Transcription
I cannot believe it. The biggest model release of the entire year, maybe of all time, happened today, just now. We are about to find out if it lives up to the hype. Claude Fable 5, aka Claude Mythos. That's right. Anthropic has released Claude Mythos 5. It is the safe version of Claude Mythos.
As you can see here, it not only wins every single AI benchmark ever made, it dominates every AI benchmark ever made. But are benchmarks real? Is it as good as all the benchmarks you see here? Because a lot of time benchmarks are a little fake. We're about to find out.
For those who don't remember, Claude Mythos was the model that Anthropic announced a few months ago, but they said, "Hey, you normies, you can't use it. A little too dangerous for you. If you get your hands on this, the entire world is ending."
It is available in all Claude platforms. So you can use it in Claude Code CLI, in the desktop version, as well as in the chat and in co-work. I'm going to show you how to set it up in a second, show you how to get more out of it. We're about to go through the keys of Claude Fable. So how you treat it different than Opus.
So before we get into actually using the model, here's how you need to think differently about Claude Fable 5. Before you ask yourself, "Is Claude doing the work right?" You tell Claude Code to do something. You tell Claude to do something. And then you go and you check to see if it did the right thing, you test it. Now it's different. This is a subtle shift in mentality. Now you are saying, "Is Claude doing the right work?" You can now be confident Claude is building things the right way. It's not building buggy code. Now you just need to make sure it's doing the right work. How do you do that? Well, through loops. And I'm going to go through loops in a second. This has been the talk of the town for the past week. I'll show you how to use loops using a few different methods inside Claude Code in just a second as well. This basically gives Claude the ability to work autonomously non-stop, right, for days and days and days on end. We'll go through that very shortly.
So, here are the three tips Anthropic gave when it came to using Fable 5.
1. Treat Claude like a thought partner. This is quite literally super intelligence. And what people were doing before with Claude was they'd begin work and they'd say, "Okay, Claude, do this next. Do this next. Do this next." Not anymore. Now you need to treat it as an equal, as an equal collaborator on what you're building. So before you build something, you need to go, "Hey, what are your thoughts on this? How would you change this? How would you do it different? How would you approach this? Would you do something completely different?" You treat it as an equal partner in what you're building.
2. You give Claude goals. So instead of telling Claude what to do next, you just give it a high-level goal and it will find its way there. Typically the way we've been vibe coding up until now is you might have a goal 10 steps down the path and you do each step of Claude one by one by one by one. Not anymore. You just give Claude the end goal and it finds its way there. So you let it be more autonomous.
3. And then lastly, this is the last tip here Anthropic was giving is just be more ambitious. You want to now push the limits of what you were doing before. If you thought something was impossible to do with AI before, go for it now. Be more ambitious. And I will show you how to do that a little bit uh later in the video as well.
Okay, so let's get into my recommendations on how to get the most out of Claude Fable. I am doing this inside of Claude Code in the CLI. I'm using a little app called Ghosty, which is a nicer version of the terminal to do this. You can feel free to download that for free if you want. I'm hearing from a lot of people Fable is not showing up in the Claude desktop app just yet. So, I'd recommend coming in here. On top of that, I'm hearing from some people Fable is not showing up inside the model picker. So, make sure to run this as well: `/modclaw-fable-5`. Hit enter. Even if it's not in your picker, even if your Claude hasn't updated yet, you can run that and you will be on the new Claude Fable 5.
I recommend going auto mode. So, shift tab, go to auto mode. It's so smart, you don't need it to ask you for permissions anymore. I know that makes a lot of people upset. You're telling people to do dangerous things. It's smart enough. Do auto mode. It's going to work. It's not going to do anything crazy.
From an effort perspective, if you do `/effort`, I'd recommend starting off on high. This model absolutely torches tokens. It does a tremendous amount of thinking. My recommendation opus was extra high. This I think you should go with high and then you can tune it up and down based on the performance you're getting. So I have it on high auto mode. Fable 5.
So let's talk about how to get the most out of Fable 5. What is the best way to put it in these loops? These loops everyone's talking about. This is the strength of Fable. It is so smart. You don't need to give it little micro steps. You give it a massive goal and it attacks that massive goal. So, you can follow along with me here. Let's do this. What we're going to do is describe what we want to build. What we're not going to build it just yet. What we're going to do is use what I call advanced plan mode first. This isn't just a regular plan mode. This is an advanced plan mode where I'm going to have Claude Fable 5 ask me a bunch of questions to fully understand in depth what I want to build. You should be doing this too. Once we have this advanced plan mode built out, then we will give it a goal, which is the big loop it runs in to build this huge idea.
So, what we're going to build today is a personal productivity app. This is everything. This is your to-do list, Kanban board, calendar, note-taking app, like your Notion on top of that. This is everything. We're building an entire complex app and we're going to build it all in one goal. So, typically when using like Opus, you would build each one of those tools out one by one by one. Fable, we're doing it all at once. So, watch how we do this.
So, here's the prompts we're going to use. I'll put this down below so you can copy and paste it. Feel free to pause right here and grab it so you can do this with me. "I want to build out a personal productivity tool. I wanted to include to-do list, Kanban board, calendar, comprehensive note-taking app, Pomodoro timer, and even a little game I can play to take a break. Before building anything, please ask me any questions you have on this idea so we can build a fully fleshed out app."
I'm going to hit enter and it is going to first start asking me any questions it has to develop this idea. This is not only a good exercise to get the AI up to date on what you want to build, but also it's a good exercise for you to kind of think through things to fully flesh this out.
"So, where do we want to build this?" Let's make this a web app. "How should your data be stored?" Let's make it local only so it's nice and easy. "What how connected should the features be to each other?" Definitely deeply integrated. "What kind of break game?" Sounds fun. Let's do a classic arcade game. I like that.
And now it fully fleshes out. And we're going to submit our answers. Now it's doing more thinking and now it has even more questions for us. Typical plan mode, just a few questions then it builds. Here we're asking to ask as many questions as it needs to get this working. "What visual style do you want?" Clean and minimal. Playful and colorful. Retro terminal. I kind of like the idea of playful and colorful. Let's do that. "How powerful should the note-taking be?" Let's do a daily journal style. "What should the calendar handle?" Definitely task and events and anything special for the Pomodoro timer? Let's do attach it to tasks as well as stats and streaks. I like that as well. And then let's hit submit and we'll submit answers there.
All right. Look at this. It gave me an entire massive command uh to build this out. Let's do this. We're going to copy and paste this. It has everything in it. And then what we're going to do is `/goal` and then paste that in and hit enter. And then now what it's going to do is put it in a loop to build out this entire thing. So what's technically going to happen here is now the agent will go start building this app out and every once in a while check back in. "Is it accomplishing the goal we gave it? Is it has it hit the goal in the success criteria in our prompt?" This is going to ensure it does not stop working until the success criteria is met of what's inside the prompt. And because we did the advanced plan mode, it has a really comprehensive view into what we actually want built. That's the key to doing any of these advanced tasks is the AI. No offense to you guys. I mean, you're really smart cuz you watch my channel, but the AI at this point probably smarter, so you should lean on it to do any sort of thinking and planning, right? You giving it a command to do something advanced. It probably can command itself to do that advanced thing better. So, you want to say, "Hey, how would you do this? What's the best way to do this? Can you plan this out for me?"
All right, looks like it is all done. It is built, verified, and running and named it Fable. I think maybe uh Anthropic might be feeling themselves a little bit because the model's named Fable and it's now naming all the apps it builds Fable. Uh let's take a look at this. Let's see how it's looking. So, I'm going to be honest. I think the UI is a little ugly, but to be honest, I also answered the question to make it colorful and playful. So, I guess it is very colorful and playful. Let's see what we got in it. It looks like it has absolutely everything in it.
So, let's do a quick task. "Edit the Fable 5 video." Hit enter on that. Boom. It's on your plate. That's very very nice looking. Let's check out the calendar. Full calendar. Let's give it an event right now. "Check video performance." We'll give it tomorrow. We'll give it a date for Thursday. We'll give it a color. "Add the event." Boom. There it is. This is a fully working calendar. Let's see. Is our week view. Week view. Can add the task there and the calendar invites there. Love that. Notes. Full note-taking section. Feeling great about Fable there. Boom. Goes into the past entries. That's very nice. Pomodoro timer. Let's start our Pomodoro timer. Okay, the focus. Look, it's got streaks. It's got everything. It's got a full tracker of how many times you use the Pomodoro timer. Let's check out the arcade. All right, it is a snake game. That's a good looking snake game. Makes the snake longer when I eat the strawberries. I don't know if the strawberries are a little dig at OpenAI there, but there we go. Let's see what happens when we die is going in there. Boom. Adds our score. Canban board. There's our task. Can we click and drag it across? It did everything. It built absolutely everything we asked for. Zero bugs. It just works. Works. The tasks are there. Everything's there. We have the nice dashboard.
The key is goals. The key is loops. If you are new to goals and loops, do not worry. You go into Claude Code. You ask, "Hey, I want to put this in a loop. I want to make a goal statement. Please tell me what goal to make or what loop we should be using." And it will give it to you.
I'll give you a quick little bit of an advanced workflow as well. If you're using other tools, other project management tools like Linear like you see here, Claude can plug into it and an interesting loop we can do is we can say something like `/loop` and then every we'll say one hour check to see if there are any new Linear tasks you can take. And now what it will do is every hour go into our Linear our project management tool and see if there's any coding task it can take and use. That's a very strong way to use loops in here as well. This is how you need to be thinking in loops because these models are so smart and powerful.
Now one last interesting point here. By the way, if you learned anything at all, make sure to leave a like down below, subscribe, turn notifications. All I do is make amazing videos about AI. One last thing I want to talk about real quick is this is only available in subscriptions for the next two weeks. On June 22nd, they take it out of the subscriptions and you have to pay via API. It is double the price as Opus. It's actually a fraction the price of ChatGPT 5.5 Pro, but it is double the price of Opus. This might be the beginning of the end of subsidized subscriptions. That's something interesting to watch here. This is the first model a Frontier Lab has released that's not available in subscriptions. So, something interesting though, if you are subscribed at the moment, make sure to test it out. You have a couple weeks to do it. Then you'll have to pay via API usage. That is Claude Fable 5. Let me know down below what your experiences have been so far and what you're building with it. I'm very curious to hear. More videos coming on using it with Hermes and using OpenClaw. Subscribe and turn notifications for that and I will see you in the next.