Anthropic just dropped the best model ever... (Fable 5.2)

0:00 This is the best model I have ever used by far. Today I am going to show you some of the most incredible demos that I have ever been able to create. And when I say software is dead, this is what I

0:14 mean. This is a Gears of War clone. It is a fully playable game. Look at this. Look, you can hide behind barriers. The lighting, the textures, everything was created by this new model. It took just

0:27 a few days to build. Okay, so we come out to the street. We can hide behind here. Pop out. I mean, this is a full battle happening right here. The AI works incredibly

0:39 well. Let's see. Reload. It has that timed reload just like Gears of War. I mean, it's not easy either. There we go. Oh, dang. I This is the first time I'm seeing this part. All right, so the

0:50 reach over the barrier looks kind of awkward, but everything else is flawless. And again, like all I told it to do was make a Gears of War clone to do one level. Yeah, obviously I'm not

1:00 very good. Okay, so I died. He's going to revive me hopefully. There we go. Oh my god, I'm getting crushed here. I think I need to reduce the difficulty. I think I'm supposed to be recording a

1:12 video right now, but I'm having so much fun. I really, really want to beat this level. This is basically one-shotted. Here's running. So easily we can run around. This is actually fun to play.

1:24 All right, I got to stop. I I can't get over how good this is. It built a single level, and yes, it used a ton of tokens, but I'm talking about making a AAA game in a matter of a few days. Okay, so this

1:39 is the actual prompt that I used, and it's only three lines long, which makes it even more wild. I like under specifying my prompts. So, check this out. Build the most accurate clone of

1:51 Gears of War Eday using Unreal Engine. You only need to build one level, but it should look and feel as complete graphics, gameplay, and characters as the real Gears of War Eday that was just

2:01 released. I put it on Ultra Code mode and just let it go. And this took 8 12 billion tokens so far, and it's still going. And if it were based on API pricing, it would be about $15,000.

2:16 Okay, so let's say you're not into making games. No problem. Here is a full feature complete clone of Microsoft Excel. This is literally the most feature complete Excel clone that I've

2:29 ever built. So in total, the Excel clone took almost 50 hours and it is literally perfect. I cannot find a single flaw nor a missing feature. First of all, all of these are actual formulas and I can

2:43 click into any of them. And by the way, this is browser based. So, I'm going to publish this and I'm going to drop a link down below so you can actually test it. It has every single feature that you

2:54 would expect from Microsoft Excel. Look at all of these here. I can even click here. I can load up this example sales tracker, right? I mean, it has all the orders. This is all fake data. And then

3:07 it has a pivot table, like a fully functional pivot table. Everything just works. And honestly, from all of this testing, I am really tired. But I'm not worried about that because of the

3:17 sponsor of today's video, eight sleep. I am so excited to tell you about the partner of this video. It is eight sleep. And if you can only improve one thing about your life and your health,

3:28 it has to be sleep. And I have struggled with sleep over the years, whether it was insomnia or just poor sleeping. And I've been a paying customer of eight for two plus years now. And I actually

3:41 reached out to AdSleep to work with them because I really wanted to tell you all about it. The Pod by AdSleep is an intelligent sleep system that includes a hub and a mattress cover. It uses

3:51 biometric sensors in the cover to read how you're sleeping and will either cool you off if you're overheated or heat you up if you're freezing. Basically optimizing for your sleep throughout the

4:02 night. And as someone who has had sleep issues for years, it not only helps me sleep better, but it helps me understand my sleep as well. It has a really nice sophisticated app and a sleep is giving

4:13 my audience up to $200 off of the sixth generation pod. So, go check them out. I'm going to drop a link down below. Make sure to let them know I sent you by clicking that link. And don't forget,

4:23 use the promo code Matthew Berman. And you can see some of the token usage, yes, was absolutely insane. We have 8 billion input read from cash and about 15 million output tokens. So, that is

4:38 about $11,700 in API costs. All right, here's another one. This is a full Da Vinci Resolve clone, and it works perfectly. And you know what's kind of crazy? Alex from my

4:50 team also did this separately and has been editing videos using his vibecoded video editor. So, here I already have a file dragged in. This is called Kestrel. And here you can see the waveform. Here

5:05 you can see the video. Here you can see I can easily split the clip. I can move it around. Do whatever I want. When I move it next to another clip, it kind of magnetically moves it right next to it.

5:16 It's actually quite nice. I can select both of these, move it back here. No problem. This is a full video editor and it really does work and it has every single possible feature that you would

5:29 expect from Da Vinci. And I feel like I keep saying this, software is solved. There's no reason anybody can't make this. Look at all the different features, everything under all of these

5:39 menus, they all work perfectly. You have full keyboard shortcuts, everything. I can also do zooms very easily. I can insert key frames very easily. All of this is just flawless. Kestrel doesn't

5:53 live in the browser. It is local on my machine and it was built for a Windows PC, but obviously we can easily make this for Mac as well. And I could probably even make it browser based.

6:04 Okay, we're not done. I know I've done a Fall Guys clone a number of times and they've been pretty darn good, but now this is another level and I'm publishing it. You can play it. Let's check it out.

6:17 This is called Tumble Beans. This is by far the best feeling game. Like the physics of how these little beans jump around, how they bounce off of each other, the complexity of every level.

6:28 So, we brought Alex and Brian in from the team. We are going to be playing this live right now. Multiplayer. And again, this was completely built by the Claude model, and I'm going to destroy

6:38 them. You're going to see >> loser pays 100 to me. >> Click the ready button, Brian. >> Oh, Matt, this is your worst game mode. Gate Crash.

6:46 >> Well, this is everybody's worst game mode. This is impossible. >> Matt, get get your own door. >> Hey, who's in second place right now? Oh, look at this. I went down the side.

6:56 First place, guys. Shoot. That was the perfect mode. >> Yes. >> No. Matt, you and I both just hit that. >> I'm so in. Wait, why is right click

7:06 jumping? Come on. Come on. Come on. >> Oh, Brian, I just bounced off you. >> No, the door. >> Come on. Four out of five are real. >> Brian is crushing us right now.

7:16 >> No, I fell. >> Oh, I thought I was not expecting the the floor to fall out like that. Come on. Wait. Wait. Why did I bounce so hard? Oh, the the bounce knock back is

7:27 crazy. >> Go, go. OH, >> COME ON. I CAN'T. >> I GOT TO QUALIFY. OKAY, I'M GOING TO

7:33 QUALIFY. >> NO, I WENT so hard at the very end. Come on. Come on. Come on. Come on. >> Oh, I'm in.

7:41 >> Oh, thank you. >> All right, we're going to have to play this for real later. So, we have a bunch of different levels, a bunch of different game types. So, it's not just

7:49 like beginning to end of a level. They have that spinning log thing. I mean, this is truly fun. And here's that prompt. Build a Fall Guys clone. Full parody. Make sure the levels are fun,

8:00 complex, and different. Build 10 levels. And then the only extra thing I added was multiplayer. And it did so flawlessly. So, this took over 9 billion tokens to build, about 71 total hours of

8:12 build time, and an estimated cost of $13,000. But we're talking about a sellable game. This is something where if you have an idea for a video game and you want to actually build it and build

8:24 it somewhat quickly and go monetize it, you totally can and it'll be fun. All right, next is a 3D simulation of a bullet going through a balloon filled with water. Watch first how good it

8:38 looks, how realistic it looks. Boom. There it goes. The rubber of the balloon evaporates. The water falls to the ground. We can move it around any which way we want. We can go through. We

8:53 can back it up. You can actually see the bullet entering. You can kind of see it right through the balloon skin. And then boom, it pops out the other side. And even the water looks phenomenal. Now,

9:06 from within that, we can actually change any of these settings. We can change the bullet speed, the caliber of bullet. Here's a 9 mm bullet. Let's see what that looks like. And every time you make

9:17 a change, it recalculates the physics in real time. We can change the sun intensity, sun direction. Okay, you can actually see the reflection of the sun on the balloon. How incredible is that?

9:31 And then look at this. As the sun is going through the balloon because the balloon is not blocking the sun completely, you can actually see a little bit of the purple color through

9:41 the shadow. So this is one HTML file with JavaScript, HTML, CSS, and WebGL2 for the graphics. No frameworks, no libraries, no external dependencies. Make a ray tracing simulation of a

9:55 bullet piercing through a water balloon. Make the making the balloon pop and the water burst, deforming from its original shape and eventually falling to the ground. Make it look as realistic and

10:04 physically accurate as possible. Allow me to freeze the animation at any point in time and view the scene from any angle. Add sliders for bullet speed, balloon size, lighting, gravity, and

10:14 other parameters. Most important, do not use 3JS or any external web libraries. I want to test your physics understanding. So, code this completely from scratch. So, this one was actually pretty

10:26 lightweight. Only 41 million tokens in total. It took about 2 and 1/2 hours and an estimated cost of $91. Okay, next. And this was inspired by RPN. I wanted to see how well this new model can

10:39 design a 3D printable tower to withstand the most weight possible. So, I gave it a bunch of parameters. I told it I have the Bamboo Lab P2S. I said4 mm nozzle PLA filament. And here are the

10:54 constraints. It has to be under 10 g. It has to fit within the printer. Obviously, for that 3D print, it took about 10 million tokens. It took a little bit over an hour and cost $37.

11:05 This is what it finally built. Let me see how much weight it can actually withstand. All right, let's start stacking a bunch of tech. I have an Asus Pro laptop. That's about 6 lb. I have an

11:16 iPad Pro. That's probably a couple pounds. Here's one MacBook Pro. Here's another MacBook Pro, a Mac Mini. Here's an Nvidia DJX Spark. Here's my little remarkable device that I use during all

11:30 my interviews. And for the finale, a power drill. And there it goes. All right. And next, motion graphics are completely solved. Like, this is by far the best motion

11:43 graphics I've ever had. And these were all single prompts. Okay. So, this is how software decompiling works. This entire thing, end to end, was created by the new cloud model. Look at this. This

11:54 is a minute and 52 seconds put together cleanly. And so not only does it have a complete understanding of what software decompiling is, but it also knows how to explain it easily and how to show the

12:08 visualization of the explanation. So it is just so impressive. So for the software decompiling motion graphics, that took about 9.5 million tokens in total, about 45 minutes and about a $25

12:22 cost. I also asked it to create motion graphics for an explainer about how LLMs work. And once again, just perfect. So here's just one part of it about embeddings. Here's transformer layers.

12:35 Look at this. Here's training. And you can actually see how training works and the loss curve, the training steps. Here's next token prediction. The cat sat on the next token mat. So yeah, I

12:46 mean we're going to be using these in a lot of videos coming up. I'm very excited about this. Previously, this would have taken so long for our team to create. All right. Next, I had to design

12:58 a gorgeous house, an architectural masterpiece in Blender. And you can actually see that here. And if it looks a little bit wonky, the textures have not been applied yet. And here you can

13:08 see it's actually populating with all of the textures. Now, I can hear my computer spinning up. So, this really was built. And here you go. You can actually see everything. Here's a

13:17 beautiful pool. This looks like the master bedroom. We have solar panels on top. A nice little terrace balcony over here. Here's the garage. And yeah, everything is beautiful here. Let's zoom

13:31 out. You can see this is the driveway. A tree in the middle of the driveway. So, there's one little mistake there. But yeah, I mean, overall, this is gorgeous. And here's the final render. Here's what

13:41 it actually looks like. I mean, it is beautiful. You can see the lighting in the pool, the reflections. Here's the reflection of this little lamp sitting by the pool. Unreal. Here's another

13:52 version of it. So, you can see inside the house. Everything has nice textures. The lighting perfect shadows look great. Even the glass. You can kind of see that this is a piece of glass right here.

14:04 Here's another glass towards the bedroom. A sliding door. And then the final one. So, now you too can build your dreamhouse with just a few prompts. A very lightweight prompt. Build a 3D

14:15 render of a beautiful house in Blender. make it so it feels like a real estate developer's dream. And so for the house rendered in Blender, we have 13 million tokens total, an estimated cost of $30.

14:26 All right. And next, I had it recreate the Roadster event from just looking at pictures and videos of the actual event. So, let's see it go. Ready? Oh my god. This is okay. This is maybe

14:42 the first one that is like just okay, but you know, decent. No, no, it's not that good. Never mind. This is actually the first bad one I've seen. I had to design a 3D European town and I wanted

14:57 it to populate it and make it feel real. So, not only does it have animals, it has people, it has cars, it has traffic, the people have full conversations. And then last, I had it create a video demo

15:10 of the entire thing. Let me show you that now. So, here you go. Here's the cows. A nice pan over these initial buildings. You can see it looks like a very old European town. We have a town

15:22 square. Bunch of people buying stuff, walking around. All of the interactions, all of the people in the town were all powered by AI. I had it basically create what it thought each of these citizens

15:34 of the town would do. And one last little Easter egg. Watch the car that passes by. very futuristic. So that is one thing I asked it to do. I said make everything

15:45 kind of old school European but very futuristic cars. Okay, here we go. You can see these two AI agents having a full conversation. You can see the dog wagging its tail right there. This woman

15:57 decided she wants to see what's at the cafe. All of these people are having a conversation. This woman just walks right through them. And this guy is checking out the sheep. I don't know

16:06 what he's doing here. Next, I had it create a full 3D visualization of Tokyo in Unreal Engine. And this is actually the first time I'm looking at it, so we're going to react to it together.

16:17 Now, this took 6.72 billion tokens, about 70 hours, and would have cost around $11,000. Let's see how it looks. Oh my goodness. Look at how complicated this is. Look

16:32 how good this looks. Wow. Okay. All right. just created an entire city with traffic. Look at all these people just walking around. I mean, obviously Tokyo's massive and it was just able to

16:45 do it seemingly perfectly. And so, let me even just try to zoom out more. And there we go. Look at this. Oh my. This is so so crazy. Here's the train station. You can actually see the

17:03 trains coming in right now. Here's all the cars, the traffic you can see. Beautiful. And look how far it goes. There's that famous tower right there. Oh my god. It has ray tracing. It

17:19 is so good. Look at the reflections in the building. Every single one of these buildings was built by scratch by this new cloud model. Alex and Brian from my team also had early access. They've been

17:31 building so many incredible games. Let me just show you brief previews of them all and we'll drop it down below so you can play them. Here's kind of a Call of Duty but block style and again

17:42 multiplayer and it just works. Here's Melee which was decompiled and he ported it over to his Android phone so he can play Melee on his phone anytime he wants. And here is a Portal clone. This

17:57 one was pretty good. I think it leaves a little bit to be desired with the graphics, but again, all of this is basically just one prompt and it's available on the phone. Here's Starlane.

18:07 Okay, this is basically like an F0 kind of game. And let's see it. Oh, this is pretty awesome. Wo, the graphics are really good. Oh my god. Okay, this is awesome. Wow. And all

18:22 the narration is generated by AI. Oh, that's so cool. Oh my god. Okay, so this is brought to you by Brian. This is Tidekeeper. This is the second version of the same game that Brian had made

18:34 with a previous generation model. This one seems really fun. The actual control of the crab is a lot of fun to play around with. Once again, we will drop links down below. And last, this is

18:44 called Dragon Rider. Oh, that's really cool. Wow, the graphics are really good. This is actually the first time I'm seeing this. Oh, cool. And so in every possible way,

18:57 this model is the best model on the planet right now. Not only the benchmarks, which I'm going to show you in 1 second, but also my own testing, and I have probably spent over a 100red

19:09 billion tokens testing this model. So, first the scores. We have all the major benchmarks on the left here. Frontier code 2.0. It absolutely dominates all the other models on the list, including

19:20 Opus 5.5, which was my favorite model. We have Frontier SWE coming in at 69.9. Again, just dominating the list. Here's Carbon. Here's Astra. Not too far behind, but still behind. The only

19:32 benchmark that Carbon wins at is Automation Bench at 53 and Fable 5.2 is right behind it at 51. Then, of course, we have Terminal Bench Science and Mythos 5.2, which also came out today,

19:45 scores a 75.3, just barely beating Fable 5.274.6. Now, this is a very important benchmark OSWorld 2.1. This is the model's ability to control an operating system. And so,

19:59 that is computer control. And Fable 5.2 scored an 89.3, absolutely dominating what I thought to be the best computer control model on the planet, Astra, at 79.5.

20:13 Opus 5.5 surprisingly scores much higher. I actually thought Astra was much better than Opus 5.5 at computer control, but according to this benchmark, it is at 87 compared to 79.

20:23 Here's humanity's last exam getting a 74.2, 10 points above Fable 5.1, a little bit less than 10 points above Opus 5.5. So, these are phenomenal scores. And besides the benchmarks, I

20:37 have tested it thoroughly for a little while now, and I can definitively say it is by far the best model I've ever used. And so I pointed it at every possible use case I can think of and you saw that

20:51 earlier. But there were a few other things that were surprising to me. This is what I noticed when I was testing this model. Number one, it was extremely autonomous. It went out, it made

21:01 decisions on my behalf and nine times out of 10, maybe even more, those decisions were right. That is what I would have done and it was just one less thing that I had to think about which

21:12 was really nice. And of course, I tested it at every thinking effort. I also tested it on Ultra Code, which basically fans out a bunch of sub aents, and yes, it is extremely expensive to run Ultra

21:26 Code. So, I really don't recommend it. But I put it on max thinking effort and Ultra Code and just pointed it at every single codebase that I had and I just said, "Do a full code review." And there

21:37 were a few surprising things. one, it found so many different bugs, security issues, little UI inconsistencies, but not only that, it finished really quickly. Previously, when I ran Ultra

21:50 Code, it just seemed to take forever. Max Thinking Plus Ultra Code, even though it was running in parallel, even though all of those agents were going at the same time, it still felt like it

21:59 took forever. But now with this new model, things that I previously would have thought would have taken, you know, 60, 90, 120 minutes take like 10 minutes maximum. And I think just overall the

22:12 speed of this model is super impressive for a model of its size, of its caliber, for it to be this fast. And I'm not strictly talking about tokens per second, which is actually kind of hard

22:25 to tell because all of that is pretty obuscated away in cloud code. But I'm simply talking about its ability to take a task and run it through completion. And it just seems like it knows the

22:37 straight line path to getting something done. And so I love this model. Now, am I going to use it over Opus 5.5? That is the question. Opus 5.5 to date is the best model. It's the one I've been using

22:50 on all my major projects. But Fable is better. Without a doubt, Fable 5.2 is better than Opus 5.5. However, Opus 5.5 is very efficient. It is very fast and it is much less expensive than Fable

23:05 5.2. And so, speaking of pricing, let me tell you what that is. They actually brought down the pricing slightly of this model by about 20%. Now it is $8 per million input tokens, $40 per

23:17 million output tokens, which again is a price reduction from Fable 5.1. Now, if you want to do fast mode, which I didn't have access to, but I can't even imagine how amazing that would be, you get 2 1/2

23:30 times the speed, but yes, you're paying a lot more. $16 per million input tokens, $80 per million output tokens. And something that Anthropic isn't known to do is they're actually giving out

23:42 Fable resets this week depending on first of all if you are subscribed, if your account would have access to Fable and how long you've been subscribed for. So hopefully we all get some resets. And

23:54 now it's no wonder that Meter just put out that report that Anthropic has a straight shot to recursive self-improvement. If they are releasing Fable 5.2, imagine what they have

24:05 internally. Imagine how good that model is and imagine how many tokens they're throwing at it to get their next generation of model even better. And I covered that entire report in this video

24:17 right