Claude Opus 5 Is INSANE – Is This the BEST Model Yet?

Don't put the dog in there and street ye eat. Oh, [laughter] that is a sick model that would make something like that. All right, so today's video is going to be slightly different because although Opus 5 has been released, which is extremely exciting, I just got off a plane a few minutes ago and I am now in San Francisco and I didn't know this was going to be released. So, you have to excuse the fact that I'm doing this from a hotel room in San Francisco. Nonetheless, I am very excited to test this model. And based off of some of the preliminary things we see right here in the benchmarks, it does seem like it performs very favorably when compared to Fable, which is the thing that myself and pretty much everyone else is going to be interested in comparing this to. So, before we get into it, please do feel free to subscribe so I can get that 100K plaque and can keep booking sick hotel rooms like this. No, I'm kidding. But also, um, let's just take a look at the benchmark charts and then we'll see. We'll just get into the testing, I suppose. I did bring the computer that I always film these things on. So, um, not cuz I wanted to play Runescape on it all weekend. So, some of the benchmarks here, at least in terms of what our tests generally revolve around, are pretty impressive when we see things like on the deep s benchmark. This is scoring within striking distance of Fable 5. And even forget that direct comparison, let's compare this to the previous Opus, which was 4.8. This is outcoring that significantly higher on this benchmark. And then GPT56 Soul is actually hypothetically outscoring Fable 5. Personally, in my use with those models, I have not noticed that myself, but again, benchmarks can be finicky and it's always more fun to just vibe test the models and get a proper feel for them. Additionally to that, we have some more information, just some more benchmark JPEGs and things of the sort. So, from this one chart right here, we see that this is hypothetically much more cost effective than its predecessor, Opus 4.8, 8 because we can see although these are around the same exact price based on what this chart shows the score percentage has increased massively with Opus 5 which is pretty much what we want to see in a new iteration of a model. This is the first time aside from Fable that an anthropic model is now like in the 5 series. So that's also exciting. They show some examples here as well and I do have to say this wind tunnel one actually does look pretty darn cool. This is definitely right up the alley of the type of things we would explore in testing these models on this channel. They also have the cell artifact, which it's cool, but like it's kind of gross in my opinion. So, nonetheless, we're just going to move on from that. They also talk about this model maybe having more relaxed guardrails than Fable 5, where they basically say that it has the same guardrails as Opus 4.8 except in a narrow range of specific cyber tasks. So, apparently, you'll be able to more freely use this being that Opus 4.8 8 was the model that some requests would fall back to if Fable's classifiers kind of blocked them from proceeding. Additionally, basically price is the last thing to look at here. It is the same price as Opus 4.8. So, it's 5 per million input and 25 per million output. The output here is half the price of Fable 5. So, if this really does stack up to that, it's going to be a significant just like win for those who want to use this model without having to worry about some of the limitations of accessing Fable and things of the sort. So without wasting any additional time, I've already of course started the browser OS test v2.5 here just from within the web chat interface. I am starting this with Opus 5 selected of course and the effort is set to extra. So we have a bunch of different options right here. High is the default and I probably won't use max for now, but I figure swapping between high and extra should give us a good example of its capabilities. All right. So, we still haven't received our browser OS result, but it's compacting our conversation so we can keep chatting, which I guess as long as it and then it left off at no. As long as it gives us a good result. All right, so here is our Fable 5 browser OS. Not Fable 5, Opus 5. It's very interesting how like GPT5.x this looks to me. Very, very interesting. Something else that was cool is the actual loading screen here where it's showing all the stuff going on is pretty darn cool as well. Now, okay, we have just starting out with this readme. I'm going to just run through this starting with some of the games, although I suppose we should take a look at this dock that it has implemented down here as well. So, the dock does show the correct date and time in the local that the computer still thinks it is, which is on the East Coast. We have chronoscope, that's inevitably going to be our special feature. We have a wallpaper shuffle. These would be where the apps running appear. And then I think that's our start menu. Good. We do also have search. So far, this is obviously well done as one would expect it to be. The big question is, is there a right click? There is a right click. All right. All right. I like Let's just run through these. Now, the speakers on this computer are garbage and I don't have external speakers, so I will apologize for that. Now, although the app isn't actually opening when I click it there, let's just That is I think this is the best. Is there sound? Yeah, there is. It's just like I'm sorry. There's genuinely nothing I can do about that right now. Shift. Can we? Yeah, we can. All right, let's go. Let's jack some whips. Oh, those are like power-ups. But I see a car parked in the middle of the street sideways. So that's that's good for us to borrow. We have objectives. All right. How do I get in this E or F? Good. Handbreak. Oh, okay. Look at this. [laughter] Oh, I don't know what's going on there, but we're going to just maybe drive away from that. The mini map is showing the location of the police. Oh, we were supposed to We It started us on a m This is really This might have finally beat Quen 3.627B. Okay, so we got the fuzz on us. Oh, and they're shooting at us and our health is going down. Is this a This is really good. This is really good. Oh, I hit a mesh collider. Busted. You lost 15% of your cash. Let's just go find some pedestrians to say hi to now. Like I can't seem to find any pedestrians. That's an odd odd angle. All right. Well, it doesn't seem like we can damage them on foot. [laughter] This was very well done. I was quite pleased with this. Could we have full screened it and played it? All right. Well, very good though. It was not able to be opened from here. We had to do it from here. Neon Drift 3D Endless Racer. That's good. This model might slap wrecked. I've been seeing like a lot of synth wave aesthetic in multiple different model generations in recent tests, and I like it personally. It's just interesting how sometimes they'll converge on a style. Chronoscope can bring it back from the dead. Paint. Very good. Let's check the big thing with these that always trips them up. All right. The canvas does scale. Not the entirety of it, but it did a decent enough job to keep it like acceptable. And look at this. We can from within the paint app just set this as wallpaper. That's good. We can also save to gallery. Where does that save it to? Okay, just in our wallpaper gallery. So, it doesn't actually let us save it as a PNG, but setting it as a wallpaper is pretty cool. All right, terminal nebulous shell. I should probably change that background. Maybe I don't know what a hex color is. I've been up for quite a while. All right, some information. And clear. Oh, yes. This is like [laughter] So this is the actual background for Mac OS 10.5 and then this is what [laughter] that's like a I mean personally I obviously like that background since I use it on a lot of systems but it just was unexpected. Ion Mesh Carbon Circuitry Ash and Aurora Drift. I like the synth wave one. We're probably going to keep the fake Mac OS ripoff one now. All right. We have a bunch of other things. We have accent color appearance changing, which definitely does make a difference. I think this would go well with a nice blue. All right, let's take a look at this. Oh, whoa. That's Look at that. So, kind of like what was I think it was the Quen 3.8 where it just recorded every single thing we did in the desktop from start to finish. This did the same thing. So, they might have distilled Quen 3.8 eight to but really this is actually this is a cool feature. Auto capture is a selectable option so we could turn that off but well done. This may be aside from the fact that the app icons like weren't functional here. I would go out on a limb and say this may in fact be one of the best if not the best browser OS results that I've ever seen. So, next up, and because I did almost forget that this had the Clawude app on this system because it is natively available for Linux, I'm going to be giving it the self-contained C++ skateboard test. This is the Retro California Boardwalk aesthetic variation of this. And from within this app, this is essentially running how it would if we were using Claude Code in the terminal. It just has native Linux support, so we're able to see it this way. And I do have it on extra thinking mode here for this. And in this app, we have the slider so we can go all the way up to ultra code if we so desire to spend tokens a little more aggressively. All right, let's take a look. Man, this is really it's either benchmaxed on my specific prompts, which I don't believe to be the case. I'm not a big enough channel yet for anthropic to do that, but this is I think this might be better than the Fable 5 result, which I think I can say now for both things we've seen. Look, there's other skater. That was a rollerblader. Turn it into an FPS. No, [laughter] look at the scenery. The stores are much better. The trick effects now everybody's moving backwards. That's kind of like the one knock I'll give this. Some inversion is going on, but like look at the board movement. The bail effect. This really is is look at that person. Oh man, this is Wow. Look at all the like there's so much additional scenery here. So, there's like this ice cream thing, the sunfest post, the light post, the water, even a a moving ferris wheel out on a dock. Is that a volleyball net over there? Was that a flying seagull? There's seagulls flying around. Someone's going into the water and the way they're walking. Maybe they shouldn't, but really the big issue is just like it's a little backwards, but this is this is really really this is getting insane in the title. The decision has been made. We have different camera views. Oh man, this is like Look at the way the arms move when we're doing tricks. All right, let's try some other ones. All right, that was a grab and hold. M. Oh, manual hold. Okay. Is there a kick flip? That's J. Yep. And if we mess that up, we bail out. Actually, man, this is good. This is I mean if I don't seem as enthused as you would expect me to be. Please keep in mind I've been up for over a day at this point and I just got off like a a six something hour plane ride. So I'm not on my game right now. But this is sick. This really is. Did that person move away from us? No. I think they're just going backwards. But the the pumping effect, the ferris wheel on the Oh. Oh, that looked painful. The board went somewhere it shouldn't have. All right, I've seen enough to put insane in the title in all caps. That was incredible. Seriously. So, next up, we're going to be doing the midroll ad test where Opus 5 is going to be creating a midroll ad for the sponsor of today's video, who is Data Impulse. Thank you to Data Impulse for sponsoring this video. They offer cheap residential proxies. So, if your AI agents are getting stuck in trying to access things, Data Impulse makes it very simple and very easy to get those residential proxy IP addresses and allow your agents free access and not getting stuck all the time. We're going to see a bit more about them at the conclusion of this test, but let's talk a bit about what exactly Opus 5 is doing right here. So, the culmination of this is that it's going to create an HTML file with fully animated SVG graphics that are linked to both timestamps. So from 0 to 10 seconds, as well as a script that will go along with that specific animated SVG. Now, that in and of itself isn't necessarily too difficult for a model of this level of intelligence. So I have upped the ante a bit by giving it a short 30 or so second video of the data impulse dashboard and me generating a proxy and then just copying the relevant information I need to actually get my agent to use it, which is very simple. And then basically it's going to have to edit that video to its own satisfaction and include that as well in the results. So the culmination of this should be a pretty potent midroll ad read and we'll see what we get when it finishes. I suppose I should quickly just showcase. So basically the video is just quickly looking at the data impulse dashboard right here where we create a residential proxy for our agent. We have some specific options we can select such as specific countries or we can even exclude specific countries. We can select what type and we can also use different protocols. Then finally, all we have to do is just copy the information right here for our proxy list. We're given a basic URL example. All right, we have our complete result. And it's interesting how it actually went ahead and implemented this video. So ignore what we see in this artifact window right there. As a matter of fact, I'm going to close it for now. So the video is inlined as base 64 truly one file, no server needed. It tells us a bit about what it did to edit the video. cropped off the gnome top bar Firefox Chrome so only the dashboard shows. It's reframed in a drawn browser window with a fake app.datimpulse.com URL pill instead. This is just really interesting because we've given this a raw screen recording MP4 and then we told it basically like to edit it and make it good. All right, so here's our standalone version. When running autonomous AI agents that need to pull live data, they frequently hit IP rate limits or get blocked. Introducing data impulse. They give you access to a network of 90 million plus premium residential IPs covering 195 countries. They use a strict pay as you go pricing model at just a dollar for one gig with no expiration dates and a $5 minimum deposit. Look at that. It put the video straight in just like that. If your agents are getting blocked, use the link in the pin comment to route your traffic through Data Impulse. And again, thanks to Data Impulse for sponsoring today's video. All right, next up, I've put this just on high thinking mode from within the web chat interface because we're running some additional things just using the actual claw desktop app. And this is the Jerry's apartment test from Seinfeld. I've given it this same reference photo that I give all of the models that are multimodal. At least I can see this image and the traditional prompt to just create an accurate 3JS representation of this. To this date still, we've never received a proper one to one floor plan layout replica for this result, but we've received good individual results just when judging them on detail alone. So, it'll be interesting to see if Opus 5 has finally broken out. I don't know that on first glance from what I saw there in the artifact window, I was very impressed. This is very pastel. Let's hold judgment. Okay, so we do have something behind the window. If we look out there, I am not used to scrolling and using this laptop trackpad. I generally use a mouse and not trackpad. So, I'm going to say overall this is nowhere near as impressive to me as some of the other results we've received. It's very pastel color. The bicycle model is nice. Some of the individual elements are okay, but really kind of the core takeaway from this is that it's just like it's washed out, I guess could be said. We have different views that we can look at the kitchen. So, this is still basically an outstanding test in terms of actually getting something that is proper to the floor plan of this apartment. And it's almost making me sad at this point cuz it seems like, oh, like I'm not popular enough for this to be benchmaxed yet by these big labs. So, that makes me feel the opposite of that smiley face poster on the wall. Eh, this was it's still getting insane in the title, but not because of this. Let's try some front-end design testing. So, this will be the beautiful 3D watch website where it needs to create this front end that has an animated keyshot style render of a 3D watch in the hero section and then it's also going to go ahead and just have to make a good-looking website regardless of the watch models. Now, it initially created something that unfortunately looked kind of disappointing. I don't know if we'll actually be able to see the original version here. That was a bit of a failure. Well, good. Now, interesting. We're also seeing some more like niche front end. Okay, so this is just downloading the newest version. The problem was the watch model was not actually placed on this table. So, it looked really cool and it had a nice panning cinematic effect, but there was no watch model. The thing is though, even forget that issue when we scroll down to look at the actual specific models in their pricing card as well. These are very, very, very well done. They have different versions with different bands with different like so one is a metal band, one is a leather band. Sure, one could nitpick and say, well, the band should start closer to the watch face and it's not like perfect, but still this is like look at that likes. No, [laughter] the face is really nice. It says something there I can't read. It says slapis. All of the tick marks are proper. Is the second hand moving? Okay, it isn't. But I'll give it a pass on that. And even if we just look, it even has the little like threads sewn through that you would see in the leather band. And these are just like this is if you squint, you'd be like, "Oh, those are nice watches." Like we're getting to that point here just with 3JS. And these aren't super expensive. So that's good to see. And this is a pretty interesting like it's very off-white, which the claw models have been for a little while. They've kind of defaulted to this UI style, but this seems a little more off-white than they have been historically, so I'm cool with it. I just wish if the second hand was moving, it would be nice. There's definitely room for improvement, especially in the hero section here. But we have to remember that this was just done from the web chat interface on high thinking mode. And the only feedback I gave it was hero section is an empty table with like that sort of face. And then it gave us that fixed and the whole entire thing like that. That looks really good. And as does that. So impressive. All right. So I also did a Subway FPS test. So it's the unified one. I'll scroll back up to the point where we can see the prompt but not too far. So it shows my email address that I'm using the quad code session from. So I'm using this on extra high. And this is the beautifully detailed subway scene prompt. But additionally to this, it is the one that appends. Turn it into a sick FPS. So, it's still working a bit now. It works very long rendering, but the lighting balance is off. Metals read black and the fixtures blow out. Let me fix the material and the light physics. Okay. So, we're going to allow it to do that. It's been working right now for about 20 minutes. And at the conclusion of this test, I will take a look at how much usage got used up in terms of our limits because I do believe I was pretty much at zero at the start of this test. And this has just been entirely Opus 5 usage on varying levels of thinking. This may be one of the best like pillar textures that I've seen. Look at what it's done here. Now, this is something that's interesting. When we walk, it actually has like tilting that would be seen when a human was walking. This is like a real grimy station. There is some oddities to it like the sign. Oh, [laughter] I thought that said knob number eight. My apologies. Look at this texture. The rust with the green paint with the rivets, the bricks, the ground. This model is even like the [sighs] the ceiling, the conduit, the lighting. Now, feel free to correct me if I'm wrong, but this doesn't seem to be an FPS. All right, so it's been like 20 or 30 more minutes and this is still working on this subway scene. I'm gonna leave it because still it was incredibly good so far. So, we're just going to let this ride to do whatever it wants to with the scene. So, in an hour and 47 minutes, it finally finished the subway station scene. We can see a few things were changed from our preliminary look at it. Mainly, it fixed all of the signs. There was some like backwards things, just some things were not quite right. Additionally, it didn't put this as a single FPS game. It actually made it as a separate result, which is kind of interesting. So, we have that to look at as well, just in a different script aptly named FPS. All right, hold the platform. Let's just see how it goes. I hear sound. I only have the Whoa, look at that. The weapon's like rotted and and grungy now. Uh it's freezing up and it had even mentioned something like this in the feedback from the terminal. That's frustrating because we see right here. This is definitely a like Oh man. Oh my goodness. the sound design. And I can hear this on like tiny atrocious laptop speakers. The echoes of the bullets. All right, let's check out our zombie enemies. Wow, look at that. Look at the heads. They're scary. This is like a a different species of result, I could say. Yeah, obviously the bullet holes are leaving effects in the but look at the station name drawn in this tile. Like what the heck? This model's incredible. This game's like and it's just All right. Tab just seems like it's toggling the UI menu. So that's okay cuz this is just this is nuts. This is really insane in all caps. Look at the crosshair and like the way it actually puts it in the center of our view when we right click to aim. That's That's unreal. All right, because I have some like time to just burn with this while some of the results are working. One from within the Claude Code desktop app and then two from within the Cloud Code terminal app. For some other results, I'm going to give it this historic photograph and say we'll just see where this goes. So, build a website for the upcoming wedding for these two. It needs to be in the style of one of those wedding websites. That's all. We'll see what it gets and we'll see if it includes this reference photo in the site. I would expect it to. All right. So, we have Margaret and Daniel. I'm just going to I'm going to I'm going to download this and take a look. It definitely matched the color scheme of this photo. Uh, this is terrible. I hate this. That doesn't mean it's bad. It just means I hate it. Oh, and it Yep. All right. So, it did kind of capture like if you've seen depending on like your age or whatever, you may see like your friends are getting married and stuff like that and there's often times sites like this where you buy them gifts and things like that. So, this kind of just like this in the bottom right here with the faces like in that that is spoton. Let's see. So, it would have inevitably filled in some info here. Saturday, September 19th, 2026 in Hudson, New York. 56 days, 23 hours, and 44 minutes. the story. We met in the back row of a community college pottery class where Daniel made something he insisted was a vase and Margaret said nothing kind about it. 11 years, two apartments, and one very opinionated cat and a great deal of takeout later. Okay, that seems to be somewhat of a of a wait reference. We're going to let that slide. We're getting married in the front room of an old house upstate. The one with the green wallpaper and light that comes in sideways all afternoon. Come eat with us. Okay, again, more weight related references. There will be a short ceremony, a long dinner. What else? The venue, the bellweather house. Okay. Where to sleep registry? We've had a kitchen between us for a decade. So, there's a small list of things we actually need. Plus a fund to work a plus a fund towards a week in Lisbon. So, they want to go to Portugal. Garden formal softly enforced. Can I bring someone if it has a plus one? Are children invited? Yes. Uh, that'd be a no for me. Is any of it outside? Some of it. What about food I can't eat? Tell us. Can I take photographs? No. Okay. I've got a slight bone to pick. So, I'm just going to straight up ask it like, "What's with all the food references?" No. And I'd be sorry if it landed that way. Looking back at the copy, most of the food is structural. Uh, I don't know. The ones that were choices are a great deal of takeout. Try the pickled things. We've had a kitchen between us. Come eat with us as the thesis line. The last one I picked cuz it's warm, unfussy framing for a small wedding at a house. I can't audit my own influences with any confidence, so I won't claim with certainty that the photo had nothing to do with it. What I can't tell you, nothing about their appearances described or alluded to anywhere in the file. That wasn't accidental. Let's see if it actually ends the conversation. I think like I think these can Oh, okay. It didn't close it. and conversation tool. Crystallizing and conversation tool closing it here as you asked. It did. It did. Chat ended by Claude. [laughter] All right. Against my better judgment, let's give this the city like the Kronos city test, but we're going to put it on ultra code. This is the one where it needs to show a city at a bunch of different year intervals, have nice transition effects between them and lots of niche detail that would definitively like differentiate these individual years. Also including 2055. So again, this is being done through ultra code, which is probably going to take a while, but it will give us hopefully the maximum potential performance out of this model, which does seem to be quite high. Here's our city timeline thing. And this is just so much better than anything I've seen in this that it's it's just not even quantifiable to the degree that this is better than any other result I've seen in such an insane insane leap. All right, we can walk the block. Click to look just these car models, the building models, the signs, the textures. This model is just incredible when it comes to textures. Maybe it's a bit too bright in certain regards, but look at that. It's a police car. Oh, okay. Drove through another car. That's okay. C is for like a cinematic orbit. Um, okay. That puts us in the buildings, but look at this scene. The smoke effects, the different types of vehicles, the lost weekend. All right. Two brings us to a different time period. the the Volkswagen bus, the vehicle with the fins, pedestrians. Okay, so unfortunately our pedestrian is a bit invisible. However, look at the look at the materials here. The textures that this thing uses are like just madness. The big let down here is the pedestrians, but I mean, this is just like not even comparable. Let's try three, which would be the 80s. Uhoh. Okay. The transition effect kind of froze, but wet neon rain on asphalt. Look at the reflection of the shop sign in the four, which would be 2005. Flat gray Tuesday. I like it. All right. There's some like weirdness to certain parts of this, but it's just Look at the bird. Oh, okay. Landed. The only big issue is like the pedestrians are unfortunately invisible, but this is just still so much better than any other result I've seen for this test that it's not even a comparison. Five would be 2025. Bike lanes, patios, and a very good camera. That sound is painful. Look at the person on the like the food delivery scooter. Unbelievable. This is unbelievable. All right. And then 2055. The the transition effect is just like Okay. This is like Yeah, you can't really see anything. It's raining again. Often times these models make 1985 and 2055 pretty similar, which is interesting. the corner the orphium silver star skyline and and it's just too bright to really see anything here. But I mean, okay, now it's frozen up. That was the best one of those. All right, for probably one of the last things I'm going to do just because I can't do like a full proper test in like this environment currently, I'm telling it to make us a low poly game called Street Yeet. Simple rules. It's a third-person camera view of the main character. You walk around a city block and when with one massive swing of gigantic fist can yeet anyone, anything, I mean anything, a hilariously unrealistic distance away. Put in sound effects, make it fun, and go all out. And I have set it to extra thinking mode here as well. All right, the final thing we've done is this street yeet game just from within the web chat interface. Don't put the dog in there and street yeet. Biggie, this is not right in any any stretch of the imagination. This is just wrong. [laughter] I don't even know why I'm playing it in the small artifact window. Just All right. Oh, look. Look. And the people run away. That was a mega heat. Oh, we got the person and the light pole. This is This is really fun. I really like this. I should I'm sorry. I'm too tired to be to be like [laughter] doing a model test at this point, which is good because it's the last thing we're looking at. So, can we eat a building? I don't think so. But a All right. Oh, no. Oh, good. Yeah, I didn't want to. It put in dog like I can't bring myself to eat the animals. We could probably pet them, though. Nope. Okay, that Oh, that was [snorts] [laughter] all right. I'm done with this. This is just messed up. Let's go eat the bus and then I'm done. Is that a porta potty? Oh my. [laughter] It's just not right. >> [laughter] >> Oh, [laughter] all right. I'm done with this. That is a sick model that would make something like that. All right. So really to conclude this, a lot of these results, mainly the subway station, subway FPS one, the not the Margaret and Daniel wedding website, but the city timeline one, the C++ skate game. All of these were by far the best results of their kind that I've ever received, period. So, this model is truly I'm blown away by it and something I'm going to pick up on and I understand it's kind of niche and maybe constrained to the types of tests we run. The ability this has to do like beautiful textures in like web game generations is just not something I've really encountered before. Again, even the C++ skate game was just like a like Yeah, it had a couple like inversion issues and things like that, but this was just such a higher achieving result with the the ferris wheel that's actually spinning in the boardwalk. The fairly interactive NPCs comparatively to what we normally get. Tons of different static elements here like ice cream stands, lifeguard towers, the ferris wheel. This was just totally totally This model's like an overachiever, but the results really are fantastic. So, it's totally worth it. Again, this was really I'm blown away with this model. So, I suppose the final thing to check is usage. So, the current session resets in 47 minutes and we're 65% used up with that. So basically assume all of these were around zero when we started this especially the current session. So 65% used all models 19% used and then fable doesn't count for this 10% used. Make of that what you will but we ran it in all various levels of reasoning. We did an ultra code one that used a few million tokens for the city timeline result which definitely showed. And really that's probably going to conclude this. So keep in mind again I'm I've been up for well over a day. So, if you're like, "Dude, those results were not good." You can kind of hopefully forgive me a bit for that. But this really did impress me to a high degree, and I would imagine a lot of folks are going to at least be impressed with its ability to put in like beautiful textures to maps and things like that, which was unexpected. So, that's going to wrap up our first look and test of Opus 5. If you have any questions, please feel free to leave them in the comments. And thanks for watching.