Claude Mythos Preview Explained: Anthropic's Project Glasswing
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I explain Anthropic's Project Glasswing and the model inside it, Claude Mythos Preview. Anthropic is keeping this model private and using it to find security holes in the software we all run. I cover what was announced and who's involved. I also explain why I think the bigger story is what Mythos tells us about the gap between Anthropic's internal models and the Claude we can actually use.
I recorded this in April 2026, a few hours after the announcement. AI tools and access rules change fast, so check Anthropic's current docs and announcements for the latest details.
What I cover in this video
- 0:00 Why I think this could be the most important AI announcement of 2026
- 1:32 A model that hacks better than human experts, and why Anthropic chose not to release it
- 3:03 The partner companies behind Project Glasswing
- 3:48 The $100 million in usage credits and the thousands of vulnerabilities found
- 4:34 The real story: what this shows about Anthropic's capabilities
- 6:04 What finding these vulnerabilities says about how well the model understands code
- 7:35 What we still don't know about Mythos
- 11:22 Why the gap between internal and public models is growing, and what builders should do
What Project Glasswing and Claude Mythos Preview are
Project Glasswing is an Anthropic initiative, and Claude Mythos Preview is the model at its center. Mythos can find and exploit software vulnerabilities better than the best human security researchers. Anthropic also shared benchmark scores, and Mythos comes out well ahead of Opus. I didn't include the numbers here, so read the source for those.
You and I can't use it. It isn't on the API, you can't pay for access, and there's no waitlist. The reason is simple: a model that can hack at a superhuman level could do real damage in the wrong hands.
Anthropic didn't charge for access. They brought in a group of major tech and security companies and are using the model to patch the internet before attackers build something similar. The partners I list in the video are:
- AWS, Apple, Google, Microsoft, and NVIDIA
- Broadcom, Cisco, CrowdStrike, and Palo Alto Networks
- JPMorgan and the Linux Foundation
Anthropic committed more than $100 million in usage credits to the project. It also put $4 million directly into open source security organizations and gave 40+ more organizations access to scan their own systems with the model.
The reason this matters: Mythos Preview has already found thousands of high-severity vulnerabilities in every major operating system and every major web browser. No human security team had found these. I see Mythos as a security layer, not something builders will use day to day to make apps.
The real story: the public Claude isn't the best Claude
The security work is important, but I think the more interesting part is what this announcement shows about Anthropic. Claude Sonnet and Claude Opus are the models Anthropic considers safe and ready for millions of people. Mythos is one they built and decided not to release.
So Anthropic has models in its lab that are much more capable than anything we can use. I'm not guessing at that. It's what they announced. The public Claude has never been their best model, but this is the first time we've seen a clear sign of how big that gap is.
Think about what this kind of work requires. Finding and exploiting a vulnerability in a major operating system takes more than being good at code. You have to understand how systems interact, how memory allocation works, and how compilers process instructions. You also have to understand how decades of legacy code create edge cases the original developers never tested. Very few people ever understand software at that level, and a model in Anthropic's lab already does.
That makes me wonder what the next public Claude will look like. Opus 4.6 came out about two and a half months before I recorded this, and GPT-5.4 had been out for about two months. Mythos is here and Spud is on the horizon. My view is that the next generation will be a big leap, closer to the GPT-4 or Opus 4 moments than a small update.
What we don't know yet
A lot of details aren't public, and I don't want to oversell this. We don't know:
- How Mythos works technically, including its architecture
- What it costs to run, which could matter a lot
- How it fails and what it's bad at
- When or whether these capabilities will reach the public. Some might come out in a limited form, and some might stay restricted for good.
- Which of the vulnerabilities have been patched, which are in progress, and which are long-term projects
Anthropic is being careful about what it shares, and given what the model can do, that's the right call. Holding back details is normal in responsible security disclosure.
What we do know is that Anthropic is now more than a chatbot company. It's helping build security for the internet, and its partners committed real money to be part of it. That gives the announcement a kind of credibility a typical press release doesn't have.
What this means for builders
Here's how I think about it now. Each public Claude release isn't the newest thing Anthropic has. It's the model they decided was ready. Above it, there's always a more capable model they're still evaluating, red-teaming, and making safe.
I think that gap is getting wider because capabilities are improving faster than the safety work can clear models for release. The public releases later this year and in 2027 will be built on internal models that are already beyond what most of us picture. The jumps between releases should feel bigger than they used to.
For anyone shipping products, writing code, or running a business with AI, that's good news. People who build AI into their daily workflow now, whether with Claude Code or with AI agents, will have a head start when those stronger models arrive. The gap between people building with AI and people watching from the sidelines keeps growing.
Key takeaways
- Claude Mythos Preview is an unreleased Anthropic model that finds and exploits software vulnerabilities better than top human experts. It isn't on the API and has no waitlist.
- Through Project Glasswing, Anthropic and partners like AWS, Apple, Google, Microsoft, and NVIDIA are using it for defense. The project includes $100M+ in usage credits and $4M for open source security organizations.
- The model has already found thousands of high-severity vulnerabilities in every major operating system and browser.
- The bigger lesson is that Anthropic's internal models are well ahead of the public Claude, and that gap is likely growing.
- For builders, the plan stays the same: keep building and shipping with today's tools so you're ready when stronger public models come out.
FAQ
Can I use Claude Mythos Preview?
No. Mythos Preview isn't on the API, you can't buy access, and there's no waitlist. Only Project Glasswing partners and selected organizations can use it, and only for security work.
What is Project Glasswing?
It's Anthropic's effort to use Claude Mythos Preview for defense. Partners include AWS, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, and the Linux Foundation, and they're working to find and patch vulnerabilities before attackers do. Anthropic committed more than $100 million in usage credits and $4 million to open source security organizations.
Why isn't Anthropic releasing Mythos to the public?
Mythos can find and exploit software vulnerabilities better than the best human security researchers, so it could cause serious harm if attackers got it. Anthropic is using it to patch systems before anyone else builds something similar.
Will Mythos capabilities ever come to public Claude models?
We don't know yet. Anthropic hasn't said when or whether these capabilities will be released. Some may come out in a limited form, and some may stay restricted permanently.
Chapters
- 0:00The announcement most people are sleeping on
- 1:00What is Project Glasswing?
- 2:30The partners and the $100M commitment
- 3:30Thousands of vulnerabilities found in every OS and browser
- 5:00The real story: what this reveals about Anthropic's capabilities
- 7:00What superhuman hacking actually means for code understanding
- 8:30Being honest about what we don't know
- 9:30The gap between internal and public models
- 11:00What this means for builders and the future of AI
Transcript
Read the full transcript
0:00Anthropic just dropped a bomb on Mythos, and this is everything that we know so far in addition to the Project Glass Wing. So, this was just released today. There is a Mythos preview, and we're going to be breaking it down on who it affects and how it will affect builders going forward. So, Anthropic just dropped something just about an hour ago, and people are completely unaware of what's going on. I genuinely think that this might be the most important AI announcement of '26. Not because of what it does for you and me right now, but because of what it reveals about what Anthropic actually has sitting in their labs right now. I'm going to break this down, and if you saw the screenshot or the YouTube thumbnail, I'm going to post
0:45some of the benchmarks down in the link below, and you could check out these benchmark scores. It is absolutely insane on how Mythos scores compared to Opus. Completely blows it out of the water. So, basically, earlier today, Anthropic announced something called Project Glass Wing, and buried inside that announcement is a model they're calling Claude Mythos preview. This model is not available to you or I, unfortunately. It's not available to anybody um that is not in like the cybersecurity realm. You know, it's not on the API. You can't buy access to it. You can't get on any type of waitlist for it, either. But, there is a very specific reason for that because this is some very dangerous technology that if it was put into the wrong hands, it could probably cause a world model can find and exploit software
1:32vulnerabilities better than any hacker, the best hacker in the world, better than every human AI expert in security ever. And that's scary. So, if you let that sink in for a second, this could potentially, you know, cause a lot of problems if it was in the wrong hands. And Anthropic built an AI that can hack at an AI superhuman level, and their response was not to release it and to charge everyone for access. Their response was to partner with basically every major tech company in the world and use it to patch the internet before the bad guys figure out how to build the same thing. Good idea. So, that is not a small thing here. That is a fundamentally different kind of move, and I want to talk about why this
2:18matters so much more than most people are realizing. Project Glass Wing is essentially a massive coalition. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full OpenClaude course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Claude code or OpenClaude running on your machine. That's it. No fluff. I built this community because watching YouTube's
3:03only gets you so far. We launched it just 3 days ago, and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. The partners are AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JP Morgan, the Linux Foundation, Microsoft, Nvidia, Palo Alto Networks. And if you know anything about the tech industry, that list should stop you cold because that's basically every major player in security and infrastructure sitting at the same table for the same reason. They've committed over $100 million in usage credits to this project. On top of that, they put a $4 million
3:48directly into open source security organizations. And they gave 40 plus additional organizations access to scan their own systems using this model. I think that's pretty cool. Why are companies writing nine-figure checks for this? Because Claude Mythos preview has already found thousands of high-severity vulnerabilities in every major operating system, in every major browser, not just a handful, dozens, thousands, things that are sitting on your machine right now, possibly being exploited by someone somewhere that no human security team had found yet. That's the situation, and Anthropic is essentially saying, "We found the holes before the bad guys found all of them. Now, let's get them fixed." This is a massive undertaking,
4:34and it's genuinely an important work. I could see Mythos being more of a security buffer rather than something that builders use to build, like that security layer for Web3. You know, I'm not going to downplay the security story here. That's why we're talking about it. This matters, and this is the most important part. But, here's what I actually want to talk about because the security story is real, and it's worth mentioning, but it's not the whole story. The real story is what this announcement tells us about Anthropic's capabilities right now. Think about it this way. The models we use today, Claude Sonnet, Claude Opus, those are Anthropic's public release. Those are what they seem deemed safe for everyone. Capable enough and ready enough to put in front of millions of people. But, Mythos is not one of those. Mythos is something they built, immediately
5:19decided this one does not go public. That means Anthropic has models in their lab right now that are specifically and significantly more capable than anything you or I currently have access to. And that's not speculation. That's just the truth. That's not me reading between the lines. That's literally what they just announced. The public Claude is not the best Claude. It never has been. But, this is the first time we've gotten a concrete signal about exactly how big that gap is. Here's the part that should get your attention right here. To find and exploit a vulnerability in a major operating system, you cannot just be good at code. You have to understand how systems interact with each other at a deep level, how memory allocation works, how compilers process instructions, how
6:04decades of layered legacy code creates edge cases that the original developers never thought to test. And that security team three generations later doesn't even know it exists. That kind of understanding is extraordinarily rare in humans. Most engineers never get there, and Anthropic has the capability running in a model sitting in their lab right now, Mythos. Think about what that means for future Claude releases. It's been a while since we've seen a generation of releases, right? Opus 4.6 was about 2 and 1/2 months ago, and we've been on GPT 5.4 for at least 2 months. Things are cooking. We have Spud on the horizon. We have Mythos. I mean, things are getting crazy here, and they're going to lead to a point where this is
6:50not only going to just grow at the pace it's been growing, it we're going to have an exponential takeoff in this AI space. And if you're an AI builder, this should get you excited even more than you should be now. You know, we've all been impressed by the current generation. Like Opus 4.6, we're like, "Wow, this is AGI." I really think this next generation of drops for these models, whether it's Mythos and Spud, if that's what they're going with, is going to change everything. Like more than we've ever seen in the last 5 years in this AI space. The models are just genuinely good that we have now, yes. But, if there's a Mythos above them that can autonomously find zero-day vulnerabilities in Chrome, in Windows, in every major operating system, like what does that suggest about where the ceiling actually is? What does the next
7:35publicly available Claude look like if Mythos is what they're working with internally today? I think the answer is that we're in some kind of stage right here where we're going to see a huge significant leap. You know, not incremental, like almost like that ChatGPT-4 moment or that Opus 4 moment. I'm going to be straight about what I'm about to say next because, you know, I just want to be straight up with you, what I think. You know, I we don't know the full technical details about like how Mythos works exactly. We don't know like the specific architecture. We don't know the cost. That could be a big one. We don't know the failure modes. We don't know what it's bad at. Anthropic is being careful about what they reveal, and that's appropriate given what the model can do. We don't know the timeline for when or whether the version of these
8:21capabilities becomes publicly available, if at all. Maybe some of this gets released in more of a constrained form. Maybe certain capabilities stay restricted permanently. It's just genuinely we don't know. And we don't know the full picture of the vulnerabilities that were found. Like thousands of high-severity issues is a real number, but we don't have the breakdown of what's been patched, what's in progress, what's a long-term project. We just don't know. That's the nature of responsible disclosure in security work. But, here's what we do know, and that's part that's so important. Anthropic just positioned itself as something bigger than a chatbot company. They're not just building assistants anymore. We're way past that. They're building the security infrastructure of the internet, and
9:06they're doing it with partners who wrote serious checks to be part of it. That's the different kind of credibility, you know, that we just don't see. That's not a press release. That's not a blog post with cool vibes. That's we built something so powerful we won't release it publicly, so instead we're partnering with the entire tech industry to use it defensively before it gets used against everyone with credibility. That's how they're releasing this project, and that's a pretty bold move in my opinion. It's just a whole new category, right? So, for anyone who's been watching AI development closely, this should recalibrate something in how you're thinking about the trajectory. My honest mental model on this, like thinking about it clearly, is every time a new Claude release, you know, goes public,
9:51it's not the newest thing Anthropic has. It's the thing they decided that was ready. And there is always a layer of capability above the public release that they're still evaluating, still like red team, still figuring out how to make it safe before it goes out to millions of people. If you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. [music] I built something called content machine. It's 10 AI agents that run on the open claw orchestration, and they handle everything. Scripts, thumbnails, X [music] posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up [music] and the content's already done.
10:37I spend maybe 15-20 minutes reviewing and approving them and I move on with my day. It works for any niche, fitness, finance, [music] real estate, marketing, whatever you are building, and it is 100% completely customizable to your [music] use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing and it molds it to you, it learns how you talk and it writes [music] so it doesn't sound like AI slop. $97 one time, it's not a subscription. I'll put the link down below and you'll thank me later. That internal to external gap is probably getting wider over time, not narrower, because the capabilities are
11:22advancing faster than the safety work and clear them for release, which means the public Claude releases we're going to see at the end of the year and in 2027 are being built on the internal foundations that already exceed what we're currently imagining. And the jumps from one public release to the next are going to start feeling like bigger leaps than what we saw earlier in these early days. This is not me hyping AI for the sake of hype. I'm generally getting tired of that and, you know, this is a concrete announcement about a specific model with specific capabilities with specific partners who committed specific dollars, and that just points in one direction. The ceiling is higher than anyone is talking about publicly. This isn't practically like, what do you do with this? If you're building with AI right now, if you're using these tools to ship products, write code, run your
12:08business, this is amazing news. The tools are going to keep getting dramatically better, not just like a marginal GPT 5.3 to 5.4 update. This is going to change everything, and the people who build the habit of integrating AI into their actual work flow today will have a real head start when these more capable public releases arrive. The gap between people who are genuinely building with AI and the people who are watching from the sidelines is getting much bigger. Project Glasswing makes that undeniable. The capabilities that exist in these labs right now are extraordinary, and even the safe version, the public version, is already changing how things get built. What's the move here exactly? Keep building, keep shipping, keep
12:54figuring out where AI actually fits in your real life and your real work, because the ceiling is not where you think it is, not at all. If you think that Opus and Sonnet are good now, it's just getting started. Anthropic just gave us a peek at how high that ceiling actually goes, and it's a lot higher than most people are ready for. Like I said, I put the link to the source down below, read it, give it to your AI to have it summarized for you. Uh check out the benchmark scores. I mean, we're in exciting times. So, like I said, if you want to read the full announcement, just check it out below. You know, not not because it's loaded with technical jargon or details. It isn't, it's very easy to read, but because of what it actually reveals about where we are actually are as builders and people who want to build businesses with AI. So,
13:40project Glasswing, Claude Mythos, just dropped a couple hours ago, I think. This is one of the moments that we will look back on and say, that's when it became undeniable. Not when some vague AI thing happened. A specific model, a specific coalition, and a specific set of capabilities that are too powerful to release, but too important to ignore. If you want to learn to actually like build with AI and maybe you have an idea and you're like, well, I don't even know how to build this. That's what we do inside of our community. We just actually hit 211 people to join this community that I launched 23 days ago. And we cover all things open claw, AI agentic frameworks, security, setting up open claw. And we actually now do nine live bootcamps every single week. So, if you're, you
14:26know, working or you're busy, there's going to be a class you could jump into and get your question answered. So, I'll put that link down below. Hey, come join the party. And if you haven't subscribed to the channel already, what are you waiting for? We're covering all things AI news and open claw. We'll see you in the next video. Be blessed.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Claude Mythos Leak Explained: Anthropic's Capybara Tier Above Opus
Claude Mythos is the first model in Capybara, a new Anthropic tier above Opus. It became public through a misconfigured data store. Anthropic confirmed to Fortune that training is done, early access customers are testing it, and it scores dramatically higher than Opus 4.6. It will cost more than Opus, and there's no public release date.
Claude Fable 5 for Builders: Pricing, Benchmarks, and Model Routing
Claude Fable 5 is Anthropic's strongest widely available model. It has a 1M-token context window and costs $10/$50 per million input/output tokens. I'd use it for hard work like migrations, deep debugging, and code review, and send everyday tasks to Opus 4.8, Sonnet, or Haiku.
Claude Fable 5.2 Explained: Pricing, Benchmarks, and Effort Levels
Anthropic's Claude Fable 5.2 costs $8 in and $40 out per million tokens, down from $10 and $50. At High effort it roughly matches Opus 5.5 at Max on FrontierCode for about half the cost per task. I'm keeping Opus 5.5 as my daily driver and using Fable for big, hard projects.
