Claude Fable 5 a prova: programació, 3D i agents
Bijan Bowen prova Claude Fable 5 amb un sistema operatiu web, C++, Python, simuladors, visualitzacions i una interfície d’alta gamma.
Claude Fable 5 arriba amb una promesa excepcional: ser el model general més capaç d’Anthropic i mantenir feina autònoma durant més temps que les generacions anteriors. Bijan Bowen el sotmet a una bateria que va més enllà dels benchmarks: un sistema operatiu al navegador, jocs en C++ i Python, visualitzacions de xarxa, escenes 3D, un simulador d’impressora, una interfície d’alta gamma i una bateria musical.
La prova és útil perquè combina tres coses que sovint es confonen: generar una aparença convincent, escriure codi que s’executi i corregir-lo fins que el producte funcioni. Un model pot destacar en una d’aquestes dimensions i fallar en les altres. Fable ha de demostrar criteri, no només velocitat.
1. Un model Mythos obert al públic amb barreres específiques
A 00:43, Bowen presenta Fable 5 com una nova categoria dins Claude. La fitxa oficial de Claude Fable el descriu com el model general més capaç d’Anthropic i el recomana per a treball ambiciós, asíncron i de llarga durada.
El repàs tècnic de 02:15 situa Fable al costat de Mythos 5. Tots dos comparteixen el model subjacent, però Fable incorpora classificadors per limitar usos de ciberseguretat, biologia, química i destil·lació. En moltes consultes marcades, la resposta passa a Opus 4.8. Mythos conserva més capacitat en aquests dominis i queda restringit a programes d’accés de confiança.
La presentació d’Anthropic publica un preu de 10 dòlars per milió de tokens d’entrada i 50 per milió de sortida. Per a un agent de programació és car per token; pot ser eficient per tasca si redueix les iteracions, comprova el resultat i evita hores de supervisió.
2. El sistema operatiu web examina estat, interacció i verificació
La primera prova comença a 04:47. Una versió del prompt demana deu aplicacions i eines de xarxa defensives. El classificador la redirigeix a Opus 4.8, fins i tot després d’una espera llarga, de manera que Bowen descarta el resultat. És una demostració directa que els filtres també afecten encàrrecs legítims.
Amb un prompt genèric, Fable crea Nebula OS: finestres, terminal, pintura i dos jocs 3D. Minimitzar, ampliar i tancar funciona; el joc de la cambra té mapa i una representació visual notable. El conjunt, però, no incorpora menú contextual, repeteix estils habituals i la tela de dibuix no s’adapta quan canvia la mida de la finestra. Bowen el qualifica com un resultat mitjà amb dos jocs bons.
També prova una versió via OpenRouter amb un clon de GTA més convincent, però no pot assegurar que el servei no l’hagi canviat a Opus. Aquesta transparència és important: sense saber quin model ha respost, la demostració deixa de ser una comparació vàlida.
3. C++, SVG i paquets de xarxa obliguen a canviar de registre
A 09:55, Fable rep un joc de monopatí en C++ i treballa amb esforç màxim. Sense utilitzar Raylib, construeix una ciutat costanera amb col·lisions, trucs, moviment fluid, càmera, vianants, botigues i una persona que camina amb una taula de surf. Bowen el considera el millor resultat de C++ i de monopatí que ha obtingut d’un model.
La tasca SVG de 12:32 combina actius, animació i un guió publicitari llegible. A 14:31, una rèplica 3D de l’apartament de Seinfeld ofereix bons materials i objectes, però no respecta del tot la planta i falla el moviment amb WASD. La qualitat visual no amaga el criteri incomplet.
A 16:15, el model connecta un backend Python a una carretera 3D que representa trànsit real: autobusos per HTTPS, cotxes per TCP, furgonetes per UDP i policia per ICMP. Els vehicles es poden inspeccionar i un ping genera els corresponents cotxes policials. És una visualització tècnica funcional i pedagògica, no només decorativa.
4. Les escenes i els simuladors separen la maqueta del programa
A 19:16, Fable col·loca la primera escena de metro dins un visor de Claude en comptes d’un fitxer reutilitzable. Convertida després en joc, queda neta i funcional, amb enemics que entren pels accessos, tot i una aparença estèril. A 21:47, un joc 3D en Python adopta una estètica synthwave no demanada, però integra mapa, so, retrocés de l’arma, traçadores i regles coherents.
El simulador d’impressió 3D de 23:11 és encara més complet: bobina que gira, tub Bowden unit al broquet, llit calent que canvia de color, capes i patró de farciment. No tanca la part superior del cub i no deixa ajustar la densitat, però la mecànica principal funciona.
A 24:49, Bowen redueix l’esforç al mínim per al simulador de vol. Amb només 460 línies, Fable implementa controls, inclinació de càmera, traçadores, acceleració, col·lisió i diversos avions. La prova mostra que no totes les tasques necessiten el pressupost màxim.
5. Una interfície d’alta gamma i una bateria tanquen la prova
La prova de frontend de 25:58 demana una web de rellotges amb renders d’alta gamma creats sense actius. Fable modela el rellotge, la corretja amb costures i una vista 3D controlable. La secció inferior té colors pitjors i frases tòpiques, però el producte principal capta amb precisió l’estètica demanada.
A 28:43, la bateria virtual està ben distribuïda en l’espai, respon al teclat, admet reproducció automàtica i diversos ritmes. Hi falta un pedal de bombo clar i no es pot ajustar el tempo, però so, animació i interacció funcionen plegats.
Bowen reuneix els resultats a 31:25. El joc de monopatí i la web de rellotges són els dos salts més evidents; l’apartament i el sistema operatiu són els punts febles. Després de diverses tasques paral·leles en modes alt i màxim, ha consumit el 49% del límit de cinc hores del pla de 200 dòlars.
Les afirmacions de capacitat i seguretat d’Anthropic es poden consultar amb més detall a la System Card de Fable 5 i Mythos 5. El document aporta proves àmplies, però continua sent material del fabricant; les sessions reproduïbles completen aquesta evidència amb errors reals.
Conclusions
Fable 5 representa un canvi d’expectativa per a la programació amb IA. El valor ja no és escriure una funció ràpidament, sinó sostenir una feina, utilitzar eines i acostar-se a un producte executable. El C++ i la visualització de paquets semblen d’una categoria superior; els errors del sistema operatiu i de l’apartament confirmen que el model encara no és infal·lible.
El model no elimina la revisió humana. La desplaça cap a l’especificació, els permisos i l’acceptació final. Per aprofitar-lo, un equip ha de preparar proves, executar el codi en un entorn segur i mesurar cost per entrega correcta. Si Fable redueix de veritat els bucles entre idea i resultat verificat, el preu pot tenir sentit; si només produeix una maqueta més polida, no.
Contrast i context
Fonts consultades
- 01
-
02
Anthropic Claude Fable
-
03
Anthropic Claude Fable 5 and Claude Mythos 5
- 04
Font de treball
Transcripció amb marques de temps
Consulta la transcripció
-
0:00
, obre el vídeo en una pestanya nova
It's getting insane in the title from this specific result because this didn't use RayLab. I genuinely... Today we're going to be taking a look at what may be the best model in the world. Now, I'm not going to be that dramatic, but I might be. So this is of course the new and thropic Fablay 5 model that has just been released. I know it's called Fable 5 and this is really exciting because there has been talk so much about anthropics mythos model and this is a mythos class model. So Fable 5 is kind of the normie version of mythos 5 if you will to put it kind of simply that's probably the best. that I can think to explain the difference between the two of these just on first glance. So to begin, let's just take a quick look at some of the interesting things in the announcement post here. Please do feel free to subscribe as I want that 100k plaque. That is one thing mythos can't get me. Probably could, but I want it like in the proper way. So, Claude Fable 5 and Claude Mythos 5. Now, Fable is a name that was not commonly thrown around, basically until like the last few days, when the model slug or just the way or name to access the model, started appearing publicly, so folks caught that. This is essentially the consumer version of... mythos class model. And what's meant by that is it has a lot of safeguards implemented in it to ensure that folks are not going to use this for malicious tasks. As this is touted as a very, very capable, perhaps the best currently existing publicly accessible model. And the fact that it's listed here in conjunction with mythos 5, we can think of it simply as Fable 5 is almost like the norm version of mythos 5. And it's listed right here that for a small group of cyber defenders and infrastructure provide. They are also launching mythos 5. They are the same underlying model, but that has safeguards lifted in some areas as opposed to Fable 5 which does not. Now pretty prominently listed here in the beginning of the announcement post is the price. So it's shown right now as $10 per million input tokens and 50 per million out which don't get me wrong. It is expensive. However, it's a lot cheaper than a lot of folks were estimating this was going to be. I'm even surprised that this is not over $100 and based on the
-
2:11
, obre el vídeo en una pestanya nova
purported capabilities of this model, it seems to be quite a decent deal. Now of course we have the benchmark JPEG which... is not a huge leap. Oh, no, okay, yeah, apparently it is. I was just looking at Claude mythos preview here, which was not publicly available to anybody. So basically, this should almost be next to these three models right here, because the leaping capability between the highest performing, at least based on this benchmark, which was Claude Opus 4.8, and mythos, or Fable 5, is significant. Additionally to that, basically, it is unscientificly speaking, putting the Mac down on every preceding model. a fairly high degree, which makes me quite excited for some of the tests that we like to run like the Browser OS, which I do have already loaded up right here, and I'm very eager to run it. I have not used this model yet. A big focus here, or at least what is prominently mentioned, is also these models being able to work autonomously. Now, it scores highest among frontier models, even at medium effort. I do like to see that. They also do have some nice demonstrations of capabilities here. In specific, I find this very interesting because 3D CAD design is something that I've tested. models with before. We did a dedicated video where they had to design an enclosure in OpenSkad. And let's just say it was kind of a disaster. It was entertaining though. So seeing this, even this looks very simple, this little lighthouse here, this is significantly more competent than I have seen in the past when using models for CAD design. So make of that what you will. Also they have a music related demonstration here, which makes me excited. Now additionally here and something I don't really commonly seen in model announcement posts even state of the art ones. They're talking about things here like drug design, biology, molecular biology, things like this. Truthfully, things that I have basically no... knowledge of in terms of any of this, but I find this stuff very cool because a lot of what folks talk about, at least in terms of the good potential in AI, is things like this, designing drugs, perhaps finding cures to novel diseases, or maybe not novel diseases, but finding cures to existing diseases should have said novel cures to existing diseases. And this stuff's just very interesting because this is a lot of what AI's power for good.
-
4:25
, obre el vídeo en una pestanya nova
like has its potential rooted and I guess could be said. So at the end here they just talk about the pricing once more and if it's possible that this is just the static pricing which would actually be pretty darn cool because that's not a seemingly expensive for the... capabilities of the model as we saw in the benchmarks. They also do list information here just about accessibility and how it will roll out things of the sort and with that because I really am genuinely excited to get started here. We're going to begin this just on high mode from within the web chat interface using the browser OS test. However, this is one that has a slightly tweaked prompt because I want to specifically go into some of the strengths of this model. So Basically, this is the same browser. Okay, I knew this was going to happen. It switched to Opus 4.8 because basically I added in here. I changed it to 10 applications, but I also said they must be white hat tools that can show info about the current network environment system, et cetera, and we'll run the more generic one. Something I want to mention as well that I'd forgotten to is usage limits for this. So this is the $200 or the max $20 plan. As we can see right here, we're starting from a completely blank slate on this $200 a month plan. So I just want to show that so we can return back to this at the conclusion of the testing to see how much is used up. I would hope that we use all of this up and then have to transition to API. But this is definitely something I also wanted to specifically note for this test. It switched to Opus 4.8 after that long. Okay, so after quite a while, it switched to Opus 4.8. So this browser OS result is null-in-void. Again, we'll just see if we can get something here that hopefully doesn't switch to. Here is our Fable 5 High Browser OS. Nebula OS first and foremost alright not bad. I've seen this UI style before very commonly seen it has a nice pretty background now keep in mind this was not able to be done with the specific GTA clone result however that's quite alright because we did get something and this is the first result from Fable 5 on high which is exciting I think so first and foremost is there a right click? No, there's no right click which is fine. Maybe it just shows instruction following is good because it's not listed in the prompt Alright, minimize to the task bar and then reopen. Good full screen. Very good and then finally close. And I do like the dock style here, although I could use some icons zoom when you hover over them and we do have a clock in a date in our location. Let's just run down the line of these applications. Hyperdrift pilot through the asteroid gate field arrow keys, WST to steer. You know what? Oh yeah, the whole breach fly again. Okay, it's impossible to play, but it's good.
-
7:04
, obre el vídeo en una pestanya nova
This is difficult. It did a really nice job of like a This is just like Maybe the mythos 5 model could play this but no human be able to Am I supposed to hit the green ones or those like help things? Yeah, I would alright. Yeah, that's okay. All right that worked and again You know, we're just gonna keep going vault 3G You locked inside the vault. We need to reach the Emerald door. Okay this oh oh wow. It's The way this is being visually shown to us is very very cool. We do have a mini map up in the top right? The way it's drawing this and rendering it is really quite something I'm going to say. Funky. Oh we found the emerald door can we go through it? Yeah we tricked the vault break in again. Oh With that link on, I'm surprised it allowed it to complete this result. Now, it was understandable that it blocked us earlier. All right, terminal type help to see what this terminal can do. Do not run matrix. Okay, of Neofetch. Good. And that's that little galaxy thing is pretty, see? Was there anything else in here that was of utmost interest? List-running apps? Cool. Let's do five times nine, 45. Very good. All right, paint application. The background in this is a bit odd. Save this PNG and then okay. Now this is an issue I've seen before. I wouldn't have expected this issue where it neglected to account for the window being made larger so the actual transparent background is a static like in the top left. I've seen that issue before just not in like the best model ever. Okay. Cool, and then we can apply a custom color. I wonder where that applies to, perhaps just the static background, but it keeps the gradient. Okay, okay, overall. the vault 3D game both the games were impressive everything else was fairly cookie cutter and standard so now I did also run this through open router but I was a little worried because depending on the way that this swaps to open 4.8 when it detects something wrong I don't know that that's going to be showcased here in open router so we did get another result here and I asked it following this which model made this so
-
9:29
, obre el vídeo en una pestanya nova
perhaps we have something but take this with a grain of salt. Okay, it's simple. I basically just want to see the GTA clone. All right, that's actually very good. So again, we can't definitively confirm that this was made by Fable 5 because we ran it through OpenRouter and it's possible that it swapped it for the same reason it did. Watch. Okay, now at the same time as the browser OS I do have this in Cloud Code here. It is with the effort set to max using Fable 5. I'm giving it the self-contained C++ skateboard simulation test. I am just really excited to test this. So we may do some more rapid fire tests or perhaps less structured than some of my traditional videos. All right, the skate games done and it made screenshots. So I'm click, oh my goodness, holy shh. It's getting insane in the title from this specific result because this didn't use RayLab. This is either anthropic watches my videos and Benchmaxed it or this model is just like I genuinely Okay, this is this is a new paradigm so to speak. It put mesh colliders on the buildings and we can see in them This is unbelievable. This really is impressive. This is done in C++, not allowing it to use Rayleigh. Look at the actual, so the lean effects of the board and things like this, I mean, the fluidity of the movement. is impressive. We have a lifeguard tower. This is just honestly the even the text on the businesses and things like that, the car there, the purple slab. The person walking, look at the walking animations with the surfboard and things like this. This is unbelievable. I'm genuinely blown away by this after seeing the browser OS, which now was kind of just mid because I guess it didn't really exercise the same degree of creativity here as it did here. This is, I mean, let's just eat it. We can do grabs. Oh, yeah. You and I is to do a grab. This is I mean Okay ours to reset look at this I mean this right here is is Gonna be put on steam for 99 cents so if anyone wants to play beach on skate sim I'll put the link in the description I Look at the birds look at the I mean this is the Ferris wheel on the on the boardwalk down there towel on the beach
-
12:18
, obre el vídeo en una pestanya nova
This model might actually be able to make GTA 6. There's always like the meme make GTA 6. Actually, no, it would it would I mean to blur that out. All right. Do you want to proceed? No, you're done. This was the best result I've received. Probably ever next up we're going to be performing the mid roll ad test where for today's sponsor oxy labs who is kind enough to sponsor the channel and allow me to spend money on tests like this as we're probably gonna go into the API We have given this a folder of specific assets as well as a fairly verbose and large prompt where the culmination of this test is going to be an animated SVG that has also like the script on below it for me to read and we'll take a look at the result we get it should basically make a little nice ad animation using the specific assets in this folder as well as putting the text on the bottom for me to cleanly read and some additional animations and things of that sort. Alright, our mid-roll ad test is completed. Let's see how good Fable 5 is. If you're building AI agents getting real-time data from e-commerce, real estate, or search results is a massive bottleneck. That's where the OxyLab's web scraper API comes in. It provides uninterrupted access to almost any public website using over 175 mil. and residential proxies across 195 countries. I dropped this straight into my NADN workflow because it seamlessly handles JavaScript heavy sites and scales instantly in the cloud. I did put this in NADN workflow. There's a video on the channel. Oxylabs also just launched a new headless browser feature specifically for AI agents plus a fast search API that feeds organic search results directly into your AI applications. Go to oxylabs.io slash Bjane for a free trial. You get a up to 2000 web results for free, no credit card required, use code BGEN for 20% off of all oxylabs plants. And I did put a video on the channel using oxylabs in conjunction with NA Dan to create a little drop-shift item detector and it really was quite cool. So if anyone's interested, feel free to go check that out and also thanks to oxylabs for sponsoring the video and allowing me to just blow money on API credits testing Fable 5. So check that out. The link is right here on the screen. Again, thank you to oxylabs. I did instruct this to do a 3JS replication of Jerry's apartment from Seinfeld, just ensuring that it didn't swap to Opus 4.8 or any false positives, which fortunately it didn't. So let's take a peek at this and see how it did. It worked in the artifact window with the heck. We may have to just look at it through there. All right. Now we have this working in the browser. Let's go around and explore it. Let's just look at some of the materials. Okay, we have the daily times. W-A-S-D should be
-
14:59
, obre el vídeo en una pestanya nova
Letting us walk unfortunately, that's not working, which I'm going to knock this for. I do believe it's maybe possible that we're stuck in one of the objects here if there was a mesh collider. Maybe not, though, so let's look. The bicycle on the wall may actually be the best bicycle model I've seen in this test and you're going to be like, well, be a child, that still doesn't look very good. Yes. So that is saying something, the spokes on the wheels look good. We can see over there to the bathroom. We also see the kitchen area is very nice. And of course a new York City style, skyline out the windows here, which is a good touch. Overall though, the shadows look good, the benches look good. We do have the cereal boxes and things like this. It's not a proper replication of the apartment. Nothing has yet. to date actually been able to get the floor plan from that image, which is interesting. We do have a bathroom over there and things like that. Judging this independently as a scene, it's well done. Judging it based on the test is this is still like an outstanding bounty in all of my tests for a model to properly replicate this. So make of that what you will. The WASD movement not working is a little... upsetting, but in lieu of just wasting like tokens on trying to get it to fix that, I'm gonna probably prioritize just running more tests. We do have a fruit basket on the table and things of the sort. I'm starting kind of a weird test that I genuinely just thought of because I find it may be kind of interesting. So I've put this back on extra high mode and I've asked it, make me an app that sees the packets coming and going from my system. Visualizes them in a 3D way like buses or cars traveling on a highway. Different vehicle pipes for different packet type. So this should be a very interesting visual demonstration or we'll see what we get here. Alright, so we've only used 24% usage. I really thought like we would have used more by now but I'm happy to have not. Alright, so it's gonna build this. It gave us a little bit of info. I'll build this as a small Python backend that captures packets with a simulated traffic fallback streams them to a 3JS packet highway in the browser. First a quick environment check. This is not a hyper-difficult task. However, just seeing how well it goes about intertwining the backend and front end as well as visually displaying these things will be fun. Okay, this is this is sick. These are
-
17:22
, obre el vídeo en una pestanya nova
actual visual depictions of the packets and traffic in the system different vehicle types being animated for different types of packets. So we have TCP, HTTPS ones. Let's see and we even have who's driving. So the city bus is somewhat common that's HTTPS. We have just the regular sedan, which is TCP other panel van is UDP other. I mean this is sick. This is so cool. Okay, police cars ICMP ping. Also, did I just see like a vest spot on the side here? Okay, that's a bicycle. Okay, so the police car is a ping. So let's just do. Let's ping Google's DNS server and see if we see. And he police cars. Okay, I just saw a sports car. Oh my goodness if you click on them you actually see Okay, I can't show some of these cuz like it book the police cars are in the median ICMP pink. This is sick now again. This is not necessarily like a horrifically difficult test to have run It's like we can even pause it look at that. Oh my goodness, and then orbit is oh and that like it's autonomously orbiting. This is so cool. Imagine if schools like taught things this way, people learn so much more specifically with math. Why did they make it so difficult? Sorry, that was a tangent. But I mean, this is like... This is so cool and again, it's not horribly difficult to implement all this. It's the way it will fit together. And the actual density of the amount of vehicles here and the fluid movement is really, really awesome. So if we, I'm positive right here, we see our police cars are pink and they do even have little light lights and things like that. This is cool as heck. And I wanted to just run something that I've never done before. And fortunately it did still work with um... I keep forgetting the models name. I want to call it mythos with this model. That was cool. Next up we're going to try the beautiful static subway scene test. This is a very exciting one. Probably my personal favorite test, at least as of this slice of time. Beautiful subway scene. Okay, so this thing, it's frustrating. Okay, that's really concerning because that's the city I'm in disturbing. It put this in like a
-
19:38
, obre el vídeo en una pestanya nova
weird artifact thing in the cloud UI. So downloading it and seeing it right here is a fault of the way that it was generated. I actually saw it. I mean, see if I can go back and try to find the chain of thought where we saw this. I'm deciding whether to use the visualizer widget or create a file artifact for the subway station. So it I believe put it in the visualizer widget, which is why when we downloaded that it was wrapped in this auditing. I named it Newton after your neck of the woods. That's scary because I don't have any memory or anything enabled with this model. So that's good. Riverside, Boston, yes, all stations close by. Next up, it's going to be like, Claude is at the door. So we're going to have to tell it then like once it turns this into a game, we'll have to yell at it to actually like make this something. Full screen to do a quick usage check we're at 37% used and I'm running things in parallel through cloud code as well as the web interface So with that I'm also going to just start the 3D printer sim just to have it ready. However we do have our subway FPS result now I will just download this and we'll take a look at it in Chrome if I can open it all right Wow, okay, let's see good okay Because that happened with one of the opus models. It was the first time I've actually seen the ammunition. Look at the way the folks are walking and like, oh, okay. Yep. Weird like groaning and moaning noises, but not bad. We can run around. I'll say it's not super visually exciting, but it is very clean. That's something I'm going to notice. It almost feels like sterile, but in a good way. And if we look at this map more now, even the way that be. dots are drawn there in terms of the do not cross line. Okay, we have a lot of enemies here. It auto reloads for us. Oh, they're hugging us. Hug. Oh, yep. Look at the way they're coming out of the entryway. That's actually nice attention to detail. So I've given it somewhat of an obscure Python test where it has to generate a 3D FPS game using Python with the following features, but it cannot use or seen a. And as we can see right here it's building this as a recasting engine FPS in pie game. Wolf and Stein doom classic style true perspective 3D. So it's going to use pie game and we'll just see what we get here. This is just a way to see how cleanly it implements these things. All right. So here's our 3D
-
22:11
, obre el vídeo en una pestanya nova
the FPS game using Python. For a second, I was expecting something else. Okay, why is this so synthwave? I like that, but nowhere in the prompted it's say it should be synthwave, but at the same time it didn't say it shouldn't be synthwave. So hello friend, this is just like, it's very solid. Then there's only so much you can do with this prompt. The sound inclusions are nice area cleared. Good accuracy and time press enter for menu. We can start again the actual ammo tracer's the weapon recoil. I don't know why I'm yelling. The synthwave sounds. The mini map well done. And the hostels are actually pretty well drawn considering the limitations. All the assets are. We have one hostile left behind this wall. Oh. Oh. Good. Alright, here's our 3D printer simulation. That's just unreal. What is the bed say? Layer Lab. This is... Okay, it started printing autonomously. Look at that. Okay, the only issue I see, it's not even, this is nitpicking, is the bed started and it was hitting the LCD screen. It doesn't matter. Look at this. It actually has the... Boden tube, which is what that's called. From, oh, look at the spool spinning tube. It has the Boden tube, and it stays connected to the nozzle. In fill is proper, although that's like a 5% infel thing. The bad looks really good. This is, yes, this is quite quite good. This test has been around for a while now after I pioneered it, but it's still. And it's showing us the actual command stream right here, which is cool. So these are the coordinates that the nozzle and bad are moving together at in conjunction. Going a bit slow, let's just maximum speed this. Oh yeah, look at that. I want to see if it closes the top of the cube or not. Most times no, and it didn't, which is fine. But look at that. We have clean layers. We have nice infill pattern. I would have liked to have seen adjustable infill density if I was to nitpick, however, oh, it's, I mean, it's very, very, it's well done. Triangle. Oh, the bed actually changes color to show that it's hot. That's I like that very, very well done, beautiful. And the school's going through like the top canteries. So if I wasn't an epic but I mean it's good.
-
24:47
, obre el vídeo en una pestanya nova
and that would be the power supply right there. I want to try something. So I've put it on low effort here and I'm giving it the flight combat simulator game because it'll be cool to see how it does in a prompt that's been run so many times. So there's a lot of frame of reference for this specific test, at least on the test I've done on the channel. So let's see how this stacks up in its lowest possible effort mode in this prompt. So this was only 460 lines of code, but let's take a look at our low effort mode flight simulator game. Okay, this is a I've not we don't normally see camera tilt like that good Q&E or for like yaw and things of the sword for low effort mode This is still pretty good look at the ammunition tracer's flying bias There is no sound in this because we would have heard that shift is to speed up all of these pretty much do similar Control schemas for these games Let's just see what happens then if we crash it should good So properly implemented everything and in 460 lines of code just to take a peek at some of the other plane models. Something stealthy. Yep, okay. Cool. All right, so I wanted to just run something on low effort because why not? Next up, I've told it that I need a beautiful website for slap-as-watch company. The site should feature a high end hero section with an animation panning around the watch, which is to be placed on the table. The entire scene needs to be created by you as I don't have the assets for the watch. Following the beautiful hero section, there should be an about section leading to a selection of available watches. There are two currently in line for the summer 2026 collection. All assets must be programmatically created and look like high end key shot renders. Key shot is a very high end and expensive rendering program that gets used for like a lot of things probably that you see day to day. So it should just make all the assets look very very high end. Well it isn't officially done. It's just doing a Q and A check and basically every time it's done that, most things have been very, very impressive. So let's take a look at our front end watch website. Okay. Geez, that's so good because I told it it needs to be that's incredible that is incredibly well done and it 100% captured what I was going for where it needs to look like a keyshot render of the actual device it made this entirely by itself and it rendered it look at the actual like the watch face effects and things like this and if I go hands off the mouse it just spins around but hands on the mouse you can kind of directionally control now I will say this right here
-
27:17
, obre el vídeo en una pestanya nova
Terrible awful like the color scheming and things like that bad, but this more than makes up for that in my opinion because you could just yell at it for that So it's designed for that to be showcased not in the white glow effect So if we look at it from this angle it looks good the watch has a crown it has a strap a watch should not chase the hour it should accompany it. Oh, that's just cringe look at that slip is drag or I mean Yeah, and it made these assets it programmatically created this captured like a static render of it and then implemented it included it in the site. Look at that. The Riviera collection. Caling automatic, doon automatic. This is incredibly well done. I mean, the rest of the site UI, why is it's kind of like cloud generic aesthetic, but it's the assets that it included here that are just absolutely top tier. and the way it went about doing this. If you know, you know, I hate that term, but I feel like it's kind of reasonable here. Even the backside of the straps have different color as they would in real life, unreal, very, very, very impressive here. And the K-Shot render really did properly adhere to what I was envisioning. Even put a fake, fake icon, fake icon, that icon. Don't ask me to pronounce that ever. So the final thing of course is the virtual drum kit sim which I have already run in the web chat I don't really like This is always just fun to run is like the final test I guess so let's peek at it and I do have the speaker on click anywhere to power on the audio engine That's very very good spatially. This may be one of the best arranged ones More often than not they're not necessarily properly arranged. Good. We can move around the scene. We have some artifacts. We have some brass effects and then it's in a nice little circular carpet stage thing. We even have a stool. So let's check the important part. Does it slap? All right, it does. Let me turn the speaker up here. Good effects. Though if I were to nitpick, we don't have a cake pedal or we do and it was kind of like placed in the drum, but you know, it is what it is. Let's check all of these very quiet. I do like that it did not make the symbols overly loud which sometimes happens. Phil Collins test. Oh, I missed that up. Phil Collins test. Phil Collins test. Spawn on. All right. Let's check our auto play feature starting with classic rock. Then we should get like a doo doo. Good, good. All right. Solid. And we have a adjustable BPM. No static BPM. All right. Let's try funk pocket.
-
30:39
, obre el vídeo en una pestanya nova
Alright, let's hear this roll. Okay, more static but good. Unless, okay, disco floor. Yes, 100% disco. Can I play along with it? Yep, good. Then half time shuffle. Alright, very good. Overall, that's going to conclude our first look in test of Anthropics Fable 5 or Mythos 5 with, didn't lack of ability to hack into things. This was very, very impressive. And some things that stood out this watch website, the way it had to do this and the way that it executed it was incredible. We even see individual stitching actually on the sides of the strap here, incredibly well done and at 100% understood my prompt denotion of like the K-Shot style renders. Now, the thing that I was likely most impressed with that really stood out to me as like a completely different... League was the skate game this was 100% the best skate result the best C++ result I've ever seen from a model the way that it implemented the fluid motion the movement the person walking with the surfboard the shops just everything here was absolutely Incredibly well done this is like I've not seen one this good ever and Folks who watched the channel consistently Thank you and you're gonna agree with me 100%. Just the lean effects and the way the camera transitions. Oh, okay. I didn't realize we could go in the sand, but I guess that's okay. Splash. If we go in the water, it force resets us and then the bail effect is obviously funny. Yeah, this was just like I mean This shows true capability either it was benchmarked on this specific test, which I don't think is the case or it's just Truly genuinely very capable, which I think is the case. So that right there absolutely blew me away. Now the thing to note is that the tests that were run from within the Cloud Code interface were done in X-Hi mode. I think I may have swapped one of them to hide, but most of them were X-Hi. The ones that were run from the web chat interface were all run in the standard high mode except the flight simulator one where we ran it in. low, just to see what it did on low mode. Really the results overall were just fantastic. It seems very very very good. I think the worst result we received was probably the sign-feld apartment because it didn't properly really won the movement and worked and two it just didn't necessarily like a
-
33:42
, obre el vídeo en una pestanya nova
match the real scene to the degree I would have hoped it would have, though it's possible if I'd done that on the maximum potential high mode it would have. The subway game was sterile but functional, same. We had a Python game, we had the packet game was awesome where it had to show the packets and then like show them as traffic with different types of cars. It was just insanely insanely cool. So if you recall, we started that from absolute zero and we've used 49% of our five-hour usage allotment which we'll reset in two hours and 47 minutes. For ARM, all models we used 8% of our usage. So, and I was running pretty heavy things, at least multiple different tasks, not necessarily large code bases that it had to chew through and do things. So, your usage may vary depending on what you're doing. But in terms of zero shot generations, the usage was, I would say fair, though I do have to recall, I did up to the $200 subscription prior to filming this, but still. I would say play with this now while you have it here in the web chat interface assuming you do have access to it. So really, that's going to conclude our first look and test of Claude Fable 5 or the consumer oriented mythos 5. Overall, it seems extremely extremely capable in specific and software development. As some of the things we saw here, mainly the C++ gate game, we're in their own league of capability. Not the browser rest though, that was kind of mad but everything else was awesome. So that's going to conclude it. If you have any questions please feel free to leave them in the comments and thanks for watching.