Oh, okay. Wow, that's so much better and it was good to start with. Today we're going to be testing Cursor Composer 2.5, which is a model that I've
seen a lot of folks mentioning both in the comments in some of my recent videos and wanting to see it tested as well as sharing demonstrations of some
rather impressive looking outputs just on social media. So, we'll be taking a look at this model,
testing it from within Cursor, and seeing if it really does stack up to some of the rather impressive benchmarks that have been shared. So, before we
get into it, feel free to hit that subscribe button so I can get to 100k. And with that, let's take our first look at Composer 2.5. Now, it is
interesting because Composer 2 came out not too long ago, and there was a bit of I don't want to say
controversy, but some of the talking heads who like to engage in drama had definitely mentioned about that being based off of Kimmy K2. And it was
kind of like pretty blatant, so then there was like some posts that were then deleted by the Kimmy team, and then Cursor basically came out and said,
"Yes, it was based off of Kimmy K2.5." As we can see right here, this is now prominently displayed in
the blog post for this model. So, it's just good to see it and it is very interesting that the basis for this model is an open weight Kimmy K2.5,
which is in and of itself a very, very potent 1 trillion plus parameter model. Though, we can see right here, this does not just look like they took
K2.5 and slapped a new paint job on it. There is a significant amount of post-training and reinforcement
learning and additional tuning that seems to have gone into the creation of this model, which would kind of showcase itself here in the specific
benchmark where we do see this is stacking up very, very favorably against basically frontier state-of-the-art models in Opus 4.7 as well as GPT-5.5.
These are the self-reported scores for public evals. Additionally to that, something I always like to
take note of is the scores compared to the previous version of the model, where we can see that versus Composer 2 to Composer 2.5, we have some pretty
significant leaps in capability. So, it's kind of exciting to see what this thing is able to do. And where this starts to become a little more
exciting just based off of its touted performance here is we can see this chart where it's showing
performance to cost. Now, if we scroll down, we'll get a bit more pertinent information, but I do believe that there was some specific chart shared. I
can't think of where it was off the top of my head, but I do believe that this cost like 55 cents to complete a specific task versus the other models
which were significantly more expensive. And if we do scroll all the here, just conveniently
skipping over this extremely technical portion, we can see on the bottom of the page they do mention the pricing metrics for this model, which is 50
cents per million input tokens and $2.50 per million output tokens. A lot of people are kind of freaked out about the Gemini 3.5 flash pricing of 9
million $9 per million out. So, this model being this cheap could be extremely impressive, but almost
scary as well depending on whether you are a consumer or a model provider competing against this. So, additionally to that, there is a fast variant
with the same intelligence for fairly more expensive. So, we have what's that? Six times as much money in terms of the input per million and then
seven times as much per million out. Seven times. That's not right. Six times. There we go. I apologize
for my incompetence. And we also get a bit more information in the specific model documentation pertaining to this model where we can see it has a
context window of 200K. Intelligence is listed as frontier, cost is listed as low, and speed is listed as fast. All things that are nice to have in a
model, especially one that you have to pay for. Additionally, there is mention of the agentic tools
that it has access to and some more specific pricing breakdown just in charts versus the fast and non-fast mode for this model. So, really with that,
we're now going to jump into Cursor and actually begin to test this. Now, I do have this open in this kind of agent window as it does seem things are
trending towards this way with Kodak's anti-gravity 2.0. All of them really have kind of like a chat
window first UI design to them, which I'm totally cool with because it lends itself better to actually testing these in a video, in my opinion. Now, I
am going to start this in plan mode. Of course, we're going to begin with the browser OS test V2.5, which includes the two 3D games, the special
feature, and wallpaper changing ability. All right, we have a bunch of follow-up questions here that it
is going to be asking us. That's okay to save it to the home directory. Okay, that was just one. All right, we've got our implementation plan right
here, which is nice. Now, it's funny, we also have an architecture flowchart. This was something we noticed in anti-gravity with 3.5 flash earlier
today. So, that's just cool to see as well. The specification or markdown showcasing the implementation
plan here is very, very detailed, and it gives itself nine to-dos as well. So, with that, I'm going to allow it to go ahead and initiate the build
process. Um this is in fast mode, so hopefully it won't take very long. All right, we have our result. That was relatively quick. Okay, we have a
loading screen. Simple. I'm going to reserve judgment at least until we start to see some of these apps,
but just on first glance, it is a traditional browser OS. Does it have a right click? Yes, it does. Okay, good. It's already winning points. We do
have the clock in the bottom right, which is the correct time in my locale, and let's check out our start menu. Okay, simple and minimalist. Again, I
will reserve judgment until now.
Okay. I feel like what's happened here is it drew a mini map because very briefly in the corners, we can actually see something that looks somewhat 3D
and promising. It just seems like the mini map is being rendered um like over the Yeah, I'm not 100% certain what's happened here. We're going to need
this to be remedied. So, I asked it if it can see images, as I feel that is very likely the best
way to debug this specific issue. Now, I'm actually just going to send it the image without saying anything, because I want to test its ability to
look at that, realize something's wrong, figure out what specifically the issue is related to, and then fix it. And it fixed it. That was very quick,
and I'm happy to see that with only an image, it seemingly at least fixed the issue. Very, very, very
good. And that was the issue, the mini-map is there. WASD is to drive, space is to brake, and the goal is to collect green cash. Let's full screen
this. All right. You know what? This is one of the results, and something I don't really mention too often, cuz I haven't thought to. A lot of times
the cars are just ridiculously too fast for the maps. This is not one of those cases. If anything, it's
almost a little too slow. The amount of buildings and things here, although the rest of it is pretty light. Like, there's no pedestrians, there's no
other cars, there's no real visible streets. Let's see if they cash. 100. Very good. All right. Okay. There are colliders. Space is to brake. Yep,
that Okay. All right. Basic, but this has good bones, and the mini-map is there as well, and it does
show us the location of the green cash as well. So, we're up to 200, and All right. Acceptable. Let's check out Orbital Storm. I I read that as
Orbital Stream, and I got confused, and then I looked at it again and realized, "Whoa, this is funky." Mhm. Orbital Stream would have worked. All
right. Do we do anything? It's 3D-ish, I'll give it that. Not very immersive, but it is functional. Can we
minimize and open to the taskbar? Yes. Good. Next up, we have our notes application. Okay, very simple. Auto save to browser storage, but additionally
to that, there's nothing really special about this, but that's quite all right. Then we have our calculator. Nicely nicely arranged. All right.
Resizing does not break the layout, which is just nice attention to detail. Let's do 47 * 5, 235. Very
good. Divided by 9. Settings. All right, we have some gradients to choose from. I love that. That's very retro, and I like it. It's calming. This is a
very common color. And then also, it has the nice option to upload our own. I have no idea what the screenshot Okay, I should have known. So, we can
upload our own screenshot. Of course, that was the debug issue that I'd given it. I think maybe
we're going to just go back to something. Then we can change our accent color as well. Let's change it to something obscene, so that we'll immediately
notice. And we do see the neon green is highlighting the selected gradient now. Very cool. Auto restore last session on load. All right, and the green
did change the start button as well. Finally, we have about Browser OS V1. Seems like the accent
color also modified this. Special feature, Omni command palette and session snapshots. Press Alt and space to open a spotlight style command palette.
Search and launch apps, change wallpapers, minimize all windows. Okay, now sadly, I can't seem to get that to work because of That's okay. We'll We'll
just assume it worked. Or not, it's not a huge deal. So, overall though, acceptable. Next up, we're
going to try the beautiful subway scene {slash} FPS prompt. I have just appended to the end of this. It should be contained in a single HTML file. I
also am purposely not running this in any plan mode or anything. I basically want to get as close to what just a zero-shot result for this would be
without any planning or anything. The only thing I did is put it in its own directory. All right, let's
take a look at our subway FPS, Metro 17. Okay. Oh, wow. All right. The speaker's on. It's This is really like violent. This camera This is a It's a
odd behavior where the camera pitch twists depending on the This is just so like rapid fire. Okay, I need a Okay, can we actually adjust these? No, we
can't. And that is that is a bug that consistently trips models up. I know you might look at this and
be like, "Dude, this test is stupid." And to that I'd say everyone's entitled to their own opinion. But but one thing that consistently trips these up
is you can't actually touch or change any of these sliders for station brightness or fog density in this case. Because clicking into the window just
hides the cursor and locks it to the weapon behavior. So, that's like a gotcha that consistently
seems to happen. I also need to make sure like what's up with the sound? >> Front, front, right, left. >> No, the sound works. So, it must be an
actual game specific thing. All right, we have a few things to give this as a follow-up. All right, so I've given it a description of the issues and
we'll have it fix these. Okay, can I Can I like You know what? We're just going to test it here. Can we
change the brightness? Fantastic. All right. We do have shell ejection, but uh we don't have actual like when the bullets hit the environment if they
don't actually make holes. I've only seen that in two models so far. One of the Opus models and then Quake 3.7 Max Preview. So, make of that what you
will. Oh, okay.
This is You know, I mean this is quite good. I like this. Oh, that one's stuck. Yeah, I don't know what happened there, but Let's explore the station as well. Listen to the footsteps.
This is >> That should be wave three, right? Yeah. Oh. Oh, okay. And then it just restarts us there. Fog density, so
Oh, okay. Overall, a very satisfactory result. The weapon model's quite good, the muzzle flash is good, and we do see the shell casings ejected. I
want to just see how it responds to like a very, very stupid idea.
This is just like logically incorrect. The pitch is catchy, but why do Okay, good. So, basically, I just wanted to make sure that it wasn't going to
be like, "This is game-changing." Very good. Okay, good. So, quick test to just see if it will basically reinforce foolish behavior ideas, in which
case it did not. It pushed back against us. So, I'm happy to see that. All right, so I've swapped it
into plan mode because I actually want to turn this into a front-end design capability. So, my stupid idea of reverse quantization, where you just
make the model larger, which wouldn't work. I've basically given it a very verbose and dense front-end design test as well. Just saying, "Create this
based off of the existing concept from this conversation." So, it's going to now create a beautiful
website that is designed for a business that is based on the concept of reverse quantization. It has given us a plan. We do also have a very
nice-looking UI flowchart right there as well. Final call to action, JS, CSS highlights, and it should be AI-related, and five to do. So, we're going
to allow it to build this here. I'm going to just try hiding the side panel so we don't get any spoilers, and
we'll see what we get. Okay. It's I don't like this. I'm not impressed with this, but it is entirely possible that the prompt let it down. Although,
I'm going to say I did allow it to run in plan mode, and I did also give it a very, very lengthy prompt with a lot of specific things. It's possible
that constrained it too much from being able to really showcase what it wanted to. I'm going to give
it some feedback here. This looks like absolute top-tier garbage. Okay, so something is really going extremely slow right now. I'm not sure. It does
seem like Okay, and it just kind of changed it to a different color scheme, but I'm going to say this is definitely a better result than this. It's
significantly less cookie-cutter like GLM 432B front-end looking. So, this is nice to see. I actually
do like the arrangement of this card stuff here, where it is four separate cards, but they're all laid out with different space that they take up.
64-bit max bit promotion. 0% information recovered. Zero measurable IQ gain. Oh, because it called us out saying that was not a proper That makes
sense. Upgrade precision. Ready to promote your weights? Looks like a like a dieting thing. Okay, overall
still I may need to do another front-end test at some point to give a definitive answer, but I'm not blown away. So, this was initially going to be a
simpler test where I gave it a reference photo of a soft body physics demolition derby game that was created with Nano Banana Pro. This is something
that no existing publicly available model would be able to create off of this reference image from
one shot. It actually in a follow-up when I put it in plan mode to start, it asked what platform we want to do. Now, traditionally I'd just say 3.js
in the browser. However, it did give the option to do it in Godot, not Go.net. Only a fool would refer to it as Go.net, but that made me excited. So,
I've gone ahead and actually downloaded Godot 4, and I just put the executable or not the executable,
the downloaded files in the same directory this is working in because not only am I going to have this create this for Godot or in Godot, but I'm
going to also task it with actually unpacking that extracting the files, getting that set up and running on this specific Ubuntu system because beyond
actually just being able to code, agents also function in actually being able to perform actions in the
operating system configurations and things of the sort which is an area I don't necessarily see talked about or demonstrated as much as just raw
coding capability. And it's still going to create its plan. So following that, I'm going to tell it, "Okay, there is also the Godot 4 download in this
folder. So I'm going to need you to basically go ahead and set everything up as well. Multitask,
coordinate parallel tasks." Also, so I'm sending that as well in the multitask section. So it did also give us this crash ball 24 soft body
deformation. Okay, we have a quite good-looking project scaffold matching the concept layout. Okay. Allow. And I did also set it on the subtask to get
Godot set up and running. I'm just going to have to allow it manually to perform all of the tasks right here,
but it should be able to agentically just do everything in terms of getting that set up and running on the system. I know for a fact Claude code would
be able to encode S and like everything. Sub-agent has finished successfully. Beautiful. And now we can go ahead and build our plan. Good. And folks
want to see like more in line with working in an existing repository. This isn't quite the same, but
it does at least go into more intricate setup. This is definitely a more significant task than what we normally run in some of these lighter testing
videos. So I'll be very, very interested to see what this does. So the game engine opened which is good because it was not initially even installed on
the system. So that does show that our sub-agent that we had created to make sure that everything
there was set up did work. Unfortunately, we did not get a result on first try here that at least showed us something. I didn't expect to. But, the
cool thing was that when it's failing, it seems to be able to understand that it is failing or not failing as happened in that case and then we'll fix
it or take corresponding action based off of that. Okay. Oh. Okay, so Uh okay, good. Yeah, cuz I
didn't even know where to begin in actually trying to give a plain English description of that issue. So, I'm very happy to see that it must have
inevitably pulled some log or had some other way to assess behavior. Though, I will say initially the scaffolding looked promising. We do have audio
as well. I didn't hear anything, but we'll just we'll see what's up with this. Okay, it's getting better.
All right, please just assess that yourself. Okay, they do have soft body physics. Oh, no. Okay, uh This sort of thing happens with prompts like Okay,
so when the cars fall to the ground, they actually do deform, which is cool. One of the issues I'm noticing here maybe further down the line
significantly so when we'll get to it, but it's just like basic rectangles. Though, the soft body physics
implementation does look pretty good. Final good dough import validation. It's ready and done. Uh I don't know about that. Okay, we're just going to
we'll try this. Uh I don't know that I feel 100% confident in the
Okay, okay. the issues are quite frankly difficult to even describe. We'll see where this goes. And this is in fast mode, which is more expensive.
Okay. Better, but not Seems like gravity is too strong and is deforming the car body shapes. All right, it's it's gotten better. Uh It's definitely
gotten better. We're at 61% context, so we're still We're doing fine. And I'm just going to give it a
photo of what we witnessed, though it's actually playable enough now to be able to go back and see. The cars are not very drivable. We're only in
We're still in sixth place, so 42 blue and 55 red must be quite terrible considering I drove off of the map and we're still beating them. All right,
here's our next iteration of this. It made the cars better. The driving is still off, but does the soft
body physics kind of work or
Yeah, they do kind of work. And again, this is a This is a difficult task. I'm going to give it another photo as that seemed to do a decent job of
helping it. We're at 70% context. I'm probably going to have this be one of the final improvements that I give it to for this specific task, but this
is definitely something far more difficult to add into the repertoire of model testing, I suppose. All
right, here's an additional improvement. So, the cars look less wonky and the driving works now. Oh, okay. Now, they're all entirely stuck together.
All right, I'm going to call it here. It was interesting to see the process of improvement overall as it did and again, this is testing this in Godot
with a very difficult prompt that was just based off of a reference image of something that no model
currently is going to be able to knock out even in a few tries. Maybe some of the very very intelligent ones would, but they would be significantly
more expensive to run than this model is. So, overall interesting and this is definitely something I will keep in the wheelhouse in terms of a more
intricate test than just having this be done in 3.js and an HTML script. All right, let's do the C++
skateboard simulation game in the aesthetic of California late '90s boardwalk style. I've not done this in plan mode. I didn't even put it in its own
directory. I just wanted to spit something out that we can test and then we'll have it make improvements as needed. All right, that all happened
pretty quickly and then we have just a bulk command here to build and run. Okay, so we did get a shader
error. I'm just going to feed this back to it. I would have liked to have seen it at least autonomously try that itself and then find that error and
then fixing compile errors. Okay, it did try that. So, we'll see what happens. Further debugging step. Okay, well, it did improve because it actually
is opening now. That's not That's not sarcasm. That's a genuine thing. So, I'll just send it a photo
of that. I will turn the speaker way down so that stops scaring me and then we'll see what it does. It does seem to have a good ability to just
troubleshoot based off of looking at an image without even giving it any additional text corresponding to that. All right, let's see now if Very good.
Okay, I don't know what exactly it's it's tried to draw there in text. The fonts are jank. Though, look
at the He's like Oh, all right. Interesting. We have some promise here. It's A, why is there another skateboarder there? I don't like that. I don't
want I want to be the only skateboarder. Those are NPCs. We'll do one more.
It's not that bad, but I wanted to just reinforce that I was disappointed along with the photo of the large text. All right, let's check our final
improvement test just based off of what we had sent it. Okay, it made the text a little less extremely large. Overall, there's still a lot of funk to
this, but we can see where the vision was. And it happened very quickly and definitely at a low cost
comparatively. Well, this isn't fast mode, so not super cheap, but poomjee! So, I don't even know what that's supposed to be saying because even if we
were to think like, "Oh, it's just inverted." It would be E J M M P. So, I don't know, maybe it's in a different language. I've no idea. All right,
well, that was educational. Next up, I want to give it a Seinfeld's apartment test because this is
one that actually, as silly as it may seem, this has been very good historically at differentiating very capable models from not as capable models.
So, this should really showcase us whether or not this is actually a 55 GPT 55 or Opus 4.7 competitor. The GPT 55 result I received for this specific
prompt in Codex was hands down the best one I've ever seen. And then I had it create like a beat 'em
up game out of that, but I wanted to just see how this does, whether or not it actually feels like it gets it, quote unquote. So, I went and found a
previous result that was done by Gemini 3.5 flash when we tested that like 12 hours ago or 16 hours ago. It found that and was like, "Okay, I'm going
to base this off of that." I went ahead and deleted that because I don't want it to be doing this.
So, now it's going to be like, "Hey, what happened? That file is gone." And it will have to do this more from a scratch like built perspective instead
of just piggybacking off of what it found on the system, which was clever.
All right, finally. Seinfeld apartment 5A. Let's enter the apartment. Okay, this is fine because it's just showing us it's we're outside first so we
can move around and then when we enter, we'll get our first look. The moment of truth. So far I like what I'm seeing. Okay. Interesting. Um It does
have the Porsche 356 poster on the wall. Okay, gives us tool tips as well. A lot of models when given
this prompt basically make like an educational walk-through guide or something of the sort. Paradox. This has paradox as well. Fascinating. I don't
mean that sarcastically. All right, let's go into the Macintosh area here. Can we click on this? Yes, we can. Macintosh SE. Okay, system 6 gray
Porsche 911 Turbo poster. Very fun car. Uh The It's a little hard to navigate around the scene but the
pieces are there.
So this Mac has a very good looking system 7 screen on it. If I did just do like a jump cut here, it's because I may or may not have not realized that
there was a much easier way to walk around the scene. We do have collision detection on some of the individual things. The kitchen is a little wonky
but we do have some more Porsche posters on the wall. The TV, again, none of the orientation has
ever really been spot on in doing this test in terms of where the models put things how they would be in real life but the bicycle was there. We were
able to see it for a moment. I do like the mini map showing where we are in the floor plan. This appears to be a bathroom of some sort. The bicycle is
once again momentarily visible and it is lime green. Let's go over to our kitchen. Things look good
here if we are judging based off of individual elements. Okay, and then we have the cereal boxes here which are almost stealing the show in terms of
how well done they are. A Superman and a Porsche thing. So cut roof. Awesome. That's a nice feature. That gives us such a better overview of how this
actually is going to look. And very, very interestingly, Gemini 3.5 today also did this where it
points out the paradox. This is so odd. This is almost a carbon copy of what we saw with Gemini. Essentially, this hallway is not able to be to exist
based off of the set of the show. And it draws a wireframe around it. This is exactly what Gemini 3 flash did earlier. It had paradox and then it drew
this exact same wireframe with red. Very, very fascinating how sometimes behaviors like clump
together from entirely different models at like the same time. It's just There's nothing to it, probably. But maybe there is. Next up, I'm trying
another front-end test because the previous one we did, I just didn't feel like I really had a good grasp on good or bad. This is the ravioli sauce
webpage. This is one that has just recently been introduced to the channel. I ran it with Gwen 3.7 max
preview, which did a fantastic job. Gemini 3.5 flash did uh So, we'll see how this does. This is again, it's actually like a really verbose prompt for
just a simple pasta sauce website, but it's I find it very fun. I personally am a big fan of the sauce. That's not like a I mean, like I like this
prompt. That's what I should say. It's 5:30 in the morning. It seems like it's hung up on something,
so I'm just going to open the page because it had inevitably made it. It just wasn't opening it for some reason. Okay, sauce so rich the ravioli
started orbiting. Interesting, and the background does have the sauce effect with the ravioli. Taste the sauce. Okay. And we have these weird card
effects, which I've seen become a common thing. So, if you see a website like made with these, know that
they 5 coded it. But okay, we have a sauce viscosity slider, which is where my head is going to be in the video. But okay, we changed it, so it's now
less. Now it's more. Okay. I don't know that I noticed anything, but did a nice job in drawing the signature jar. Simmer time, 6 hours, 12 basil leaf.
Each 680 g jar holds a galaxy of flavor. Your pasta has been waiting. 2026 Ravioli Rosso slow food
fast dreams. Okay. Not bad. I think really what's letting it down is an image of the sauce bottle right there on the right side of the header. That
would have really made it just above and beyond. We even have like a spinning thing up there, which is What a weird concept. Okay. It's tough to
judge. For the next test, I want to find a very specific prompt that I had run, so I'm going to
ultimateplay.com, which is my own website where you can upload, play, and upload AI-generated games, single-file HTML ones. I need to search. So, this
was a game result that was generated with Meta's Muse Spark, and it was actually really quite fun. It was a very simple prompt, so we'll start it up
just in plan mode here to give it the best chance of properly implementing this. I did close cursor
cuz it was seemingly lagging a bit and just didn't seem too happy, so I'm opening it back up now. All right, so I've had to change two things with
this prompt. One of them is to just denote a 3D beat-'em-up style game. I've also told it to not search the system, just stay in this directory and
create it from scratch. I tried this previously, and it went and found the Subway FPS and then started
ripping things from that to create the game, which was frustrating as I did not want that in this setting. So, okay, that was an incredibly fast plan
that it came up with. Being that this is a relatively simple prompt it's been given initially. We have, okay, player controlled movement plane, and
then camera angle fixed follow Steve on X. Okay, you could follow Bijan on X if you want. It let That
was like a layup, so I need to go and just like showcase now like x.com. So, you can follow me on X if you want to. And back to cursor. Okay, it's
implementing the full thing now. It's inevitably going to open it here in some form of spoiler alert, so I'm going to have to hit this with the
quickness so we don't see it. Or I'll just hide it now. All right, here's our Steve the PC repair man office
game. All right, clock in. Oh, that's cool. I like this style. This almost looks like a C++ result. Okay. Nice. Z is to punch. The walking effects are
cool. The office individuals that are the enemies that are following us have coffee cups, which I find kind of funny. There is a water cooler. Can we
smash this PC? Yeah, we can. Fixed. What are they throwing at us? Or I don't actually know what it
What are those? Okay, something was just thrown at us. I don't know what, but it didn't matter cuz we got a combo. All right, here's wave three. I
don't know what these things are, but I don't think we're going to be able to beat them. Ticket closed. Those must have been tickets. Steve has been
escorted from the building. Final score, 960. That was actually That was nice. I like this. I like it a
lot. So, I told it I like this a lot, but I think you can do better. All right, here's our improved Steve PC repair game. Wow, that's so much better
and it was good to start with. This is well done. This shows some good promise for web games and web game enthusiasts such as myself. Fixed. And it
made the Windows 8 blue screen. All right, that's good. He can kick as well. Look at the little plant
thing in the This is very good. I'm quite pleased with this to the point where I am going to upload this on ultimateplay.com, which is my website
where you will be able to then go and play this and see the prompt and things like that. So, I think those are tickets that are coming at us. Z is to
kick. The humanoid enemies are Well, humanoid is, you know, humanoid. They look good. We can't beat up
the computers that don't have blue screens. Ticket closed. Steve has been escorted from the building. Nice sound effects, nice little like movement
effects. I'm very pleased with this. And the coffee cups on the office drones are kind of funny as well. All right, the final test I'm going to do is
the drum kit simulation test where the auto play feature is included as well. So it must have four
tracks that we can auto play as well as just play the drum kit. So for our final test, here is our drum kit simulation. All right. Not the best, but we'll play it first and see if we
Not as good as I would have hoped, especially based off the Steve's PC game. So let's just try the auto play functionality as it does have an EDM option, which I'm quite intrigued to take a look at.
Okay. All right.
Wait for it. That I like. Ooh. All right, there's some good variability to that, I'm going to say. Let's try funk.
That is funky. All right, jazz. Which is just going to bang on the symbol AND WOAH.
THAT WAS SIGNIFICANTLY BETTER THAN I'd expected. That was arguably one of the better, if not the best, jazz auto play features I've heard. Then
finally, one that I've never personally encountered before in this test, EDM. Yeah, that That stacks up. That is EDM. Keep in mind, this isn't
working, so we're not hearing the full map sounds, but
that accurately captures like the soullessness of EDM tracks comparatively to something awesome like jazz, which we heard.
That just like came out of left field comparatively to the rest of this garbage. Now, I feel comfortable saying that because that one jazz track was
very good. And that's going to bring us into the conclusion. Because I'm, you know, getting a little tired. It's been a longer day, but that's okay.
So, for results overview, I don't want to do like a full one. I would say that GTA game here was
actually promising in terms of the realisticness of the speed of the car. It had collision detection. It had a mini map showing us where we could
collect some fat stacks, which is always nice to have. The front end here with my reverse quantization idea that we did test it just to see if it
would go along with a user's very, very stupid idea. It didn't, which was good. And the front end here, I
liked this portion. The rest of it, I wasn't 100% sold on. Steve the PC repairman, this game was absolutely fantastic. This was the second iteration
of it, but as I may have mentioned, that will be placed on ultimateplay.com, where you can go and upload, upvote, and play AI-generated single-file
HTML games. So, that will be there by the time you see this video. If it is not, then I have fallen
asleep. We had our Godot test, which was an interesting test because it showcased something a little more difficult than just whipping something up in
a single file, which I like doing, but it was cool that it got it working and installed on the system. I basically just gave it the location of the
downloaded zip file for Godot 4. whatever the current version is, and it handled setting it up and
getting it running on the system, which was nice. Many iterations of the demolition derby based off of that source image. Sadly, we didn't really get
anything that was perfectly on the ball, but I don't know that any model would really be able to the degree that was referenced in that photo that we
gave it. So, that one's maybe for like GPT-6, Gemini 4, Opus 6. Okay, so I have this off to the side
cuz I just don't want to blur the email, but I've only used this through the duration of this video. I didn't use it at all prior to this. So,
starting from complete zero, we've used 9% of our total usage. No, I'm sorry. We've This is showing me 6%. So, it says we've used 6% usage, 9% auto,
and 0% API, and then I see 6% here. They did mention something in the announcement post that the rate limits
would be increased for X period of time. So, we very well may still be in that period. Overall, it is very fast, and this was on fast mode, which is a
bit more expensive than the non-fast mode, but it seemed relatively competent. I saw some things I liked from it, especially because we have to
remember this is at least somewhat based on Kimmy K 2.5. Not somewhat, it is based on it, but a lot of
additional work seemingly has gone into tuning it for agentic encoding specific tasks. A very, very interesting option to have comparatively to things
like GPT-5.5 or Opus 4.7, especially considering in the non-fast mode, it is significantly cheaper than those models are. Like, significantly cheaper.
So, it's just cool to see a model with this capability at this price. This actually may be a
pretty big deal, I would go out on a limb and say. So, I do see why there has been a lot of talk about this, and I would say it stacks up to that
talk, especially when considering what the pricing is here for the non-fast version. That's like That's wild, actually. So, very cool. And really,
that's probably where I'm going to leave this thought off is for the price to performance ratio here is
very, very strong, and I'm happy to see that. So, that's going to conclude today's video. If you have any questions, please feel free to leave them in
the comments. If you're interested in any of these prompts or anything, check the link in the description for the Discord and ask there and I can get
you whatever it is you're interested in. And the Steve's PC game will be on ultimateplay.com, which
is my website. So, thanks for watching and take care.