Ep 796: New Claude Fable 5 and Mythos 5: Anthropic's Boldest, Riskiest Launch

This is the Everyday AI Show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life.

Anthropic just released the most powerful AI model the public has ever gotten its hands on. It's called Claude Fable-5 from the Mythos family, and it's a genuine step change in AI capabilities. Not just another incremental update for most casual users. In use cases shared at launch, Stripe actually revealed that it used the family of models to migrate a 50-million-line codebase in a single day, work that Anthropic says would normally take a team like 2 months. So that's the good news. But here's the part that no one's saying out loud: the Fable and Mythos rollout might be the first real step toward AI that's not democratized, because after June 22nd, your average Claude users will likely lose access entirely, leaving the true frontier power to big enterprises that have big API budgets. So intelligence for thee, but not for we. But that's not the only con to the massive pros that this model brings. What else is there to be worried about? Well, Anthropic can decide where this model is allowed to work. It quietly downgrades you without telling you, and it actually keeps every prompt input and output that you send it for 30 days. So yeah, we're breaking down today not just the hype and the capabilities and the biggest upside, but also 4 big catches that should make every business leader nervous and we're going to cover everything else in between as well. All right, let's get into it. So here's the big picture. Fable 5 is, yes, definitely a capability shock. So Anthropic shipped the most powerful public Claude model yet and the most capable the world has seen. So this is Fable 5 from the Mythos family. So if you're like, wait, Fable, Mythos, 5, 1, isn't this the first? Fable, yes. All right, so essentially Fable 5 is the made-for-the-public version of the also new Mythos 5, but with guardrails. So Mythos 5 is now available to the Project Glasswing elite there, but you have to be essentially invited to use the full Mythos 5 model, which is also released. But for everyone else, you have the Fable 5 5, which is the made-for-the-public version with more guardrails. But even so, Fable 5 is a genuine step change. It's the biggest capability leap that Anthropic has shipped. So the upside is clear and it's enormous. It executes real long complex work, not just better chatbots, right? And you can not only use this in, you know, Claude AI and Claude Cowork and Claude Code and all of those things, but also on the API side. So the flexibility and the Capabilities are a big step forward. It's enormous, but it arrives with 4 big catches: gated access restrictions, a lack of transparency, and required data retention. For the most part, these are 4 new bottlenecks, roadblocks, or obstacles that we really haven't seen from a frontier model before, which is why today's show is gonna be less about showing you hands-on some of the capabilities and more about actually discussing the pros and the cons and the facts behind this new model. So on today's show, stick with me for the next 25-ish minutes and you're going to learn why Fable is the most capable public Claude model ever released. You're going to know what the step change actually does across coding science and workflows. You're going to know those 4 catches, the shrinking access usage restrictions, hidden downgrades, and data retention. And I'm going to let you know how business leaders should adopt this frontier model without getting burned come Well, 2.5 weeks from now. All right, let's get into it. Welcome to Everyday AI. My name's Jordan Wilson. We do this thing every day and it's for you. This is your daily livestream podcast and free daily newsletter helping business leaders keep up with the nonstop feature upgrades, model drops, and everything else. I tell you what's important, what's not, and help you grow your company and career. So if that's what you're trying to do, starts here. But make sure you go to youreverydayai.com. That is your free generative AI university. Almost 800 episodes. You can go listen, watch everything else, but make sure you go sign up for the free daily newsletter. We're gonna be recapping the highlights from today's show as well as all of the other breaking news that you need to know that's gonna impact your career. All right, so let's talk Claude, Fable 5, and Mythos 5, which I think are Anthropic's boldest and riskiest launches yet. So here's some of my thoughts. All right. And notice this model came out Tuesday. I decided not to do the show Wednesday because I wanted to actually spend some time on this, right? You know, maybe you saw on YouTube people within 30 minutes of it being released that didn't have early access are like, here's a full review. That's not me. I think especially with the model of these type of capabilities, that's probably irresponsible to put out something within an hour or two, right? So any hot takes that you saw, unless it was from people who had early access, which a lot of people did, I think it was probably premature. So I'll quickly even walk you through some of my first impressions before we break everything down in the show. So is this Fable 5 the best model I've ever used? In some use cases, absolutely. But for me, I'm not personally blown away by Fable. And here's the reason why. And I think this is actually an important distinction because we're going to be looking at benchmarks, both that Anthropic released and from Artificial Analysis. And for the most part, Anthropic does obviously very, very well with the benchmarks. But one thing that literally no one talks about is almost every single benchmark out there is not against OpenAI's most powerful model. So what do I mean by that? Uh, so if you see something called like xHigh, that is the thinking version of OpenAI's GPT-5.5 with extra high reasoning. And that is what almost every single benchmark uses. Sometimes Anthropic and competitors will even just take the default 5.5 model, which isn't very good. All right, so that's important to keep in mind. For me, I use FiFi Pro. That thing is not benchmark. It is expensive, right? So if you're on the $200 a month plan and you're someone like me, I use FiFi Pro at least 20 times a day, usually, usually more. All right, that's why I pay a lot of money, uh, for some of these subscriptions. So when I'm doing some of my, uh, kind of anecdotal testing, uh, between FiFi Pro in Fable, I mean, it's hit or miss. Sometimes Fable is better, but in many instances, FiFi Pro is better. So if you're just looking at this from a pure intelligence standpoint, at least if you are someone that uses something like FiFi Pro all the time, I don't think you're going to be blown away. But I do think that is the, that is the very minority of our audience. I would say the majority of our audience is not using FiFi Pro for every single thing. But that's what I'm saying. Is this a huge capabilities gaps in terms of what was available? Not necessarily. Is it a big capability jump in terms of what people, the majority of users use on a day-to-day basis? Absolutely. So I'm, I'm not trying to take away, right? But I've had a lot of people asking me, what do you think of Fable? And that's my honest answer, right? It's, it's better, you know, I, I'm getting better results right now in Claude Code, um, with Fable than I am in Codex with Phi-5X-Hi. But that's because in Codex you can't use Phi-5 Pro, which is a much more capable model. You know, there's also some anecdotal things like I just saw, uh, Dylan Patel of Semi Analysis, right? One of the most kind of trusted sources in the whole compute and, you know, technical side of AI. So he actually just shared a tweet. I'll read this. He said, usage share of OpenAI grew versus Anthropic yesterday despite the Mythos Fable 5 launch. Multiple power users at Semi Analysis tried Mythos or Fable. They got refusals for nonsensical reasons. Got pissed off at Anthropic, gave Codex a legitimate try. Now they actually prefer Codex to 4.8 Opus. And I do think that's actually going to be an ongoing issue that Anthropic is going to have to address. Anyways, let's get into it. Let's go over the Fable Essentials. So when it comes to any benchmark, Fable 5 is a big step up from where we were with Opus 4.8. So essentially, even let me break this down for our non-technical audience. Now you essentially have 4-ish tiers of Claude models. So you have your base. So we're going from least powerful and fastest to most powerful and slowest. So you start with your Haiku, that's your fast, not very smart. Your Sonnet, that's kind of your, uh, little slower but smarter. Then you have your Opus, which previously was the most powerful yet slowest model. And now you have your Mythos family, which includes Fable. So might use kind of Fable and Mythos interchangeably in certain respects, but that's just to let you know that Fable 5 at least is a big step forward, specifically from Opus 4.8. But you might be using Opus 4.8 even if you think you're using Fable. That's because some of the new strict safeguards block risky topics and you might not even know. So if you're doing things like cyber in biology, those queries might actually fall back to Opus 4.8. So Mythos 5, though, is with fewer restrictions. But right now, that's only available to selected partners through the Project Glasswing or the trusted access. So they are kind of the same in terms of Mythos 5 and Fable 5. But Fable 5 is kind of what's available for anyone in the public. And then your Mythos 5 is essentially your trusted partners version. Both models though are priced at $10 per million input tokens and $50 per million output tokens, which makes it twice as expensive, uh, as Opus 4.8. And that does bring it much more in line with the 5.5 Pro model. All right. So Fable right now is included on all your paid plans, but here's the caveat, only until June 22nd. And then essentially Anthropic said, yeah, we're going to pull it. Maybe we'll bring it back in the future, maybe we won't. But essentially, if you want to keep using Fable 5 post-June 22nd, uh, yeah, you're going to have to fork over the API costs, which are going to be astronomical. All right, so my recommendation is start using Fable 5 as much as you can right now while it's included, right? Uh, reality, you might, if you try a very complex prompt, even on a $200 a month Pro Max plan, you might not even be able to finish it. You know, I get tired though, FYI, of people saying, oh, you're anti- Anthropic, right? You say all these things. No, I'm not. Uh, there's plenty of other people and I have a list of people that did one, they tried one complex prompt with Fable 5 because they really wanted to push it on its capabilities on Max plans paying $100 to $200 a month and said, I hit my rate limit, right? The first prompt I tried, obviously on a $200 a month plan. Hit the rate limit. It couldn't do it. So I would recommend you take some of your most complicated problems, run them now through Fable-5 when you can still have subscription access, right? Because it might be taken away after that. And then compare it to GPT-5.5, right? Which is for the foreseeable future going to be still open for all. Because I think as a business leader, many people are gonna have to start making that decision. They're gonna have to start saying, are we going to start, you know, maybe 5x or 10x our AI spend, right? Because that's probably what it would cost when you talk about the difference between what is included in a subscription. It's subsidized, right? Versus what you're paying for the API. I mean, many people have said it's 100, 200, 500x more expensive. So that's why I would start using Fable 5 now while it's still included in the subscription. I would make sure that you have good access to GPT-5.5 that might, you know, force you to pay the $200 a month plan. And then I would throw your most complex problems at it and then work with your team to see which ones are the best. Because after June 22nd, yeah, Fable-5 is leaving. And I do say maybe it's going to be Fable-5 is going to be the best model for you and maybe you can justify the cost. But I think you don't have too long to, you know, go head to head. So reading from some of Anthropic's release here, a couple, couple things I think that are worth pointing out. So they say today Today we're launching Claude Fable-5, a Mythos-class model that we've made safe for general use. Fable-5's capabilities exceed those of any model we've ever made generally available. It is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas. The longer and more complex the task, the larger Fable-5 lead over our other models. I would have loved to put that to the test, but hit my, hit my, my usage limit. All right, next, here's what they said about the fallback on Opus 4.8. Releasing a model this capable comes with risks. Without safeguards, Fable-5's capabilities in areas like cybersecurity could be misused to cause serious damage. We've therefore launched the model with safeguards that mean queries on some topics will instead receive a response from our next most capable model, Claude 3. Claude, at Obis 4.8, to release the model both safely and quickly. We've turned these safeguards conservatively. They'll sometimes catch harmless requests, though they trigger on average in less than 5% of sessions. With more capable models arriving in the coming months, we're working to improve our safeguards and reduce false positives as quickly as we can. But AI moves too fast to follow, but you're expected to keep up. Otherwise, your career or company might lag behind while AI-native competitors leap ahead. But you don't have 10 hours a day to understand it all. That's what I do for you. But after 700+ episodes of Everyday AI, the most common questions I get is, where do I start? That's why we created the Start Here series, an ongoing podcast series of more than a dozen episodes you can listen to in order. It covers the AI basics for beginners and sharpens the skills of AI champions pushing their companies forward. In the ongoing series, we explain complex trends in simple language that you can turn into action. There's 3 ways to jump in. Number 1, go scroll back to the first one in episode 691. Number 2, tap the link in your show notes at any time for the Start Here series, or you can just go to starthereseries.com, which also gives you free access to our inner circle community where you can connect with other business leaders leaders doing the same. The Start Here series will slow down the pace of AI so you can get ahead. Here is the thing that I think all business leaders need to pay close attention to. This is the new data retention policy, and this is extremely troublesome. And I hope this is something Anthropic reconsiders. I do assume they'll start to get a lot of flack as the mainstream media reads the fine print or starts paying attention. So they said, finally, we're making a change to the way we handle business customer data for Fable 5, Mythos 5, and future models with similar or higher capability levels. We will require 30-day retention for all traffic on Mythos class, uh, models on both first and third-party surfaces. We don't use this data to train new Claude models or for any non-safety-related purposes. And we've instituted new privacy protections, including logging all human access to the data and ensuring its deletion after 30 days in almost all cases. The data will help us defend against complex and novel attacks, including new jailbreaks and attacks that operate across many requests, as well as help us identify and reduce false positives. All right. This is a first ever. FYI, this is the first time a frontier AI company has decided to not make data retention optional. So if you want to use this, you have no choice. Even if you— there is no turning off, you know, the concept of turning off model training means, you know, big company can't access my data. Nope, not anymore. If you're using Fable-5, Mythos-5, or any future capable models, Anthropic is keeping your inputs and the outputs for 30 days. Here's what that means. So Fable 5 forces the 30-day retention on everyone. So even if you previously had a no-retention deal that you signed, which is most enterprises, nope, nope. You are, by using this, you are giving Anthropic permission to access everything. And it's the first time a lab has chosen this. And I saw some people not really, uh, paying attention to what this means, and they're like, oh no, this is part of the new, uh, the president's new executive order. Order on AI. No, it's not. Uh, this is something that Anthropic chose to do, and there's huge ramifications here. I mean, talk about healthcare, right? So that lets, uh, Anthropic legally handle patient data that requires no retention. So the, the whole, you know, zero-day retention, that's out the window. No such thing anymore. HIPAA zero-day retention, no longer a thing, at least according to, what Anthropic released. So what does this actually mean? Humans can review flagged chats. Yeah. All right. So this is just risky. It's not just risky for legal, confidential, privileged client material. It's risky for just about anything. Giving a human the ability to review your enterprise chats, I think sets an extremely dangerous precedent. And I think and hope it's something that gets called out, uh, among other, uh, leaders in the AI space. And I hope and assume that OpenAI, Google, and Microsoft will not follow suit because I do think that this is an extremely dangerous precedent to set. Uh, and I'm going to tell you why. Well, another thing that's worrisome is Memery actually pulled old chat into Fable sessions. Uh, so this has been confirmed anecdotally. I confirmed this myself. I've seen other people that have confirmed this, that they literally can't even use the model because if they've been working on things that were considered, you know, dangerous or risky by today's precedent, but not by yesterday's precedent, they literally can't use the model. And I do have some examples of that. But here's why this is worrisome. Anthropic, y'all, and please don't, don't be like, oh my gosh, Jordan, you're so anti-Anthropic. No, I'm not. I am pro-facts, pro-truth. right? And the fact and the truth is Anthropic, more than any other lab, has had this little problem with leaking. Okay, so you know that Claude code thing, one of their most successful products ever? Yeah, the source code, whoops, got leaked. Okay, you know what else? This Mythos model, uh, this, the, the most powerful model that Anthropic has spent months hyping and You know, putting all this marketing around as they prepare for their, uh, IPO to go public, right? This Mythos thing, it's so good, so powerful. Guess what happened? Whoops. It was leaked, right? Literally Discord users guessed and got access to it when they weren't supposed to have access to it. So I'm not saying that your chats, your company's personal data is at risk of leaking. I'm not saying that. But if you look at this new rule that you can't opt out of, and if you look at Anthropic's recent history, these things are in the last couple of months, right? This isn't when they were a young startup 4 years ago and didn't have everything in place. This is happening like now. Again, I'm not showing bias. I am trying to inform you of making the right decision for your company. That's all I'm saying, because Anthropic does not have a stellar record, especially recently, of keeping private matters private. All right, back to the fable show. So Anthropic is kind of framing, uh, both of these as mythos class models. So Fable is not necessarily a step down from Mythos, but it is in certain ways, right? Uh, I, I did post about this online that I didn't necessarily like how Anthropic put out their benchmarks, which I'm gonna show here in a second, because what they did is they labeled Fable 5 and Mythos 5 in the same column. And presumably, you know, when they put them in the same column, they took whichever benchmark was better and applied it unilaterally to both models. It's not the way this works. I would like a little bit more transparency from Anthropic when it came to posting on their release site on the benchmarks, because a model with more safeguards is obviously going to have lower benchmarks because in some instances it is going to default or fall back to Opus 4.8, which is now, you know, a good generation behind, right? Because of Mythos and Fable. So I think that's an important thing to keep your, your eyes on that even though Fable is part of the Mythos family, it is in theory less powerful. But overall, the capabilities are undeniable of this new model, right? In, in some of my initial testing,, right? I was able to, um, kind of break through on some, uh, individual use cases where Phi-5X-Hi and Opus, you know, 4.8 max reasoning couldn't get through, uh, especially when it came to those longer running tasks. So, uh, on agentic capabilities, right? Obviously big capability leaps. Um, you know, Anthropic says Fable's lead really grows on the longer, more complex tasks, but Again, it's tough, right? Yeah. If you're on a $20 a month plan, good luck. You know, really realizing the utility here, which I think is unfortunate. You do have to unfortunately be on that $100, $200 plan to really, I think, realize the true benefits of this class of models. But the shift shows up first in anything coding apps. You know, one thing that I think has been going very viral in the first, you know, 48 hours of release its ability to create interactive and immersive 3D worlds via code, stitching together, you know, hard and long workflows, things like that. So from a benchmark standpoint, here's what I was talking about previously, you know, kind of sticking Mythos 5 and Fable 5 in the same column, even though they would— if you went through each of them individually and you did see this a little bit more in their model card, So I know they did test them individually, so I don't like that they put them on the same pane. Anyways, you know, when it came to a lot of the big benchmarks, you know, big, big steps, big steps, uh, not just, you know, from their previous model, Opus 4.8, uh, but also GPT-5.5, uh, Gemini 3.1 Pro, Gemini 3.5 Flash, right? Essentially in almost, not almost every, but in a majority of meaningful benchmarks, Mythos 5 and Fable 5 shot ahead. But not every single one, right? There's a handful of pretty important benchmarks where you would have thought previously, right? I think a lot of people, especially the way that Anthropic was hyping this and they've spent months hyping it, you would have thought that every single benchmark that existed aside from speed and cost,, right? But when it came to raw capabilities, you would have assumed that Mythos was gonna wipe everything out, but it wasn't like that. Um, right? So if you look at like GPQA Diamond, um, not, not the top, right? It's, it's behind most models from OpenAI. It's behind Gemini 3.1 Pro. So GPQA Diamond is a difficult science reasoning benchmark made up of expert-level biology, physics, and chemistry questions. So it's often used to test whether top AI models can handle PhD-level reasoning, not just general knowledge. Much, which to me it's interesting, right? Because Anthropic made all this hoopla about its use on bio, biochemistry, all of these things as one of the safeguard reasons. But OpenAI's and Google's models are far ahead of them in this benchmark. Same thing with the Artificial Analysis Long Context Reasoning benchmark. So this is the benchmark that tests whether a model can reason across very long inputs., right? And multiple documents, not just retrieve a single fact from a long prompt. And I think this is the type of, um, for a lot of general, uh, knowledge workers, I think this is an important benchmark that tells you something. And this is where I think I've seen anecdotally where GPT-5.5 Pro and, uh, Fable, I'm like, no, 5.5 Pro is still much better because these are the type of types of prompts I'm using, multiple document dumps, having it retain context, go through multiple rounds of research, creating multiple artifacts. And at least for me, when I can even get it in on a single prompt on a $200 plan, eh, I'm like Fable 5, not that good. Right? If I'm being honest, not that good. You know, it, it may be if, if, if I'm saying Phi Phi Pro gives me A level on this, um, you know, Fable 5 gives me C. Whereas before, you know, Opus gave me a D, uh, but it's still, not, not close. And again, that's just not my personal opinion and that I've confirmed anecdotally. That's according to the artificial analysis long context reasoning benchmark, which I think is an important one. So yeah, there you go. Fable 5, very far behind Gemini 3.5 Flash, Gemini 3.1 Pro, even some, you know, open models, Minimax M3, you know, and then the best there is obviously GPT-5.5. So when you talk about using, you know, different models like inside of Codex, inside of Claude Code, right? Those are the types of things when you're giving it access to all of these files and folders and you're having it run a longer task that requires it to maintain that context with a very long, hard input. It's not frontier, right? It's not top class. It's not state of the art. Similarly, and this is one that I've talked about recently, again, maybe some of these benchmarks just hit my personal use cases, but the IF Bench, right? So this is the instruction following bench from Artificial Analysis. So this tests whether an AI model can follow specific instructions and constraints, especially ones it has not seen before. So this helps show whether a model is truly good at instruction following or just optimized for common benchmarks patterns. And this is another thing I'm not even talking about, like SWE Bench Obviously Anthropic's models always do good there, but Anthropic admitted that Claude cheats on SWE-Bench. It looks up the answers and then it tries to give the answers in a way that disguises that it actually cheated. Right. So something like instruction following for me is extremely important. And I think for a lot of knowledge workers are. So if you are someone that, you know, doesn't just give these open-ended, you know, here's some documents, go solve this. I take a lot of time because when I'm running longer, longer running tasks, especially using things like goal mode, using planning mode, I'm putting a decent amount of time into giving a model proper instructions. And I saw this with 4.8. Opus 4.8 was absolutely terrible at instruction following. It just refused to follow simple instructions that I laid out because right away it came up, you know, kind of this new honesty thing that Anthropic says it's pushing. You know, it's like, Jordan, honestly, what you're asking for doesn't exist. And I had to go through back a couple times, tell it to browse the web. I'm not going to, Jordan, because you, you don't know this, Jordan. You're a silly human, but this doesn't exist. And I had to go through multiple times to finally make it do follow instructions. So Anthropic's instruction following are absolutely horrendous. So, you know, on this chart here that our livestream audience can see, its models are consistently the absolute worst. All right. I'm not saying models like, you know, OpenAI's and Google's are the best. Actually, some of the best at instruction following are, you know, Groq is really good. Minimax is good. Mimo, Quen 3.7. So actually some of the open source ones are really good at instruction following, but Anthropic's are a huge step down from not just those, but also the OpenAI and Google models. It's just not good at following instructions. So before you just look at Anthropic's blog post and say, oh my gosh, this is the best model ever and it crushes every single benchmark, any company that puts out a new model, they're obviously going to cherry-pick the best benchmarks. But at least for ones in my use, because when I started first using Fable, I'm like, okay, some things I needed to code a game. State of the art. Amazing. Right? If I need to create a 3D world, fantastic. Right? All these, you know, very specific singular use cases, great. When it comes to maintaining long context and following instructions across general knowledge work, compared to Phi-5 Pro, I'm like, I don't know if I'm going to use this. Right? Even up in between now and June 22nd, I don't think I'm going to use it very much. Aside from testing it, at least for now, for me, GPT-5.5 Pro is better. There's certain instances where it's like, oh, I need to do this in Codex because it involves more automations, it involves computer use, browser use, et cetera. I might be using Fable for some of those instances because in my testing, Fable is a little bit better than 5.5x High on some of those things, but not across the board. And then obviously it's so freaking expensive, right? Actually, I had someone, you know, some shared the artificial analysis, uh, here thing on intelligence versus price. It completely reset the quadrants, right? So now it's like business leaders when they're looking at these things, these certain graphs, right? You have your quadrant on the left side, you have, uh, tasks that are cheaper. This is essentially the, uh, amount of compute money that it takes to complete the artificial analysis test, which is a series of like 15 benchmarks. So you, you know, you always wanna be in this upper left-hand quadrant. Which means it's the smartest, the most intelligent, but it's the cheapest, right? And there used to be like 3 to 5 models that were in that quadrant. And now because Claude Fable 5 is so freaking expensive, it completely reset the graph. So actually someone from Artificial Analysis reached out to me like, hey, what suggestions do you have? You know, and I'm like, I don't know. It's, yeah, because Fable 5 is so, so expensive., right, to get the same level of intelligence. We've been talking about that a lot here. And when we went over it on our Start Here series, talking about token maxing to token efficiency, I think you have to start using as the subsidy era kind of starts to fade out, right? Anthropic is stopping it. Google is stopping it. Microsoft stopped it with, you know, GitHub for now. We'll see what else they do. OpenAI luckily hasn't really stopped it too much yet, right? But essentially they're starting to restrict where you could get, you know, thousands of dollars of, of inference or, you know, tens of thousands of dollars of use if you're just looking at the API side for, you know, $100 a month if you're on a, you know, higher tier. That time is going away. So I think, uh, business leaders need to pay more and more attention about the cost per intelligence. Um, you know, there's certain units you can look at, like a cost per intelligence unit is extremely important. And now Fable 5, yes, it's, you know, uh, beats everyone else on the artificial analysis. But not— I don't think it's worth the price, right? It is more than twice as much as the next tier, which were technically expensive, right? The 5.5 High Series, OPUS 4.8, right? Those are kind of the leaders in terms of the most intelligence. So you're paying more than twice as much for only a roughly 6% increase in overall intelligence. So I don't know, business leader, if you're out there spending I, I don't know, $1 million a quarter on AI or even $500,000 a year on AI. Are you going to want to double that for incremental increase? I would say probably not. Uh, but again, that's for you to decide. So let's talk about some more areas where it did really good. So in Anthropic's biology exercise, Mythos, assisted generalists beat world-leading specialist teams. So yes, in a setup they gave, you know, essentially generalists, uh, you know, people that had a baseline understanding of biology but were not experts. They gave them Mythos family levels and they had them go against world-leading specialist teams. And the generalists with Mythos, you know, crushed the specialists. And I think that's really indicative to what having a state-of-the-art model can do. If you speak the language, there's no, right? And we talk about that with GDP Val all the time on this show. There's no more, there's no more, um, comparison, right? Uh, if you know what you're doing, if you understand the language, and if you know how to work with a large language model, you're going to beat the world's leaders every single time. And Anthropic's, uh, kind of experiment here showed that. So graders said that a 6 in this, uh, experiment, uh, graders said that a 16-hour output normally needs 40 to 95 working days. So what the generalists were able to accomplish with Mythos in a 16-hour output compared to what would normally take expert teams multiple months. Okay. So that shows you the step change here that a model like Fable-5 and Mythos-5 unlocks. So this is the very power though that Anthropic called dangerous. Days before shipping it, which is interesting. And we're going to get to here in a minute. And the viral, like the viral nature of this model can't be denied, right? If you go look on Twitter, LinkedIn, wherever, you're going to see some extremely impressive examples that were just literally not possible before. Right. And I'm not just talking about pelicans riding a unicycle or whatever that SVG is, right? The demos now that you're able to write, whether you're creating full games, 3D worlds, simulations, websites. I do think for individual artifacts, especially on the creative end, Fable 5 is now in a league of its own, even compared to, I think, GPT-5.5 Pro when you're talking about those single artifacts, right? Coding a game, you know, creating a 3D world. But I mean, here's my question to a lot of people. Maybe that's part of your job, but I'd say so many of these things that you see online that people think that means that a model is good. What percentage of people are creating 3D worlds for their job, right? If you're a video game designer, sure. An artistic director, sure. But I don't know, I would venture to say like 0.01% of the US population is using AI models for things like that. So, uh, you know, a lot of, um, Outputs were, you know, people that shared about where Mythos or Fable, you know, tested, fixed, shipped, and documented the entire thing. So non-engineers can now prototype work that once needed teams of engineers using AI, right? So even if you look in the AI phase, if you go back and look pre-reasoning models, what a non-technical person can accomplish with a something like Fable used to take a team of AI-enabled engineers. So Fable's ability to kind of run those long tasks with singular outputs that are very focused and don't require a ton of instruction following, you know, people have been sharing like, you know, go clone this Pokémon game, go clone A, B, and C, right? Go clone Replit, go clone anything. It does a wildly impressive job doing those things. But it's not all amazing demos, benchmarks, and roses, uh, because I think we have to now tap into the 4 downsides. And I'm gonna go through these quickly, uh, 'cause the show's already running long. So Fable, obviously on the cost, $10 per million input and $50 per million output. So it is more than double the cost of Opus 4.8, and now it puts it more in line with GPT-5 Pro, which is the world's most expensive model. So subscription access though, ending soon. So yeah, this whole democratizing AI, yeah, it could be out the window. We might now be on the AI for thee versus AI for we side of AI. So beyond who pays though, yeah, so number one, good luck because from an affordability, I'm not paying for GPT-5.5 Pro on the API. Absolutely not. I'd go broke, right? Same thing. I'm not going to be paying for Fable 5, especially if GPT-5.5 Pro is continued to be offered in a subscription. So beyond the democratization, there's another big downside. Actually, let, let me talk about this. So the downside number 2 is Anthropic decides where Fable can work. So things like cyber, biology, chemistry, and model copying prompts can be rerouted to Opus 4.8 without telling you. And Anthropic says that over 95% of Fable sessions avoid this fallback. But still, 5%. That's a decent percentage where you're just not going to be able to use Fable even if you're paying for it. And some of these limits trigger with no notice to you at all or just get outright refusal. So I talked about, you know, how Dylan Patel shared something about Sumo Analysis, some of their engineers getting refusals. You know, here, very well known, and hopefully I get his, his name right, Daria Úna Tamás is a well-known professor and biomedical researcher working on AI capabilities. He couldn't even use, right, he shared this online, he couldn't even use Fable because it brought in the memory and the history of his chats, right? That's another thing. So even if you've worked on things in biochemistry, testing models before, and if you try to start a new chat using Fable, you might get outright denied. So that's downside 2. Anthropic decides where Fable can work and who can't. And you also have to think about your chat history and your memory. If you're using it on the, uh, Claude on the website, might be impacting your ability to even use it at all. All right, so downside number 3 is the invisible downgrades can quietly break trust. So those hidden safeguards can silently weaken Fable on AI and machine learning research work. I do assume that eventually I will run into this because I'm constantly doing some meta-prompting inside large language models, showing them my input, showing them the output, and then copying and pasting the chain of thought to look at the tool calling and then trying to make my prompting approach more token efficient. I assume once I do that a couple of times on Fable, it's not going to work. I'll share that if I actually have the limits to do that testing. But Anthropic estimates that these silent limits touch about 0.03% of traffic. But critics say that this quiet weakening is more dangerous than an honest refusal. So yeah, they're just going to roll you back and not even tell you that. Again, unprecedented move from Anthropic. So it's not like it's saying, oh, you've been downgraded in this case to using, you know, Opus 4.8. It's just not going to tell you. Right. And Anthropic says in its model card, this is more about anti— or sorry, it's about stopping model distillation or competitors from gaining an unfair advantage by using their models. But to me, it's absolutely asinine. Obviously, you want to do everything you can to protect against distillation, but do an outright refusal. Say you can't use our model for that versus giving them intentionally, right? You're in theory, you are intentionally making the outputs worse and not telling the user. That is a huge miss, like a huge trust barrier, right? Uh, and this is again what Semi Analysis, uh, you know, shared about when they were trying to do these things. Um, so they said Anthropic's model will not help you if it thinks your ML research, ML engineering is interesting, and/or will secretly degrade its IQ so that the average engineer won't notice. We are already seeing Anthropic's latest models, moderation filters, our GPU inference research and programming. All right. And downside number 4, the last one, data retention. We already touched on this, but I think this becomes a deal breaker. I do hope Anthropic does the right thing and revisits this because I don't think this is a good idea, especially with Anthropic's recent history of keeping Private data, private. So Anthropic does retain all Mythos class prompts and outputs for 30 days. So that breaks through any zero data retention or HIPAA-style setups that you had previously. But Anthropic does say it doesn't train on those input and outputs. But I mean, those 4 catches right there are huge. But the model itself actually adds a 5th risk, and that is Fable confidently faking its own work. Work, right? So Anthropic, to their credit, does share about these things in the model card, but most people don't read the model card. I created like 20 different NotebookLM audio overviews on different aspects of the model card. Yeah, I'm a dork. But in an audit of 886 real internal uses, Anthropic studied how Mythos falls short in this regard. So in 16 of those sessions of the 886, Mythos or Um, Fable claimed untested work had been fully tested. So not just, uh, skirting around the issue, but said, nope, this has been tested. You're good. It's correct. This is verified. We tested it. So yeah, even though it didn't. So intentionally lied so it could skip a human review phase. That's wild. That's wild, right? So human verification and not raw capability, I think becomes your team's real bottleneck. Also, as we wrap up here, you have to keep in mind about Anthropic's IPO. All right. And I think you have to keep this in mind with all companies that are about to go public. But surprisingly, we haven't seen this from SpaceX as they go public here in like a couple of days. We haven't seen this from OpenAI, but we've seen a lot of this from Anthropic, which is surprising. Uh, but if you don't understand the storyline I think your business and your budget just becomes another character in Anthropic's story. I think you have to follow the money, right? Because Anthropic just came out with this big blog post, this big statement about how, uh, AI, uh, development needs to slow down. Yet they launched what is the most powerful model in the world, and they've shared their progress on recursive self-improvement. So if you have models that are more capable than anything else and they're improving themselves, but you're the ones calling for a slowdown, but you're also trying to go public, right? I think the Mythos and the Fable really just weaves into this long-term narrative that Anthropic has been painting this picture of, oh, Mythos is so powerful, it's so scary, the world is not ready, don't worry, We are the answer, right? They painted the story of the villain and how dangerous it is so they could be the hero. And then they came out with Fable 5 and they said, don't worry, enterprise, don't worry, Wall Street, we are here to protect you. This is Fable 5, right? So you can go have a little taste, but we're going to start charging you through the ears for this model, uh, here in 2 weeks, just in time to ramp up our revenue right before we go public. So hype sells capability, safeguards sell control, credits sell their future revenue. But all I'm saying is maybe Fable 5 is going to change your business. Maybe you have access to Mythos 5. I'm not saying that that's not the case. All I'm saying is don't believe anyone's hype. Not Anthropic's, not OpenAI's, not Google's, not Microsoft's, not anyone else. When a new model drops that promises world-breaking capabilities, you have to do your due diligence. Is it safe for your company to use? What's the ROI? What are your use cases? What are your internal benchmarks? So before we all go just see everything that's going viral online, take the time, understand exactly what these new models and the new capabilities mean for you, your company, and your career. So that, my friends, is a wrap on a longer show. I hope this one was helpful. If so, please let me know, uh, by by number one, subscribing to the podcast if you're listening. So if you're listening on Spotify, appreciate that. Apple Podcasts, make sure to follow the show, leave us a rating if you could, leave feedback. All right. That I always go through the comments when I can. And then when you're done, make sure you go to youreverydayai.com, sign up for the free daily newsletter. Thanks for tuning in. This was a longer one, but I still hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'all.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.