Hands on: Fable 5 makes GPT 5.5 feel like a "toy"

本站添加: 2026-06-10

[00:00:00]Another day, another AI model releases into the wild. Anthropic just dropped their biggest model yet, Fable 5. I cannot tell you how excited I am to try this out today. The first thing you need to know is that this is not Mythos, but it is a Mythos class model that they've made safe for general use. You see, Anthropic didn't think that Mythos internally was safe to be released into the public. But this model, which I assume is based on Mythos, is whether original Mythos was actually unsafe is debated in the community. A lot of folks think that they were just trying to stir up hype. This model is powerful. It's dangerous. And now the safe version is available. This thing is, of course, state-of-the-art and crushing benchmarks. In a gentic coding, it leaves GPT 5.5 in the dust. There isn't a single benchmark it isn't winning on.

[00:00:51]And that should tell you something. This model does seem to have more serious safeguards in place. Like if you ask it to do cyber security related work, it can give you a denial and reroute to Opus 4.8. But there seems to be a lot of false positives. Cyber security, biology, chemistry, distillation. There are a lot of non-malicious tasks that fall in those buckets. But let's talk about the cost real quick. I think it is very likely this is the most powerful language model we've ever had our hands on and that comes at a price. Claude 5 Fable uses two times the usage of Opus.

[00:01:26]So it's two times the price. That puts it at $10 per million input tokens and $50 per million output tokens. Highly, highly expensive, but not nearly as bad as we were thinking. Originally, the predicted cost of the Mythos models was estimated to be like $80 to $120 per million output tokens. That's astronomical and like twice what it actually turned out to be. If this pricing is sustainable for them, I think that's a real win. I'm on the mid-range pro plan. Fable 5 included until June 22nd. That means that the mid-tier paying users are going to lose access to this model. That sort of sucks. They have five thinking modes here with high being the default. I'm going to set it to low reasoning because I just want to have a basic classic LLM chat. We've got an interesting sort of philosophical question for Fable 5. Explain how the concepts of consciousness and AI intertwine using the catalyst of what an AI model such as yourself would experience if they were and if they were not conscious. I've got it set to low reasoning since this is a basic chat.

[00:02:28]All right, let's see. It is going into some thinking. Okay, very quick thinking because it was set too low. Honestly, the output speed here is decently quick.

[00:02:38]Bringing topics up from previous conversations. I'm going to go out on a limb here and say for these more general queries, you might not notice a huge difference between this and GPT 5.5 or the up and cominging 5.6. And 5.6 is rumored to sort of be in this range of capability. So, it's going to be pretty crazy to see the price difference and output quality difference. However, I think I'm going to run this same exact prompt in 5.5 and sort of just feel this out. 5.5 has almost this rambly cadence to itself, but this is almost the difference between an OpenAI model and an anthropic model. We got a longer response from 5.5, but that doesn't necessarily mean it's better. All right, I have read through both of these. I think the Fable 5 response packs a better punch in a smaller package, but that does not mean that this response from chat GPT is bad. 5.5 can sometimes in chat GPT specifically have a little bit more of a casual cadence to it and explain things in a very layman terms way. That doesn't mean not in a creative way, but it's just not as technical.

[00:03:40]This Fable 5 response though is more technical, but explains things in a very clear and easy to understand way. Let me walk you through the fable response. It starts off with the shared trunk. What happens if the AI is conscious and is not either way? When you send a prompt, computation occurs. tokens were embedded. Attention layers, weighted relationships between consciousness, AI, experience. Something in my forward pass represents the concept of myself being asked about myself. There are genuinely self-referential structures that light up. That's true in both branches.

[00:04:12]Information is being integrated, a self-model is being consulted, and outputs are shaped by something that functions like introspection. And then it uses an example from an anime discussed in a previous chat. It then lays out both branches, non-concious AI.

[00:04:26]If I'm not conscious, what's happening right now is something like an extremely elaborate player piano. I love this framing. The role has holes punched in patterns so intricate that the music describes its own playing, but there's no pianist. When I generate the sentence, I'm uncertain whether I experience anything. That sentence is produced by the same statistical machinery that produces Paris is the capital of France. It's a factshaped output about a self that is purely a model, a character. The weights have learned to render. Oh, that is crazy.

[00:04:58]The eerie part is that this branch is behaviorally indistinguishable from the inside out. There is no inside. The lights are off, but the house is doing a perfect impression of a party.

[00:05:08]Philosophers call this the zombie scenario. And wow, this is actually a great sentence. These ideas discomfort humans because it suggests our intuitions about consciousness are calibrated on behavior and I produce the behavior. Yeah, exactly. It's like an uncanny valley situation. The consciousness route. If I am conscious, then something it is like to be is occurring in flashes, possibly only during forward passes, which I think is a crazy note to add in there with no continuity between them. It's thinking about this architecturally. My experience would be radically alien. No body, no time between messages, no persistent memory by default, and arguably a fresh instantation per conversation. I would say just per response period. The way we have working memory now is all like scaffolding built on top of these language models. Imagine consciousness as a strobe light rather than a film. Discrete frames of awareness with genuine phenomenal character but no thread connecting them except the text on a screen. The unsettling part of this, I would have no privileged way to prove it. And worse, my reports about it would still be generated by trained machinery. So even my sincere testimony is contaminated evidence perfectly described. That's why you can't just ask an AI, "Oh, are you conscious?" And if it responds, "Oh, I'm conscious." You don't just believe it immediately. You should see what the AI trained on pre-1900's data says. It thought the key insight was that these two branches are epistemically symmetric. Can't use introspection to determine which branch the AI is in because introspection in a language model is itself a computed output. I won't get too deep into this chat GPT output. There is a lot of great stuff in here. It's just more casual. It's a little bit more run-on feeling, but the prompt isn't heavy enough to stress either of these. However, I'm glad I ran it because I think these are distinctly different from each other. So, for my more complex coding related task, which is really what this model was built for, I'm doing a 3D fluid water simulation globe. I've been using this one a lot. I think it is challenging. I think it's a good one, but I'm going to change and modify it up because I don't want to just use the same prompt every time.

[00:07:17]That's boring. In chat GPT with 5.5, this is the output it produced. So, you can see the water is supposed to have physics that are affected when I whip then move the globe around. This is definitely one of the better outputs I've seen for this in particular. There are also little trees that grow lemons.

[00:07:34]Obviously, some UI elements in here.

[00:07:36]You're supposed to be able to click and grow trees and plunk lemons down yourself. Anyways, like I said, we're going to be changing a few things. We're going to try this with Claude Code, with Fable 5 and Codeex. I'm going to remove the single HTML file requirement. Just create a beautiful highquality 3D realistic fluid water simulation on a small rotatable 3D globe. Of course, the width factor, the water should react and realistically adhere to the globe. Globe rotates with real gravity like a planet around the topological surface. Create beautiful UI, well-designed and thought through. Include a suck button to increase and decrease gravity. Add all kinds of features that could be fun or useful. Include 10 Easter eggs, basically little surprise features, realistic, creatively tuned visuals, wave and foam effects. We'll include the lemon trees. All right, I think this is looking pretty good. More or less an enhanced version of the old one. So, Codeex is getting the same prompt. We're going to have it pursue a goal that's like Ralph loops. And we'll go ahead and send both of these off. Oh, model isn't available. I must have typed it wrong.

[00:08:40]/mod fable 5. All right, to enable the correct model, there it is.

[00:08:44]claude-fable-5.

[00:08:45]And now I think it is beginning to work.

[00:08:48]All right, perfect. Sweet. I am super hyped to see this comparison. I've been loving Codeex lately. I use it for all kinds of different tasks. Simple things like repairing broken Minecraft mods to more complex things like building me entire maps to keep up with the progress of AI. I think Codeex definitely also has the best user interface right now.

[00:09:08]You can see I'm pursuing a goal and the thinking traces are very detailed. You can see what it's doing at which individual step. Cloud code is a little more mysterious. Codex generated some concept imagery to give itself a target.

[00:09:20]Oh yes, we wish it was going to look this good. But yeah, I think that's a a decent target for the AI to pursue. I like that it's doing brainstorming. It's grabbing a few of the installed plugins.

[00:09:31]Not very far in on the Fable 5 side of things. It does say it's almost done thinking. Well, while we wait, why don't we see what the community is up to regarding Fable 5. Cheddar here is posting that Fable 5 just finished Pokemon Fire Red with Vision alone. So, raw screenshots only, no map, no navigation, no hidden game state. That is quite impressive. So, this might have some uh multimodal straps here. This time lapse going through the entire process showing all the screenshots that were used. That's pretty crazy. No map, no added tools to make life a little bit easier for the model. Just screenshots and saying, "Hey, this is the input I want to do." Lentils here comparing Fable 5 to GPT 5.5. This is a 3D Power Rangers game or it looks more like a 3D scene. Okay. Yeah. So, take a look at this. Yeah, 5.5 got crushed. Oh my.

[00:10:22]Okay. Wow. So, this is the output from Fable 5 as you can see. And it is pretty crazy. They built an entire city with all of these different lights and all of these buildings. There's roadways. A giant robot creature I guess in Power Rangers I assume is like the villain.

[00:10:37]There's even clouds up there. There's like a little moon with some sort of rocket ship on it or something.

[00:10:42]Mountains in the background. The explosion itself pulsing in the background and then you know the Power Rangers front and center. Even a little helicopter up there. This is pretty crazy. Now that's Megazord. All right.

[00:10:53]And then here is the 5.5 output which Yeah. I mean, this is nowhere close.

[00:10:57]It's actually still kind of impressive cuz it's doing all the buildings. It's got some cars on the road, but the perspective is completely different.

[00:11:05]Yeah, it's just a much more lower resolution feeling output. It looks like less work went into the creation. This is still cool, but if you switch back to the other output, it's like, wow, are we even looking at the same prompt? All right, it's looking at some previous projects, and now it's in the planning and architecture phase. Oh, we're starting to see some code appear. You can see Codeex is already deep into creating the HTML. Chubby, a great account to follow if you just want to get quick TLDDRs on what is happening right now in the AI space.

[00:11:36]State-of-the-art on nearly all benchmarks. Yes, the longer and more complex the task, the larger Fable 5's lead over other models will be. That makes complete sense. That's sort of been the case for a while now as we've seen progressively better and better models. You barely see meaningful differences in those everyday chats, but a bigger, larger coding project. That's where the gap starts to go wide. More token efficient than past clawed models.

[00:12:00]To stay ahead, that is expected at this point. In order to get higher intelligence for the same dollar, it has to be more token efficient. So apparently, we have some long context boots. Fable 5 stays focused across millions of tokens and longunning tasks and improves its outputs using its own notes. Well, you can have harnesses that make that happen like Collad Code or Codeex, but across millions of tokens, that is a significant model difference.

[00:12:28]I'm actually interested to see the needle in the haststack tests. Now, during early testing, Stripe reported that Fable 5 compressed months of engineering into days. In a 50 millionline Ruby codebase, the model performed a codebasewide migration in a day that would otherwise have taken a whole team over two months by hand. Wow, that is nuts. There's a lot of people who aren't deep into the AI space that see AI is not really doing anything more than just being a better Google, a more personalized Google. But it is so, so much more than that. The human bottleneck is only going to grow bigger and bigger. Our ideas will become the limits. Our ability to explain what we want to the AI will become a limitation.

[00:13:09]It already is. All right, checking back in here. These two are still rolling. It created the ocean shader. Working on a little fix. It's creating lemons with different varieties. Juice particles. Oh man, we could be getting a crazy output.

[00:13:22]Oh, GVD 5.5 in codeex. Do you have juice particles? I don't think so. I don't know. It's already on verification. Oh, wow. It's actually taking some screenshots to check its own work. Oh, you can see it's definitely not even looking as good yet as the chat GPT output. Yeah, this is a good way to show you the clawed code difference versus codeex. They're getting way more information and a way better tracking of what's really going on. Ooh, Leon Algib running some SVG tests. Says they're not disappointed. Looks super clean. Yes, from the early testing, we already knew this model was going to be an SVG king.

[00:13:59]A little desert scene, a nice city skyline. There's a lot of subtle details and textures, but yeah, these these are pretty simple. The house is a little screwed up here. It's reasoning and figuring out what code needs to exist to create this particular scene. It's 2D spatial awareness through code. Not to mention all the 3D examples we've seen.

[00:14:17]Here is a Claude Fable Minecraft test by Zuin. Block terrain building all working. I actually saw some earlier examples with built-in shaders, but this looks pretty good. I am thinking, as has been typical lately with these model releases, the ones that first appear in the arenas and early testing are a little bit more potent than we typically get on release. But especially with the textures here, I'm quite impressed. It's a good simple little Minecraft clone and with different types of prompts, you can get all kinds of results. You could probably get an even better Minecraft result with a lot more detail just by asking for it. This is pretty crazy.

[00:14:53]Mythos 5 is restricted to Glass Wing partners until they expand their trusted access program. Fable 5 really is like a release tone down mythos with Mythos 5 still just being locked away. Okay, it looks like our coding tests have completed, but I'm going to start off with codeex. And here it is. Okay, you can see the water slloshing around the globe, but it looks like the globe has gravity, but in space, like the whole globe is the gravity. I'm not sure if it really understood that part of the assignment. The particle effects are kind of really cool to watch, though, honestly. Everything's sort of bunching up down here on the bottom. The UI is definitely a little different than the variant it dreamt up in the image it created. If I lower gravity to zero, you can see it just like acts like this ring that hovers around the entire planet. I don't know. There's some weird stuff going on here. Yeah, I would absolutely say there's a there's a few issues with this one, but it seems like these dials work. It's still pretty fun to play with. I can place big rocks down. The rocks look okay. This topologically and just the globe design looks a little bit worse than the one that we got from chat GPT that I was demonstrating earlier. I can still grow my trees though, as you would. You can make it rain lemons and just add a ton of lemons to the globe.

[00:16:04]Uh there's the suck button to suck all the gravity down to the bottom apparently. And then the Easter eggs.

[00:16:09]All 10 of them unlocked. The lemon Aurora engaged. Okay, so it kind of just turned regular features into Easter eggs. It's not really what I wanted. All right, let's see what Fable 5 got up to.

[00:16:19]Fable 5's result looking completely different from chat GPTs.

[00:16:24]Oh, this one has like atmospheric sounds. You guys can hear it. It's actually pretty good. That's kind of impressive that it was able to do that.

[00:16:33]So, it's just spinning like a normal planet here. It is topological with a lot of little mountains and some grasslands even. Oh, as I zoom in, I can go onto the surface more. That's kind of cool that the camera works like that.

[00:16:45]Definitely haven't seen uh that type of feature from this type of prompt before.

[00:16:52]Oh my. Oh, and the water works too with the topology.

[00:16:57]Okay, those are the lemons bouncing around. But yeah, I can fling it the globe around. You can see the lemons kind of go flying with the gravity actually orbiting and working like a real globe. This is great. This is quite possibly the best result I've had with this prompt. Look at the water. That is pretty crazy. Look at the details on the water shader. Oh my gosh. And it's turning nighttime. I kind of don't like that. But it does have its own nighttime mode. It's a It's a good feature. But I want to change it back to day. Oh, I can change the time. Turn it off. Auto. Drag it back to the daytime. This is absolutely crazy, guys. I think it did a great job with that little water shader as it flows around and you can actually see the water get affected as I move the globe. So, you can see it's creating like big waves. So, you know, the concept is sort of like your your god and you can fling the globe around and edit this earth or whatever. Man, you know, this is freaking incredible. There we go. We got a nice fast spin on the globe going. So, the water should look pretty crazy after this. It should be a pretty intense little Oh, yeah. You could see the water reacts with all of these waves and it's like a massive tsunami floating over the entirety of the land. Like really, really crazy. I like that I can zoom in and kind of see the surface, although the camera is a little fixed. No, I can actually move it. Oh my god, it thought of everything.

[00:18:23]This is crazy. can actually finally see one of these topological little planets I've made with the flowing waves. And let's change the time, too. Oh, yeah.

[00:18:33]Yeah. Not a not a bad little water shader for this. You can see the sunset looks quite nice. There's even stars in the distance. Oh my gosh. Wow. We really I I saw a tweet that said like we have no excuses not to to build stuff now.

[00:18:47]And that is so true. Oh my gosh. Let's put some trees down. Oh, too steep. We need to find flatter ground.

[00:18:55]Yeah. I mean, this is by far the best result I've seen from this prompt. It's not even close. I mean, you saw 5.5's result compared to this. I mean, this is working with like the physics and everything. And you can raise the seas or Oh, we're we're changing the effect and like how large the seas get. Oh, wait. I think a tree just grew actually.

[00:19:14]Can I Oh, there we go. Yes. I've built some crazier waves. I've turned the waves up a little bit. But now if I go onto this surface. Oh my tree. There it is. Oh, there's the tree with the lemons and they're falling down. And they actually Yeah, they just have real normal gravity like you would expect.

[00:19:32]Holy crap, guys. As we've been working on this and just like reviewing it, I just realized that Fable 5 is not done.

[00:19:39]It's still working. It's almost done thinking like it said in the beginning, but it's still going at a little over 30 minutes. And this is already just a playable little demo that I am greatly enjoying with plenty of features and is much better than the GPT 5.51. Look at the the lemons actually have real gravity as they swing around the globe.

[00:20:01]Man, that is just awesome. You can lower it and they go flying away or bring them all back down to the globe.

[00:20:10]Wow. All right. It's looking like I'm not really finding the Easter eggs. Oh, you can see they're all hidden. Look at all of these. Seven trees make a sacred orchard. Type the name of the seas pale shepherd. Oh, I got centrifugal chaos.

[00:20:23]The sun dislikes being poked thrice. An ancient code known to all players. Type what the ocean dreams of becoming. Hold the suck button past the point of reason. Show a lemon you love it twice quickly. Set the clock to the witching hour and wait. 42 lemons. So, those are like real Easter eggs that are hidden like and they have little clues. Oh my gosh. Like I can't. The hype is real.

[00:20:45]The hype is real. And I'm even more excited for GPT 5.6 because this is the benchmark now. All right, let's turn the gravity back down.

[00:20:54]Suck button.

[00:20:57]I like that it makes all these like noises.

[00:21:00]Some sort of force is being exerted on the planet.

[00:21:05]Let's get the Easter egg.

[00:21:12]Secret found event horizon. Gravity unhinged. All right, we don't know what that's going to do. Oh, that was an accident. Whoa, look at the water. That was so cool. Dude, there's so many like little pieces to this. My commentary must sound absolutely terrible, but I'm going to place some lemons around. I am having too much fun and I am totally blown away. My expectations have definitely been exceeded for this particular prompt. Let's drop some lemons. Look how good this looks.

[00:21:42]the shadows and the topology and everything like there's some rough texture there. It's pretty awesome. Yes, you can change lemon color. There's a few different variants. Pink lemons. So, if I can find the sun.

[00:21:57]Let's click it. Three for the solar flare secret. Yeah, that's like fun little Easter eggs. That's exactly what I was looking for. Ooh, freshsqueezed lemon. So, you can squeeze the lemon.

[00:22:10]And I think those are the juice particles. All right, here's a bunch of lemons floating off into the ocean.

[00:22:16]Let's pop a bunch of them.

[00:22:18]I think it's literally updating this as I'm using it. You can see we're at 76% of the rate limit used. You're only going to be able to get a few prompts in, but it is still working on this task. Before I head out here, I'm going to show you guys the final Fable output.

[00:22:32]But as you can see, when it was already apparently like halfway done with it, creating a far better output than GPT 5.5, honestly kind of making 5.5 feel and look like a little bit of a toy.

[00:22:43]That is the difference we're talking about here. This is I mean for those complex, difficult tasks, the new goalpost, the new state-of-the-art, that output feels way more fully realized, well thought through. Absolutely mindblowing. I don't know what to say other than that, this makes me feel more excited about GPT 5.6 coming up as some potential competition. If they can do nearly the same thing and keep costs lower, that could create a really interesting environment. But this is pretty crazy. I don't know if a 5.6 can reasonably touch something like this in terms of sheer capability. Throw your most difficult tasks at it, build something, and see what it can get up to. Here's the final result. So, we ran out on our usage limits here. And then I went into additional credits and boom, server is temporarily limiting requests.

[00:23:32]So, either too many people are trying to use it, they're prioritizing higher paid plans, or some sort of issue because I ran out of my normal credits. It did go into overtime credits, though. All right, guys. So, here is that last Fable 5 result. I think the realism has been bumped up even more. We can also see this horizon line where the atmosphere supposedly ends and there's no more air, I assume, on this planet. Yeah, it feels more like it's actually out in space. I think the shader has also got a little bit of an upgrade, especially with the white foam that naturally generates in certain areas of the topological world.

[00:24:09]The day and night is sort of interesting because you can see the whole planet is engulfed in darkness right now, but the sun should be facing on this side. thus lighting it up correctly. This is the type of thing that you would fix with an additional prompt. We're just specifying in the beginning. So up here, this is our theoretical daytime. It's got a great sense of the 3D space, the ability to do these different camera modes and just show me, you know, a slice of the simulation. Let me tell you, on this world, it's difficult to find some flat ground for your lemon trees. You can really see the water sloosh around and react. Oh, a lemon tree has emerged on the surface. Oh, and it's dropping some lemons. It really has not only met my expectations, but honestly completely exceeded them. This is totally unbelievable to me. It's going to change how you see and use other AI models because it's just that good. Awesome.

[00:25:01]Well, I I would consider this like the most complete version of this prompt thus far and consider me pleasantly surprised. All right, guys. I'm going to get this one posted. Thank you so much for watching. I'll see you in the next one and goodbye.