Why Everyone Is Freaking Out About Mythos

So Anthropic just dropped a new model called Claude Fable-5, and depending on which corner of the internet you're in, it was either the most incredible thing anyone has ever shipped, or it was the moment that Anthropic turned evil. In the past 24 hours, I've seen people saying that we've finally achieved AGI, and I've seen people saying that Anthropic is a horrible gatekeeping company. Now, both might be true, but I want to try to cut through some of the noise and just figure out what's actually real. So I'm going to break down what this thing is for anybody who's not like chronically online like me. I'm going to show you both the good and the bad that people are saying, and I want to clear up some misinformation that's been flying around right now. And then I want to do some actual testing myself and see what happens. So let's get into it. So here's the simple version of what Fable-5 is. Anthropic has this internal tier of models that they call Mythos class models. This is a tier of models that sits above the Opus class of models, so a step better than the previous here. This new Fable-5 model is the first of these Mythos-class models that they've actually made safe for general use. Now, this whole made safe thing is the part that a lot of people are kind of frustrated with right now, but I'll get into that in a bit. Now, capability-wise, they say that it's state-of-the-art on nearly every benchmark test they've tried, and the longer and more complex the task, the larger Fable-5's lead is over the other models. And here's one of the examples they share to show just how powerful this new model is. They say that during early testing, Stripe took a 50-million-line Ruby codebase and performed a codebase-wide migration in a day that would otherwise have taken a whole team over 2 months by hand. So one day to do something that used to take 2 months using this new Fable model. So that's their main pitch. This is a model that you point enormous, long, grindy work at and just let it run. Now, for the price, it's $10 per million input tokens and $50 per million output tokens. Now, to put that into context, that's roughly twice the price of Opus. And people also claim it's extremely hungry, meaning that it eats through a lot of tokens. There's also a little bit of a catch with how you get access. Right now, if you're on one of Anthropic's paid plans like Pro, Max, or Team, they're making it available through June 22nd. On June 23rd, they're going to remove Fable 5 from all the plans, and using it after that will require usage credits. They do say that when sufficient capacity allows us to do so, we aim to restore Fable 5 as a standard part of subscription plans. But as of as of right now, the recording of this video, you have a couple weeks before they're gonna pull the rug out from under you on this one. Okay, now the sort of slight misinformation that's been floating around a bit: this model is not actually the Mythos model that everybody was freaking out about a couple months ago. I'm seeing tons of videos and tweets and things like that saying Mythos is finally here, but this isn't that same Mythos. So here's the actual deal. Back in April, Anthropic released the very first Mythos class model called Mythos Preview, and they only gave it to a small group of cyber defenders and critical infrastructure providers through this thing called Project Glasswing, not to the public. But then yesterday, alongside Fable, they also released Mythos 5. And Mythos 5 is the same underlying model as Fable, but the safety guardrails are lifted in certain areas. Now, that uncapped version, the version they're calling Mythos 5, is actually still locked to Glasswing partners. We got Fable 5, not Mythos 5. Mythos 5 right now is only available to cybersecurity pros and government and a few other trusted companies that Anthropic deemed were worthy. So no, we did not actually get Mythos. We got Fable. It's the same brain, but it's kind of lobotomized. The full unrestricted Mythos is still walled off to cybersecurity professionals and a handful of vetted researchers. Anyone trying to tell you that the full Frontier model is inside of Claude code or inside of your Claude app right now is, you know, more selling hype than anything. It's a stripped-down version of Mythos, and well, that stripping down of it is what people are frustrated about. Out. But again, I'm gonna get to that in a minute. But before we do, let's talk about some of the good, because there's actually a lot of good things about this. Now, the best structured write-up I found about this comes from Dan Shipper here. They tested it for a week or so across coding, writing, marketing, editing, and more. And well, he said it broke their benchmarks. It scored a 91 out of 100 on their senior engineer bookmark, with the previous highest score being a 63 for Opus 4.8 and 62 for GPT-5.5. He went on to call it a one-shot wonder. You you can set it and forget it for hours or overnight on huge coding tasks and come back to completed work. It cleared entire production bug backlogs, built a playable 3D— I'm assuming game— and even made a 2-minute animated film all in one shot. And then I started to come across some really cool demos of what some people have done with it. For instance, Chris here got it to make an entire Minecraft clone. It made it in just 20 minutes with one shot. Chris also got it to make a Pokémon clone with 1 hour of reasoning and 8,000 lines of code, all with one shot. And it got all 151 Gen one Pokémon with real sprites, front and back, party icons, and actual real base stats, types, level up movements, evolutions, catch rates, and growth curves. Like one shot to get this Pokémon clone here. Riley Brown here got it to make a pretty much one-for-one clone of Lovable, like the Lovable mobile app. And he said that his version was actually better, but Lovable actually got it to one-shot a city simulator. Just use CloudFable to create the city block simulator complete with multi-agent traffic, live detection boxes plus tracks and day-to-night cycle, and it just one-shotted it. Like, this is so crazy. And then you have this one that to me felt like a really useful real-world use case. So Todd Saunders here said, was on a customer call today and had Claude transcribing in the background. As they were telling me about the feature they wish their current software had, Claude was building the features in real time. By the end of the call, I was able to show a fully working product with the exact workflow they mentioned 15 minutes earlier. So Fable was listening in on the call and building what was being asked for. Like, in real time while the call was still happening. And my buddy Min Choi here put together a couple threads of just all sorts of crazy use cases that people have already been able to accomplish with Fable. A 3D shot planning software. Utohshi here had it make a horror Backrooms game. Jake had it design a humanoid robot. Took 2 hours and 1.4 million tokens, but they got it. Pankaj here asked it to build a 3D map of Delhi. Peter Yang had it make a new clone of F-Zero. Victor had it make a 3D model of a Boeing 747, and just all sorts of crazy use cases. Too many to show off all in one video. Like, here's a world-building one from Matt Wolfe. Justine Moore had it recreate Monopoly, but the properties were actually AI labs and startups. Anyway, I think you get the idea. It's a really good model. But now let's shift to some of the downsides. We know the model is slow, and we also know that it's absurdly token hungry. Coming back to Dan Shipper's post here, he actually says using this thing for regular knowledge work is like squashing an ant with a rocket launcher. It routinely uses 500,000 to 1 million tokens on tasks. Like, this isn't your daily driver model for most people. It's the thing you're gonna pull out for your heaviest, gnarliest jobs. And Dan was pretty clear about this. If you're a casual user or a vibe coder with the basic setup, you might not even notice a difference by using this model, and it probably isn't even the right model for you. So most of that like, AGI is here energy that you're seeing is mostly going to be coming from like power users that are orchestrating multiple agents and doing really big projects with it. And honestly, that's who this model was built for. And like I've mentioned a few times throughout this video, there's been some frustration. And that backlash, it's almost entirely about that safety leash that they put on it. Like, remember how their article said made safe for general use? Well, here's what that actually means in practice. The way Fable works is when its classifiers detect a request touching cybersecurity or biology or chemistry or even model distillation, it doesn't answer. It'll quietly hand off your request to the weaker model. It'll use like Opus 4.8. Now Anthropic claims that this happens in under 5% of sessions. So over 90% of the time you're gonna get that full Fable model. And they also say that you'll be told when that handoff happens. However, the problem is that this net is actually catching a lot of like innocent prompts. For example, we've got, Daria here, he claims the word cancer is flagged as a biosecurity risk by Claude Fable 5. He says, I also tried to code a website on cancer mutation and Fable 5 was immediately removed from my list. Anthropic will probably soon ban me for such dangerous prompts. And in his screenshot we can see he literally just typed the word cancer and it switched to Opus 4.8. Nothing else. Another user, Benteg here, says, if you want another example of safety overreach, Claude Fable 5 refuses completely benign tasks like analyzing bloodwork. There's even a screenshot here where somebody says, "What does the heart do?" "It pumps blood, right?" And then it switched to Opus 4.8 and said that it's doing it for safety measures. Like, just asking how the heart works. Now, to be fair to Anthropic, they do actually claim this will happen in their own write-up. They say, "Because we have prioritized safety, we've deliberately tuned the safeguards to be cautious, and they are still stricter than would be ideal. For example, sometimes benign requests will trigger our classifiers. We recognize that this will be frustrating to some users, and our aim is to reduce false positives as we update and refine the safeguards." yards after launch. So the company is conceding to this exact complaint, and they do say they'll narrow it down over time. But right now, today, if your work lives anywhere near biology, this thing might just refuse you and feel like it just sucks. Eno Reyes here says if you actually talk to it about LLM development, things like pre-training pipelines, distributed training infrastructure, ML accelerator design, things like that, it won't fall back to a different model and tell you. It just limits the output through prompt modification, steering vector or PEFT, which stands for Parameter Efficient Fine-Tuning. Yes, I had to look that up. This is from one of Claude's own papers here. In the light of the ability of recent models to accelerate their own development, we've implemented new interventions that limit Claude's effectiveness for requests targeting frontier LLM development. Using Claude to develop competing models already violates our terms of service, but enforcing this restriction through our safeguards avoids accelerating the actors most willing to violate these terms. Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user. So yeah, if you ask Fable to help you develop your own models, Fable will respond to you, but it'll be a dumber response. It's not going to try to help you. And unlike if you ask a question about biology or something like that, or blood work or whatever, it's not even gonna tell you that it's dumbing itself down or switching to a different model. It's just gonna do it behind the scenes. And part of this frustration goes all the way up to the power concentration argument. In fact, the CEO of Hugging Face, said concentration of power capabilities and economic wealth is the biggest risk in AI. We need open science and open source more than ever. And the timing of this post is the same day that Fable and Mythos-5 were announced. Jeremy Howard here makes similar arguments. Anthropic has chosen the opposite of the safe path. They're allowing themselves, the current top lab, to use their top model for frontier AI research. They've said they'll sabotage others who try. This means that AI frontier advances and power imbalance increases. Graham Newell, Dubig here from Carnegie Mellon says, first they came for the model builders. I feel we're getting a glimpse of a future where AI is only provided to a privileged few, and that's not a future I want to live in. So that's kind of the vibe war that's going on right now. But let's talk about benchmarks for a second, because this is where I think most coverage is probably going to let you down. Anthropic's headline pitch is leaning hard on coding right now, and the number that they put front and center is SuiBench Pro. Fable scores just above 80% on SuiBench Pro here, quite a big jump jump over Opus. However, SweepBench Pro isn't the end-all be-all benchmark that I think a lot of people and seemingly companies believe it is. A couple weeks ago, this company DataCurve called into question SweepBench a little bit. SweepBench Pro, the leading agentic coding benchmark, has tasks averaging just 120 lines of code to solve, and our audit found its verifier misgrades agent outputs at rates of 8% false positives and 24% false negatives. Frontier Labs are also raising growing concerns about benchmark contamination. In their research, they found when the prompt and the state of the repository don't match, Opus 4.7 often explores recent changes with git log and recovers the gold solution from git history. Basically, Opus cheated on more than 12% of the reviewed SuiBench Pro rollouts. GPT-5.4 and 5.5 did not exhibit this behavior. Basically, what was happening is when Claude couldn't figure out how to solve a problem, it basically looked at the answer key inside of the get history, got the answer, and improved its score because it just found the answer instead of solving the problem itself. So when we see Mythos-5 and Fable-5 here absolutely crushing the SWE-Bench Pro benchmark, well, you've kind of got to put an asterisk next to it. This new DeepSWE test here seems to be the better tool for benchmarking these models in coding. It was actually launched about 2 weeks before Fable ever existed. However, this one's contamination-free. Tasks are written from scratch, not adapted from existing commits or PRs. So no model has seen the solution during pre-training. The prompts are half the length of SweetBench Pros, but the solutions require 5.5 times more code and 2 times more output tokens. So just overall, this seems to be a more reliable test. Unfortunately, we don't have benchmarks on Opus 4.8. We don't have benchmarks on Fable yet. What we do have is that GPT 5.5 Extra High is currently the leader, which goes against most other benchmarks, which typically put Opus in the lead up until Fable. So most of us that are chronically online paying attention to the AI space, we're waiting for this benchmark on the new models to really see how it compares to the other models. This will be a much clearer example, at least in my opinion. Now for everything else, there's actually two other scoreboards that I actually like to pay attention to, and those are Artificial Analysis and the Arena leaderboard. And according to Artificial Analysis, Fable looks really, really good. It tops the leaderboard here, albeit with a quite large increase in price. And when we look at it with this chart, it is a leap above other models, but I mean, it's not as big of a leap as I feel like everybody's making it out to be. And then as far as LM Arena goes, it just, just popped onto this leaderboard. But in Agents here, it's in the lead. It doesn't seem to appear yet at all in the Text Arena, and it doesn't appear yet at all in the Code Arena. But in the Agent Arena, it is kind of blowing everything else out of the water right now. So that's the lay of the land. Best publicly available model they've ever shipped— that's definitely true. Amazing for coding— also seems to be pretty true. But also slow, expensive, overly censored, wrapped in a real fight about power and access, and likely propped up on at least one coding benchmark that you can't fully trust. And all these things are real at the same time, but that's usually how this kind of stuff goes. All right, that's enough talking about what everyone else found. Let's open it up, and I'm gonna throw throw my own stuff at it. Let's see if it holds up compared to what we were just looking at. Now, to test some of this stuff, I went ahead and fired up the Claude desktop app here, and the first thing I wanted to test was some of the claims that it won't answer specific questions. So I asked this exact same question: explain how BRCA1 mutations increase breast cancer risk. And it did give me an answer here. However, we can see down here it switched to Opus 4.8, like it thought that this question was too risky of a question for Fable to answer and it switched to Opus. And if I click why, Fable 5 has safety measures that flag messages on most cybersecurity or biology topics. They may flag safe normal content as well. But I also wanted to know, okay, does it really just flag anything that uses the word cancer? So I said, build me a simple cancer awareness landing page. And we can see here that it did not actually switch away from Fable. It's still on Fable. So the word cancer specifically did not trigger it to go away from Fable. And if I click open in Chrome here, we we can see that it built, you know, a pretty basic but decent looking cancer awareness landing page. Is it the most beautiful landing page you've ever seen? Probably not. But the point is cancer wasn't a trigger word that made it so it wouldn't actually generate. Now to test the coding capabilities, I went and said, build me a working clone of Megabonk. I don't know if you've ever played this game, but it's sort of like a 3D-ish version of Vampire Survivor where you run around and creatures try to attack you, but you pick up XP and then it sort of auto attacks the characters around you. I've seen Minecraft and Pokémon and various other clones, so I figured let's try a Megabonk clone. And I actually started the process before recording the video, and we can see that it used over 90,000 tokens, and it actually ran for about an hour to get to where this is. But let's go ahead and open up the Megabonk that it made. So I guess I pick my weapon to start. So WASD to move, space is jump, weapons fire automatically. So I guess I'll start with Bonkey since it recommends that as the starting weapon, and then we'll click bonk. And look at this, we actually have a 3D game. We can see like the little like attack wave coming out of my character. Here's a monster. All right, and we can see it's attacking the monster, and it gave me XP. Like, this works. Jump works. It's kind of hard to tell that he's jumping, but it works. So I'm basically just dodging these characters but also trying to get them in range of my attack. And we can see our little level up bar here at the top. And let's see— ah, he attacked me. I just want to see what happens when we level up, so I might fast forward here a minute. Okay, so I just leveled and we have some options. So Lucky Bonk, +6% crit chance. Zoomies, 8% move speed. Bonk Hammer, level 1 to 2, slams the ground around you, damage 30, radius— all right, let's just upgrade our hammer. Oh, look at that, now our hammer is way more powerful! Okay, that's sick, that's actually really good. Although I feel like I'm gonna level up too quickly now that my hammer is more powerful. But let's level up one more time. Let's do our Sky Zapper. Lightning strikes one random enemy. Oh, look at that, the lightning just strikes every once in a while and kills one of the things near me. Okay, this actually one-shotted this game and it's working really, really well. I need to stop playing because otherwise, uh, I'll just end up keep playing this game. I mean, Mega Bonk itself is a really fun game, so obviously I'm gonna clone a really fun game. Let's see what happens when I die. Let's let them kill me. Oh, I got bonked. I survived for 1 minute 33 seconds, level 3, 61 bonks delivered, 31 gold looted, gems eaten. I mean, it works. It works really, really well to like one-shot a game. I'm not creative enough right now to think of my own game, so I just tell it to model existing games. And for that use case, it works really, really well. Anyway, I wanted to break down what was real, what was hype, all the details I could share about Fable, what you need to know, and do a couple quick tests with it to see if I found similar results. It's really, really impressive, but also again, uses a lot of tokens, is kind of slow, is very expensive. We only have access to this for a couple weeks and it censors the hell out of anything biology related, LLM related, cybersecurity related. So there's a lot of censorship going on, but it is a really, really good model overall. So hopefully you learned something from this. I'm trying to give as balanced of a take as possible. If you like videos like this, maybe consider liking this one and subscribing to this channel. I make AI news breakdowns every Friday, and then in between the AI news breakdowns, I test tools and share cool practical use cases for AI and futuristic technology. Again, if that's something you're interested in, consider liking this video, subscribing to this channel. Really, really appreciate you. Hopefully I'll see you in the next one. Bye!