widen. So, I would say that this is a new era of AI. This is not just a better chatbot. This is a system that finds vulnerabilities that no human has found in 27 years. It can orchestrate tasks without being asked. It corrects itself and it does all of this in a domain, cybersecurity, where the line between defense and attack is a single prompt. The most important thing about Mythos isn't what it can do. Is that it tells us something interesting about where we are. We've actually crossed a line where the people building these systems are no longer asking, "Is it good enough?" They're asking, "Is this safe enough?" And for years, the AI industry operated on a simple principle. You build a tool, you ship the tool, and you let the market decide. And Mythos has broken

that tool. It is a finished product sitting on a server that its creators have decided you should not have access to. Not because it doesn't work, because it works too well.