GPT-6 Sol Is VERY GOOD – Is THIS an Opus 5.5 Competitor?

0:00 Cool. That's a different um model than we've seen before. Oh, okay. Open AAI has released two new additions to the GPT6 family aimed at a better price to performance metric. So,

0:14 the one we're going to be focused on solely for today's video is GPT6 Soul. However, Luna also represents a very, very interesting value proposition, and we will be testing that as well shortly

0:26 in its own video so we can get more views on the channel. Now, before we get into it, do feel free to subscribe so we can get that 100K plaque. And let's get started by taking a look at the two new

0:36 additions to this family, mainly focused, of course, on Soul. And I do like this background cuz you can interact with it. Now, sort of the TLDDR of this announcement post is that these

0:45 models are both very good values when it comes to price toerformance ratio. Whereas GPT6 Astra is the most performant but also most expensive version, Soul and then also Luna

0:56 represent a very very good value in specific Luna. And we'll of course save this for the dedicated video. But 10 cents per million in and 50 cents per million out makes this basically one of

1:07 the most competitive models from a large lab in existence at least currently. So that's quite exciting. However, soul is also cheap and it has become cheaper from its predecessor where 56 soul was

1:19 the absolute basically top dog and now soul is no longer the top dog. This is under GPT6 Astra and we can see the price kind of reflects that at $2 per million in and $10 per million out. Half

1:32 the price of its predecessor and also half the price of Opus 5.5. I almost have a bit of trouble following Opus 55 with this video because that model was just absolutely phenomenal and very very

1:45 good-looking, I guess could be said. But we need to keep in mind that when we're testing this, this is half the price of it. So, we do have some benchmarks here. And this is probably one of the worst

1:54 charts I've seen ever just in terms of like the things that are used to specifically I just I don't like this. That may be a personal preference, but I'm not a huge fan of having to discern,

2:04 oh, which tinge of yellow is this sun model representing? Makes me mad with that. They also talk a lot again about lower cost. And for soul right here, they showcase it's 3.9 times

2:16 more expensive to run Astra on low. And hypothetically, Astra on low is being beat or matched by GPT6 Soul on either max or Xi. They also do include the Deep S. benchmark, which I find interesting

2:29 because I do believe there's almost Oh, I didn't realize you could hover over these. That makes this graph far less infuriating in my opinion. I do believe the deepsw SWE score is here as well.

2:40 Yes, it is. Now, something that I noticed that is a bit odd is it seems like there was a degradation in score from its predecessor to six soul. However, Deepsw SWE has also been

2:50 showing some perhaps not in line with reality scores lately for some models. So, I'm not going to harp on that too much. Additionally, kind of like what we have seen with the Opus 55 announcement,

3:00 it is apparently just nicer to speak to. Personally, I don't really care as much how they talk as long as they work well, but it's also nice to see, I guess, if there's less AI isms in the speech. So,

3:10 for some technical information, as we saw, this is $2 per million in and $10 per million out. It is multimodal in that it can accept images in as well, so it can see images, which is always a

3:20 good feature to have for some specific types of troubleshooting. It has a little over a million context window, 128,000 maximum output tokens, and a knowledge cutoff of late April 2026. So,

3:32 relatively recent in the scheme of things. Now, really, there's not a whole lot else to say about this. It was a relatively light introduction post as well, considering it was two models that

3:40 dropped at the same time, which I'm quite pleased with. So, of course, we're going to jump into our first test, which I have run. It has finished, and I have not at all looked at this yet. So,

3:48 that's why we have this little green icon here. So, this is, of course, the browser OS test v2.7. I was concerned it was going to show it to us in an artifact window here, and it didn't.

3:57 Okay, it named it Astro OS, and it worked for 23 minutes and 28 seconds on maximum effort. This is an option where if you don't see this in your app, you actually, I believe, have to enable it

4:08 in settings. So, it's possible if you just see extra high as the highest available one, you go in and then you enable this in settings. There is also ultra, but for now, we'll just run some

4:17 things on max. So, let's take a look at our browser OS. Oh, okay. I do believe this is pretty similar looking to the GPT6 Astro result. All the cache for every single

4:29 run and everything on this computer has been removed even since the Opus 55 video. So, just to make that point as well, welcome to your space, the universe at your fingertips. Drive a

4:40 city, defend the stars. Oh, that gives us a spoiler of the second game. Now, I may be mistaken, but I'm pretty sure this background is not moving. It is. It's just that's on me. I'll own

4:54 up to that. We do have a date in the top right and pulse. Okay. Seems like some form of chronological saving point. Animated wallpaper motion. Let's speed this up a bit. Okay, it is moving. Kind

5:09 of like a very sparse option list for wallpapers. apply seed. That's very interesting. So, you can use words to create the seed. Uh kind of like Yeah. All right. That's like some

5:23 Bitcoin hardware wallets or crypto. You can use like seed phrases and things like that. I think there was like an issue with some of them recently with that. Additionally, let's check if

5:32 there's a right click. There isn't. Perhaps playing Koi. Do not be accused of benchmaxing. Let's check our start menu. All right. Everything looks good here. And this is pretty similar for

5:42 what we saw with six Astra. All right, let's just get started. So far, I'm a bit concerned. Okay, this is just

5:55 traffic collision. Police alerted. Okay, well, we got hit. Oh. Oh, no. This is bad. Right here, what I'm seeing is not impressive. This is the smell of a small model. When the police snap to

6:07 your location and they just like hop over you, it showcases that it didn't really put some logic into. Okay. And this is inverted. And it's it's clean, but it's bad comparatively.

6:26 And we can't Okay, we can. It's just like I said, it's it's a little difficult to follow up the Opus result, but I would have expected better from this. Next up, let's just get this one

6:42 over with cuz it's our second game. Okay, not bad. Again, I do feel like there's some Astra like spice in this, if you will. Male.

6:54 All right. Acceptable looking. Nice color palette. some transparency in the background. And then just matter references to things on this specific operating system. Notes. Let's create a

7:05 new one. And we can also title them. Okay, that's nice. Something you don't actually really see very commonly here. So that's just some nice attention. And then we can delete it. 27 * 6

7:18 182 minus 20 * 5 810. Good. Pulse. A new star joined the desktop. Oh, no.

7:31 Wait. Oh, no. Wait. There we go. Okay. Interesting. So, maybe these are representative of more than just a background. Our pulse engine. Let's look

7:46 at settings. Oh, we already did look at this. This one's relatively simple, but you know, that's acceptable. Good. And we have a log of all of our specific events right here.

7:57 Interesting. I I don't know how we interact with these, but they seem to represent specific actions that have been logged or taken in the browser OS. So, that's interesting.

8:10 Overall, I'm going to hold off judgment till we see some tests that will allow it to better showcase its capabilities or lack thereof. So, one thing I did, and I ran this on ultra, this is

8:20 something I did just to test to see how it was doing. I said, "Make me a Runescape 2007 clone of Player Versus Player at the Grand Exchange." It did work for a little bit. So, we see that

8:30 basically took 22 minutes. I said, "This looks like a cheap app store ripoff. I want this pixel perfect and 3D." Then, it worked for an additional 23 minutes. This is actually, I would say,

8:40 respectable all things considered. It's not pixel perfect, but Oh, we got attacked. Vex the bold. Can we speak? Sit. Oh,

8:51 sit. Okay, the speech bubble's not showing, but it seems like we can swap our combat style. And it does. It seems like our player's using a crossbow now.

9:03 Magic special attack. Oh, that was the safe zone. Interesting. Did that just declaw speck him? I can't tell because it's kind of hard to see. No, it Oh, well, get seated. Oh, cool.

9:22 Skills. We have inventory. Are these anvils? Oh, it's a shark. We have prayer. Does it show the icon? Okay, it doesn't. But our prayer points may go down here if if anyone like knows

9:35 what I'm talking about. Good. Run is space. Okay. And we even have hot buttons there. And then these are just kind of some of the similar things in a mini map.

9:46 Overall, this is it's not bad. World 318 PvP. I just don't quite know what I expected and I have no prior like reference for a result here. So, I

10:00 wanted to do something just totally random. And I think this definitely fits that bill. Something we've not done before. Look, look, it changed us into like magic robes and then it just

10:09 changes our All right, let's special attack this individual. He's about to get seated. Next up, we're going to do the self-contained C++ skate game. This is

10:24 the one where it has to be the New York City block and it cannot use ray. And also, we're running this on max. I suppose I should have taken inventory of my usage prior to running these tests.

10:34 That is on me. Let's just look at it now to see. All right, so currently our weekly limit is at 92% left. Keep that in mind. I do have some resets available, but we're at least from here

10:45 where I ran the Runescape test. I also ran the browser OS and it's currently doing the robot arm test. So, we're at 92% left after those. And I don't know what I was at prior to beginning the

10:55 filming. All right, I came back to my computer to find this, which I have to say I'm actually somewhat satisfied with on first glance. It's I noticed also with the GTA game,

11:10 it's almost seems like it's doing cartoonish style aesthetics. Interestingly though, as well as with the GTA game, the controls are kind of inverted just for the left and right,

11:21 which was also the case when driving the car in the GTA game. So, that's kind of interesting. Q and E. Okay, the spin is a little off, but I will say this is simple. However, it's competent. So,

11:34 I'm okay with that. I don't have the ability to snap onto these rails. I'll say this is competent. It's simple, but it's competent. And I think this

11:47 would really benefit from a follow-up prompt just saying, "Okay, I'd like you to maybe like go hard on this now." So, I've given it a follow-up just basically saying, "Make this better and fix the

11:56 grinding logic not working." And this is on ultra. All right. I'm just like over here trying to work on this. My friend gave me this old Alienware laptop and it's I think the

12:08 motherboard's toast, but that's what I was doing the whole time. So, all right, back to this. This definitely looks like a visual fidelity overhaul. I do believe I also heard some sound effects.

12:20 Yes, but they're very quiet. So, let's turn them up. Let's see if the big the issue that we had. The only one was the grinding logic didn't work.

12:38 Let's try a different rail. That may not be registered as a rail. It's good. All right. This is definitely much

12:54 better. It's still simple, but that is okay. Are there any Okay, so we still can't collide with the pedestrians. Something like that is what I would have liked to have seen in a real overhaul

13:06 like this. However, it definitely did make the graphics much better. Like that right there was downright impressive. We have a fountain with some water and there are colliders on it, so we can't

13:18 actually go in the water. That's acceptable. I hate to say this, but I like the sound effects it chose to use. All right, come on. Come on. No.

13:34 All right. This was definitely a proper visual upgrade, which is what we wanted. And again, the thing to drive home is like not so long ago, this would have been like a, oh my goodness, this is the

13:46 best model yet. So, keep that in mind as well. And something I do want to notice just off the bat here, this does not seem to really take a very long time to produce these results just in the small

13:56 sample size we've seen so far. I also did the robot arm test and this made good progress, but something I noticed is that unfortunately it kind of misgripped the car. So, it did manage to

14:06 almost pick it up. The problem was it did not close the gripper far enough to make it a successful grab and then after that things didn't necessarily go very well. It moved it in a position

14:17 accidentally where it dropped. That was truthfully rather difficult to get it. So, it worked well. I think after about an hour and a half, I basically called it as it was still trying to pick the

14:26 car up from where it had dropped. I believe towards the end, I put it on fast mode and set the thinking to the lowest effort. But at that point, it was a little too late to have made like much

14:36 of a change because the car was in a bad spot. So, overall, not bad. It didn't get it 100%, but I noticed kind of in the way the big full Astro model had done, it was more kind of like, okay,

14:48 I'm going to move it, take a look at it, move it, and then just like gently grab it based off of visuals. And it's interesting to see the different approaches models take about that. Some

14:57 of them, like Fable 51, just went like, I'm going to make like simulation of this entire scene and then map it out. And that's cool, but it may not always actually be the best approach for a task

15:08 like that. So, overall, not a full success, but it did kind of pick the truck up, so not horrible either. Now, I'd like to try the Subway FPS. This is the version where the subsequent waves

15:19 get delivered via train. Now, I'm a bit concerned because trying to follow up the Opus result for this specific one may just be a lost cause for anything to look really impressive after that. I may

15:30 opt to following this, have it transpose this using Blender and GDAU just to see what we can get out of it in some different styles of test. All right, in 21 and a half minutes we have our

15:40 result. I will say, okay, I guess I'm going to not say anything until we look at it. It again reminds me a lot of the skate game where it's simple, but it's well

15:54 done. Well, if you don't care for bullet holes in the environment. All right, we do have some other weapon models. Yeah, just very simple. I have to say

16:10 though, I kind of I feel like this would be a result I'd expect from Luna and not necessarily from Soul, although it only took 21 12 minutes. I feel like GPT 56 Soul probably did a

16:26 better job on this. Okay, this is what 56 Soul did, so I'm going to eat my words on that one. This This is bad. What we just got from Sixole was significantly better than this, keeping

16:38 in mind neither of them were really that good all things considered. However, I think Oh, I tried to close the tab in the video. All right, so I'm definitely more bullish on this now. Actually,

16:48 after taking a look at what 56 sold did, perhaps I was looking back and expecting that to have been better. It definitely was not. So good. But again, something I kind of noticed, and this was the case

16:59 with Astra as well, I believe. Giving it two prompts just really pushes it over the edge in terms of its capability. And I'd like to do that with this. I know I've prompted this in kind of a

17:11 weird way, but I want it to seem like I'm asking it what it thinks of this. We know obviously it's going to be like, "Sure, let's do it." But I want it to feel responsible for having made that

17:21 decision to make this better using Blender and GDAU. I agree the current scene has reached the limits of simple geometry and animation. All right, that took a bit longer than I had

17:29 anticipated, which I'm happy about because I want to see this thing go all out. That took 47 minutes additionally to do what this did. Now, it's using GDO, but it's in the browser. My

17:41 knowledge of GDAU is probably going to stop here because I'm a little Let's just see. That is significantly better. Okay, at least the weapon model is. Yeah, that I like this. It still

17:59 keeps its weird like cartoonish aesthetic though, but I like it. Check these are there. Nope. I'm going to say this is actually quite

18:15 interesting because it kept the aesthetic and just the overall vibe of the game so so similar, but it just massively graphically overhauled it. And perhaps someone in the comments can

18:26 explain to me how exactly this works. So I would assume then like you can just make a game with GDAU that runs in the browser. Hopefully I'm right. Train model is much better. Let's check

18:42 the doors opening. Good. We can't go in it, but there is an interior. All right. This is definitely an improvement. The biggest nitpick here is that there are no bullet holes in the environment, but

18:53 that's okay. The sound design is terrible. I'm going to just say as well, tickets. Cool. That even has like a little screen. We do have some individual tile effects and

19:10 things of the sort. And we have one individual here left. Cool. That's a different um model than we've seen before. Oh, okay. You went down. That was

19:24 interesting. I was actually okay with that. And it's so fascinating that like here's the one that it based it on. It kept the same style so so similarly. It just massively improved the visuals by

19:35 using Blender and GDAU. So, that was actually kind of cool and I'm happy we did that. And that was on ultra and it took about 50 minutes past that. All right. Next up, we're going to do the

19:45 watch website test. And I want to just put this on ultra now, perhaps for a few of these. I'm now seeing also the large amount of spelling errors in some of these prompts, but that's okay cuz the

19:54 AIs can understand intent. This is also, of course, the prompt where it needs to include this photo in the face of a watch dubbed the Beijan as a special edition. So, we'll see what we get here.

20:05 In just under 20 minutes, using ultra code, our watch website has been completed. Okay. So far, I'm a bit distracted just by this white artifacting at the bottom. And it does

20:15 seem to be lagging a bit. Oh, no. Okay. The issues I'm seeing here, especially with the dial, the numeral indicators being mismatched. That is a little

20:25 disappointing and something I would not expect from a model of this caliber. One below Astra is not necessarily a bad model, or it shouldn't be. So, first impressions, not the best. A little

20:36 Audacity. Looks good on time. All right, summer 2026. Things get a little better down here. We still can't actually see these. However, can we select a different one and change the

20:48 one that's appearing in the header? Okay, we can our point of view and it brings us here. I will say front end, not bad. Just ignoring like the actual models, this section, the off-white, not

20:59 bad at all. Now, let's see if the exploded view looks good. Very simple. not hyperdetailed. However, it is competent and together

21:14 and then of course we have the ben which I'm always going to like. So maybe this is a bad test because it's just far too subjective. But I will say the big issue I'm noticing is the numeral

21:26 markers are just messed up. And that's something we see on like cheaper openweight models, not necessarily something that still should be a Frontier model. So again, this one kind

21:35 of reminds me I'd mentioned in a previous result where it's like what I would have expected from Luna, not from Soul. So I'd like to try something next that's just drawing more and it's 3D

21:46 design capabilities. I know, surprise, surprise, but I like this stuff and it's fun, at least for me. I'm saying I'd like you to create a highly detailed 3D model of a 1980s Saab 900 Turbo in

21:56 Blender. then port it to a website where users can go and interact with the model. Opening the doors, looking inside, opening the trunk and the hood, etc. I'd like this to be what a high-end

22:08 website for this car would look like were it to be on sale today. Realism is paramount. Allow different color combinations and trims to be selected and go all out. All right, so I had seen

22:18 what this was doing and I was going to head out for a second. So, I basically sent it a follow-up that was just cued until it had finished, yelling, saying it looks jank. Now, it didn't do a bad

22:29 job. This has elements of it, and honestly, this is cool in and of itself. The big problem is this looks nothing like a sub 900 turbo. The wheels options are more or less correct, so I like to

22:42 see that. However, I feel it prudent to respond to this with a meme, and we'll see how the model assesses my feedback in image form. All right. So, I'm using chat GBT just from within the web chat

22:53 interface to generate the response I'm going to give this model where I've looked up a photo of a mid1980s Volkswagen Shurocco and I've also taken a photo of this Saab 900 model that it's

23:05 given us and I asked it to create the meme from the office where it shows two pictures. It's like corporate wants you to find the difference between these photos and it's basically they're the

23:14 same picture and I'm going to send that back to it to see how it assesses my feedback. Spot on. It's pretty much exactly what I need. So, let's send it this now. And keep in mind, this is only

23:23 the stroke of like sarcasm that could come like after 4:00 a.m. when you've been testing models all day. So, we'll see how it resp we'll see how it responds to my

23:34 feedback. Fair point. Yeah. Okay. So, it did it did understand my meme response. I like that. Overall though, I mean, like I think the the hood actually does open that way.

23:49 And I'm going to be honest with you, the actual engine is is not as bad as I mean, it's well done, at least in terms of that. All right, so after 46 minutes and then an additional power nap

24:01 on my part, we have our updated model. And I will say it's still funky, but this is definitely more in line with what I'd hoped to have seen, especially like the front bumper and things like

24:11 that do look a lot more realistic. and the curvature of the windshield, which is a real thing on that specific vehicle. So, overall, it did definitely get some of it. So, let's just take a

24:20 look at this natively in the browser. You know, it's actually not horrible. I would say I kind of dig this. Now, let's open the hood. Very good. And it didn't change anything in there, but

24:34 that's fine because that was actually more or less pretty decently done. Though, the battery is on the wrong side. Should be over there. But um we can open the doors, we can look

24:44 inside of it, we can open the hatchback, and there is a little cargo shelf thing there like to hide your cargo. I think this is kind of cool, right? And then we have presets for different views. The

24:58 engine, the hatch, the way the camera moves. I think we can also like you can download the photo. Hey, and it saves it in a partially transparent background, but the shadows are still there.

25:10 Interesting. That's a feature that's nice. I didn't necessarily ask for it, but it's good to have more, not less. Let's take a look at some of the interior options. All right,

25:23 not bad. Um, I would be curious to see what some other models would do with this, namely like the full-fledged Astra. Share your spec. Configuration link copied. Be the judge of that.

25:34 Okay, it did. It just had it like tacked onto the URL. Overall, I think that's kind of cool. So, for the next test, I gave it the Blender and GDAU prompt where it has to create the 3D wrestling

25:44 game that is 1980s themed. It worked for just 31 minutes. That's not a lot of time. And we basically have a preview right there. But fortunately, that's not spoiling too much of it. So, if I go to

25:55 this directory and run play.sh, we should see our wrestling result. And this was done on Max as well. So, here is our wrestling game, Ring Riot 88. So far, it's simple, but that's not

26:06 necessarily a bad thing. I do have music on. Okay, that's all right. Maybe it plays when we get in. You know what? These models are actually better than I was expecting considering

26:17 the ring and everything like that is kind of simple. They're cartoonish style, but they're not actually bad. All right. It's kind of simple, but it's actually

26:34 not bad really because the models it made are are acceptable. And it would have had to have rigged these as well. There is some simple movement and things like that. Admittedly, I'm having a bit

26:43 of trouble. There we go. Oh, okay. So, I think we just got knocked out. Keep the crowd loud. This is Oh, E is for chair. All right.

27:11 Yes. Y to swing. Oh, nice. Up. All right, that was It was competent. Kind of in the theme of all the other things we've seen where it's just decent. It's nothing special,

27:34 but it's overall competent. Now, something else I did because I would like to test something that's a little more on the software side and maybe less of a visually striking demo. I gave it

27:45 and yes, we've swapped to a Mac just for this one because it has the full tool chain for this task. I wrote to it and said, "I am at a hackathon for iOS and have 90 minutes to make an app for the

27:55 new iPhone 18 Pro Max, something that takes advantage of its new hardware. It started 3 minutes ago and I legitimately have zero inspiration or ideas. We are allowed to use a coding agent, make

28:06 something insane that will do very well. I will attach the phone to this Mac when it's ready to live transfer the app. We also need to ensure it has an app icon as they're scoring heavily on good

28:16 aesthetic as well." So, it worked for basically it asked me one question just are there any rules or anything like that. I said no, just use the newest iOS and I don't actually see how long this

28:27 worked for specifically. It was run on fast mode and also on max effort. So, keep that in mind. I don't have unfortunately a static time value here of how long this took, but I'm

28:38 going to tell it now. The phone is attached. So, let's actually get this running on it. I don't actually still know what the app is yet. I guess we should read a TLDDR. Echo frame is built

28:47 and ready for the Pro Max. It turns the live camera into a dragable window through the last few seconds. Four metal effects reshape the history. Metal being like a Apple like a hardware thing.

28:57 Think of it as like a like CUDA or something. I think ghost mode uses ondevice person segmentation to keep someone in the present while the room falls into the past. Has a custom app

29:07 icon and a sharable still image capture. Apple lists the A20 Pro's first 7 core GPU and dual neural engine. Okay, sick. This is running on the phone right now. I accidentally hit that new button on

29:19 the side that I'm not used to that I keep hitting. So, I don't know how well this is going to be visible and I will perhaps start a screen recording on the phone. Although, if this app uses the

29:28 camera, I don't know if that will work or not. Let's just enter live camera. Okay, now I think something's gone a bit ary here. It had asked me a question and I didn't answer in time.

29:40 So, it's continuing. But it is also possible because I'm screen recording. So, let me stop the screen recording and let me reopen this app. Let me just close everything. Okay. Yeah,

29:52 unfortunately, it's still not working. So, let's check. All right. It's trying to capture frames to confirm the image. So, it said, "Please point this at a lit room." And this is also just something,

30:02 this is a harder thing, I guess, to test on video, but just showing its usefulness in actually performing tasks like this where, oh, I'm debugging this real device right now for this app that

30:11 it just made, and it's just basically like QA testing. All right, it's saying check the Apple camera. That's an acceptable request. Son of a That's black, too. I've noticed

30:25 that's a glitch that this phone has in software. I've noticed that independently of this app actually doing anything. So, let me Yeah, look at that. And it this has nothing to do with the

30:37 phone. I mean, with this app. Oh, I only have 4 seconds. I'm going to restart this phone. This phone so far has been like genuinely a piece of junk. I'm using my 15 Pro Max still to record this

30:49 video. The whole reason I bought this was to have a higher quality camera for cinematic mode. And nope. All right, I've restarted it and the camera is now working. This is the default camera app.

31:00 So now I'm going to go back into Echo Frame. Enter live camera. Nice. Good. So it wasn't even a problem with this. It was a problem with the actual iPhone camera. That's upsetting,

31:09 but it Whoa. Oh my goodness. I don't quite know how to describe. Actually, this is kind of impressive because do you see these funky effects right now that it's doing? I'm going to

31:23 try to like blindly click on these things. This is actually entirely being run by the like phone's GPU and processing these effects, which is why it may look kind of laggy and things

31:37 like that. This is actually downright cool. All right, so check this out. It's like, and when you move the time depth up, this is the ghosting one. So, it keeps things like where they were and

31:50 overlays it. Prism again with time depth and this is using the phone's actual hardware. This is something that would actually be difficult to process from a hardware standpoint I imagine. So, and

32:02 then we can really make this like crazy then rift. So, low time depth and then high drag to bend time. Oh, I think time depth is is drag. Hold the moment. What

32:18 happens if we Oh, a slice of yours to keep. Interesting. Let's check our info. Move. Drag anywhere to move the portal

32:34 through the frame. Huh. All right. So, you're saying that I can Oh, yeah. That's That makes a lot more sense in terms of how this would be

32:48 interactive, doesn't it? It's like warping the that is prism ghost. Let me try it. Like

33:11 I don't know what this is going to look like. And then ghost. I think ghost was cool. Ew. I see myself when I turn it because it's lagging.

33:31 Or I did. Oh, this is funky. Cool. And then we get like that. That was actually kind of cool. So let's take a look at the readme and a bit more specifically about what exactly this is

33:50 and how it works. All right, so here's the cool like technical things of this metal rendering. A full screen shader blends 20 recent camera frames in real time. The A20 Pro has a seven core GPU

34:00 on device vision. Ghost mode generates a person matt and comp and compo composits them against a timeshifted world. It needs no account or network. The visual memory is held in GPU texture slices

34:12 avoiding perframe image uploads or remote processing. designed for the Pro Max display, the live canvas, precise rectacle, tactile mode controls, and custom icon form one visual identity.

34:22 And then we have the 45se secondond judge demo right here. So, this was actually pretty cool and I wanted to showcase it in something that was a little less of a game and more of like a

34:31 piece of software and also even a bit of troubleshooting. As we saw there, the camera wasn't working and that was not this thing's fault. It was the phone glitching out, which seems to happen a

34:40 lot with this 18 Pro. So, with that, that was cool just to see it like, okay, I'm going to reinstall the app and seeing some ability to troubleshoot a connected device through the terminal or

34:50 whatever connection. Overall, I'm actually quite pleased with this and it makes me more impressed with this model. All right, next up, I'm going to put this on ultra mode, and we're giving

34:58 this the brand new prompt, which is the guitar store game where you go around, you can play all of the instruments, and then it turns into a beat them up if you make too much noise. This was something

35:07 we first ran with the Opus 55 and it was just really hilarious and a lot of fun. So, we'll see what we get. I'll be very interested. And this is on ultra. All right, so our Guitar Store result is

35:18 done. And what I'm noticing right here just off the bat from this preview is the UI, at least the start screen looks very, very similar to the Opus result. I would like to make it a point that these

35:29 were run on entirely separate systems. So, the Opus result was not even run on this physical computer. So, anything similar here is just coincidential and based off of the prompt. Interesting to

35:40 see sometimes when there's like convergence like that. Now, with that, let's just look at this natively in the browser. I will ensure that the sound is on.

35:50 E is to demo gear. Okay, we can jam notes. Click is to punch and Q is to dodge. Okay, it's given us kind of like a firstperson view here. Definitely a simpler take on this. However, I'm going

36:01 to be honest with you. The individual models and things like that are respectable. I think this overall, especially considering the time, I would say this

36:12 is actually a decent result. And we can there's like hotspots so you can Okay, it's muted. So, we need to good store noise. Okay.

36:34 Okay, I like this. Power cord has to recharge. Some drum kits. This is well done. We have the acoustic guitar wall. Now, we can't pick them up, but

36:52 Okay, so space is also to punch. Let's just look at some of these. Oh. And they have different sounds as well. I like that. Cool. We can play all of these.

37:14 Okay. It doesn't seem like we can do anything with the amps. I like this. Now, yes, this is much simpler and less detailed than the Opus result was yesterday. However, we have

37:28 to remember, one, this is half the price, but two, it's also fun in its own way. It's a cartoon firstperson style. Definitely more emphasis on the beat him up. And judged individually, had this

37:39 been the first time I'd ever run this prompt, I'd be like, "This is this is awesome." So, I like this. It's simpler, but I very much like this. It's fun to play. And in some ways, perhaps simpler

37:49 can be a good thing for this. All right. I like that a lot. So overall, that's going to conclude our first look and test of GPT6 Soul. I have to say the model is basically just

38:04 competent in any of our tests. It didn't make mistakes really. It didn't omit things. It didn't do a bad job. But the flip side of that is it didn't do anything that was really jaw-droppingly

38:14 impressive. We have to remember pricing wise, this is half the price of Opus 5.5, and it really comes down to the individual users judgment on whether or not the cheaper price is worth it. And

38:26 that will be very dependent on your specific task. The model's competent and it seems to follow instructions decently well based off of what we saw. I find that like with Astra, giving it a

38:37 follow-up prompt to be like, "Okay, you made a good base, but like can you go really hardcore with this now tends to produce significantly improved results and sometimes I wish it would just do

38:46 that off the bat rather than needing another run to actually create what I think it can have the ability to from the gate. So, let's take a look at our usage as well as the final thing. So,

38:58 our weekly usage now is down to 87%. I am fairly certain that it was at 94% or so when we started. I will just ensure if I'm wrong, I'll put that value on the screen right now. That's really not a

39:09 lot of usage that got drained. I believe 7% or so. And we ran a bunch of different things here. Some on ultra, some on X higher, whatever the max, whatever is under ultra. So overall,

39:20 this seems to provide a decent amount of usage, at least considering that the plan didn't go down too much. Pretty good. And again, I don't have a whole lot to say because one, it's tough to

39:29 follow Opus 55, and two, none of the results really absolutely blew me away. However, overall, they were competent. And I think the biggest takeaway is it represents a decreased value with

39:40 reasonable performance. And this is something where 6 months or so ago, I believe this would have been a jaw-dropper state-of-the-art model. Also, I will say that all of these

39:50 results were done very quickly. Even on ultra or maximum reasoning mode, this model does not seem to take a very long time to complete its tasks, which I like to see even on those higher reasoning

40:00 efforts. So, that's going to conclude our first look and test of GPT6 Soul. Next up, we'll be testing GPT6 Luna, which I'm very excited about because it is just so darn cheap that any decent

40:11 level of performance in that for that price is going to be absolutely awesome. So, if you have any questions, please feel free to leave them in the comments.