GPT-5.6 Luna First Test – Hands-On With OpenAI’s CHEAPEST Model!

summarized

TLDR

OpenAI's GPT-5.6 Luna is now its cheapest model after an 80% price cut, costing $0.20 per million input tokens and $1.20 per million output tokens. In hands-on testing with coding and 3D tasks, the model delivered competent but inconsistent results—impressive for its price on some tests (like a 3D watch website and a skateboarding game) but poor on others (like a browser OS and a 3D-printable engine). The total cost for all tests was only $1.34, making it a viable option for cheap, repetitive tasks where top-tier quality isn't required.

Key points

  • GPT-5.6 Luna's price was reduced by 80% to $0.20 per million input tokens and $1.20 per million output tokens, with higher rates for prompts over 272k tokens.
  • The model scored its highest intelligence on 'Max Reasoning' mode, which was used for all tests in the video.
  • The browser OS test produced an unusually weird, dashboard-like result that the tester found confusing and low-quality.
  • The C++ skateboarding game was competent with decent animations and tricks, though it had a bug where the skater got stuck on rails.
  • The 3D watch website test was a standout, producing a clean, aesthetically pleasing site with good 3D watch models for a very low cost.
  • The 3D-printable engine test failed because the model ignored the 'little to no supports' requirement, making it impractical to print.
  • The Chrono City test in C++ showed simple but functional time-period transitions, though pedestrians were static and the soundtrack was poor.
  • The Street Eat game improved after feedback, with decent low-poly graphics and physics, but had issues with audio timing and tiny vehicle sizes.

Tools mentioned

Techniques

  • Max Reasoning mode
  • Multimodal coding with image input
  • API-based testing for cost tracking
  • C++ development without RayLib
Transcript (captions)
We spent $1.34 on the entirety of everything we tested here today. >> Okay, >> today we're going to be testing OpenAI's cheapest model, which was just lowered in price. So, this is GPT56 Luna. This came out alongside Soul Luna and Terra as the middle tier model. This is the lowest end one. And just a couple of days ago, they dropped the price in this model rather significantly. I do believe there was an 80% reduction in price bringing this to 20 cents per million input tokens and a$120 per million output. Now there is a bit more nuance when the output becomes or input becomes over 276k tokens. It gets a bit more expensive. So there are some additional like gotchas right there. As they say prompts with 272k input tokens are priced at two times input and 1.5 times output for the full request. So, just keep that in mind that this is not a static price across every single potential generation, but more or less this is a big big reduction in price compared to how this was on release. And because it's so cheap, I want to really test the capabilities and see how well this model can perform given the fact that this is probably the cheapest Frontier like state-of-the-art model from a big lab. Well, at least from OpenAI. Now, as we can see here, and before we get into it, please do feel free to subscribe so I can get the 100K plaque. But a couple days ago on July 30th, they did this price decrease for GPT56 Luna, which we're testing today, as well as Terra Soul has maintained its pricing that it was released with on I believe July 9th is when those came out. So, we can see here though that this model right here on maximum reasoning mode is pretty decent when factoring in the cost per task as well as performance, seemingly outco competing GLM 5.2, which is a very well reggarded openweight model, which this is not. So keep that in mind as well. And one thing that is missing from here being that it did just come out a day after this price decrease is the new DeepSeek V4 Flash0731 model, which we did just test. And at its maximum potential cost from the official Deepseek API, that model's around 56 cents per million output tokens. So it's still going to be cheaper than this, but it'll be interesting to see how this model stacks up being that that Deepseek model is the most recent one we tested prior to this video coming out. So, at least the results should be fresh in folks minds and mine as well. For a bit of information about this model and then in the specific way we're going to be testing for today, it has a maximum context window of a bit over a million. It is multimodal, so it can also see images, which is something that the Deep Seek V4 flash model definitely could have used to improve some of its results. So I will also be doing some specific image input tests as well to see how this model is able to transpose images into 3D space or something like that. And really the only other thing to mention is how we're going to be running this today. So I am opting to use this just from within the codec that I have installed on this system. I'm going to be running this only from within max reasoning mode. In that page we were just on which showed the chart of the cost for performance for this model. This scored its highest potential intelligence score on Max Reasoning. That is the highest for this specific model and I want to test it full performance. Now, I'm going to begin this browser OS test fee 2.5. But before we do that, I want to make specific note that I'm going to be using this today entirely from the API, not through a subscription or anything. So, we'll be able to see exactly how much the entirety of today's test video actually does cost. All right, in just under 5 minutes, we got our result. Now something interesting I noticed is this basically named it like ominous again which the deepseek one did as well. They were both being used through this codeex install. I like completely deleted any open stuff when installing like the deepseek patch that they had. I am almost wondering if it's because the organization name is ominous as we can see down here in the bottom left if that's what is attributable to that behavior. But I did just want to point that out there because I found that to be quite curious. So, with that, we are met with the weirdest browser OS result I have ever seen in my entire life. I'm genuinely going to need like a few seconds here to make sense of exactly what we're looking at. It's so weird. This is almost less of an OS and more of like a dashboard, but we'll just try to go through it as we would normally. I'll turn my speaker on. We have the correct time and date as well as a 82% battery percentage which is not necessarily um the specific one for this computer. So that's just emulated in the top right app launcher. Very very weird. It's almost like was it Windows 8 that was kind of like this? It had like a much different approach to like All right. So I believe this is the menu bar. Yeah. I need a guide. Yes. Okay. One tab, a small universe. All right. Let's just click on the Launchpad here. Oh, okay. So, these are just files. So, this would be like a a file browser. Okay, not bad. I'm actually worried if we open some of these, it might just open some of the applications here. Okay, here's our notepad that we're going to be able to take a look at. I'm going to make note that because of how busy this background is, again, this is really without a doubt one of the most different results I've ever seen for this test ever, the transparency is a bit distracting because of all the elements here in this background. Can we make a new folder? Okay, we can. Can we open it? Well, but we can make a new file as well. Markdown. Can I open that? I can. Okay, that's cool. Next up, notes. We already saw that. Pulse. a quiet inbox for your signals. Oh, so it made like a fake messaging app. This is really quite a terminal. All right, here's our GTA game. Let's full screen it. All right, the sound is on. I'll just like quadruple check that because it's been a known point of contention in some of our tests. This is Is that a car that Okay, it is. So, we're in this car and the people are gigantic. Perhaps this is a remote control car we're driving. This is different, but I'm I'm not going to sugarcoat this. This is quite bad. It's just a total like I can't say that but it looks like this model took like some form of mindaltering substance before actually performing this task and giving us any of these results. I also so don't see the ability to actually close this window. Now I didn't even check if there was a right click because I've been so confused by this. No, there isn't. But all right, let's just get through this skyline drift. Okay, perfect line. Hole scrape again. We have like a very very micro thing that is showcasing us and then we need to avoid these obstacles while picking up some of these. Okay. Settings. Sundown deep forest graphite accent color. Okay. It changes the text right there. The last thing is the workspace snapshots. Restore it from settings note session. Okay. So, it did properly do that. I don't really I don't know what to make of this. This is just so odd. Wow. Glad you like it. Interesting. All right. So, I basically told it like, "This is crap. Goodbye." A lot of times models will be like, "Hey, man. I'm really sorry about that. Like, can I try to improve it? Like, sometimes they'll just start trying to improve it. This one was like whatever. Like, adios. I kind of respect it more for that, I have to say. All right. This was um All right, so here's the self-contained C++ skate game. This is the California Boardwalk aesthetic one. I have not specifically mentioned it can't use Ray Lib. I'm assuming it won't. I'm going to go rinse my eyes out after this browser OS. All right, it seems like our result is all set. And something that it was doing as well while creating this is just taking a screenshot of it and then verifying that everything looked good, which it said it did. So, let's take a peek. You know what? This is actually quite good. I'm okay with this, especially considering the size and price or cost of this model. This is a fairly competent result. The clouds are decent. The water effect is there, albeit a bit simple. And one of the big things is like I don't see any Oh, okay. If we maybe close up on some of the ways that these folks are walking. Okay. And and also that it's not perfect, but it's definitely competent for a model of this caliber, judging it based off of its size and intelligence score from the intelligence index. Okay. Kick flip combo. U is to do a shove it. That was kind of cool. Look at the way the legs twist like that. Perhaps somewhat painful, but our skater is composed of entirely code, so they may not feel things the way we do. Yeah, that walking is looks like spaghetti noodles, but still, it's actually not bad. The shops have text on them and everything looks decent for the text. Playability is fine. There are some umbrellas there, some benches. Really, one of the biggest issues, I think, is just the issue that happened when we got on the rail there, where we kind of got stuck to the end of it. And that's something that's like it happens more often than you would think. So it's a pretty consistent bug to see in these results, but overall this was respectable. Wow. No, I'm just kidding. I'm not going to glad it landed. Yeah, this time it was it did. So let's do a price check real quick just in terms of what we use so far for those two results. The browser OS and then the C++ skate game. Okay, so we're up to 16 cents so far. So those cost about 16 cents. Next up, I'm giving it the unified subway FPS prompt. So, it has to create the station that looks good in and of itself. And then following that, it just needs to use that beautiful station to create a sick FPS. So, interestingly, that took the longest so far out of all the tests we've run, which is interesting. But then again, I guess we're hopeful for a good result. Idiot. Don't tell me to run it like that if you've already used the port. Okay. Nightfall line07. Very James Bondish, I suppose. Maybe just cuz that looks like Skyfall and then that made me think of Double7. Nonetheless, you know what? I kind of like the vibe of this. I don't know that it's like the best, but it's definitely not the worst. And it does have almost like a synth wave looking effect to it in my opinion. Let me make sure the speakers up. Can we shoot? Yeah, we can. Okay. There's no bullet holes in the environment. And the controls are inverted. Oh, are they? I guess not. All right, let's see. Yeah, they're they're a little funky. Nice. Target neutralized. It's very clean in terms of like there's no, you know, like Okay, the reload was kind of cool. This is almost giving me like Halo vibes. I don't know why. Just the weapon. I think the LEDs on top of this weapon have reminded me of something. Platform clear. Next contact. So that means we have some time to look around. All right, we have tracks. We have the do not cross line. There are some posters there. A lot of times this happens where the actual backing that the poster should be placed on is oriented the wrong way. That's just like a small model smell kind of thing, I guess. All right, let's see what happens if we lose or biometric vitals. Okay. All right. And we do have like flash. Okay. And then it just restarts us. Nothing super special, but nothing bad either. So again, just a competent, albeit simple result. Okay, we're at 44 cents total now. I'll figure it out in editing what just happened there. So just keep in mind we're at 44 cents total for today. All right, so I want to try to give this a multimodal coding task in the form of the Jerry's apartment from Seinfeld. I've put it in its own empty repository. except well, I guess it's not fully empty because it has this screenshot that I consistently give these models when giving them this test, assuming they do have vision capabilities. And I've also just appended to the prompt that you should reference the photo if you need help. So, we'll see what we get here. All right, let's take a look at Jerry's apartment. That was 6 and 1/2 minutes, so not a lot of time. And I noticed that it was just give it a second here. I noticed that it was doing a lot of Q&A checking. You know what? This is not very true to the actual apartment as shown on the show, but I have to say as a scene, okay, maybe the more I look at it, the less I was going to say what I planned on saying, but I do like the way this presents, I guess should be said. Maybe truthfully, I think that I like the color that it chose for the floor. And I think the background blue here also is making this look a little more high quality or elegant, I suppose could be said. So, keep in mind that colors and things like that can influence my perception and judgment of any one of these results. With that, though, I kind of like what I'm seeing, even if it's not all 100% right. There's some odd behavior here. And the way these couch faces are drawn almost reminds me of some of the oddness that we saw with the GTA clone result where it just has some weird like faces and things. Are these like points of interest? Maybe if we just click like this. Yep. Okay. And it just spawns us to those. The bike is actually pretty well done, even if it's the wrong color, even though it's pretty iconic that that's a lime green bike. But I guess that's okay. We have the different places here. That is one heck of a bathroom. Looks like something that would be the first thing they go at on like a home renovation show. And then I think that's supposed to be the computer nook. Again, the viewer can judge as well because, you know, let's see how much that one ran us. But we're currently at 44 cents and if we refresh we're now at 58 cents. So that cost 14 cents. All right. So next up I want to give this a more directed 3D model CAD test where it needs to create a 3D printed model of an inline 6 engine that will fit inside of it an N20 motor. Now an N20 motor is a pretty tiny little RC motor. The reason being, I want to actually print this because I'm asking for a bit more detail and I'm giving it to a specific reference item aside from just a generic inline 6 motor. All right, by far the longest amount of time we've waited for a result, a little over 17 minutes to get our 3D model of the motor. So, we went from 64 cents total to 89 cents. So 25 cents on the dot it took to perform that specific task, which I mean a quarter for a 3D model that may actually be something that could be used in real life like a RC car. That's not bad. So let's peek at it. Let's just look at the open SCAD. And I did see mentions of like a turbocharger as well. So it's possible that it created one of those as well. Okay, you know what? It's a bit simple. I'm going to hold off on judgment until we see some of these parts individually and maybe not as a whole there. So, let's just look at them individually. So, here's our base. Okay. Um, yeah, maybe. All right. I'm noticing a a few issues here that are quite perplexing considering my prompt said it needs to be printable. This is going to need a ton of supports. Not on the sides here, that's fine. But the bottom is just a disaster from a support standpoint. I mean, this should All right. And then also, this may not be an issue because depending on how long this valley is, we can probably just slip the motor down in here and then push it forward so the shaft would come out there. But and then that didn't need supports, too. All right. And this is just to check if the motor fits. Okay, that's fine. Here's the head, which goes on top of the motor and the base that we had open before. Again, not bad, but supportwise, it's going to be a nightmare to print this. I mean, it wouldn't be that bad, but still, this is not little to no support. Okay, another cap there. Maybe holes for the wire, so that's good to see. And then finally, the turbo. This is what a turbocharger looks like. And it's just not quite I'm not going to print this. This is just bad. It's It's not what I wanted. And it's not even about that. It's just that the support adherence, like I said, little to no supports. This in and of itself is already like a disqualifier for that. And then add just not quite there. Dimensionally, it may very well be fairly accurate to the specific thing that we're trying to accomplish. But detail-wise, it's quite quite poor. All right. Next up, I'm going to be giving this the beautiful watch website for the Slapest Watch Company. It needs to just create a high-end website with a 3D model of a watch, a cinematic panning effect in the hero section, and then just a good-looking overall website. All right, in 3 minutes and 13 seconds, we got our result. And we're going to notice that I have made a mistake because I did not specifically denote that this must be 3D, which is on me. Though, I'm actually going to be honest with you here. I actually think this is quite good just from a front-end aesthetic standpoint. It's refreshingly not black and gold, which we've seen for like every single model I've run this test with just does that. So, if we kind of omit the fact that I messed up by not denoting these must be 3D renders, it actually did a good job. And even for like 2D SVG-ish looking renders, this is actually pretty good. So, we have the Soul, which is an interesting choice for a watch name, and then Orma. And they're not super expensive either. I like the way it has put them here in these renders. I actually am quite pleased with this result. I'm going to try it again just with 3D included in the prompt. But I I like this UI. So, I'm just going to do that as a follow-up here, saying this is fantastic, but I need the watches to be 3D models rendered and visible on the page. So, hopefully it keeps this pretty good-looking aesthetic and just adds in some 3D watches. There we go. Okay, so it's swapped. And this is actually quite well done. For a model of this price, this is absolutely by far one of the better 3D watch models I've seen. Okay, now that I say that and I'm looking a bit closer, we can see some of the tick marks for each individual numeral get a bit funky when rotating around the circle, but this is still actually not bad at all. It has placed it on a table. The crown is very visible. It even has a little date selector thing. The reason I'm not like using the mouse is because it's orbiting by itself, and if I do that, it kind of messes things up. But this is actually overall quite competent, I would say, especially considering models of this size and what I've seen. This is a really nicely done result. Let's take a look now at the two individual renders. And honestly, these are almost arguably a bit more impressive. Maybe I just like the green one, and I'm biased, but I think this is very nicely done. And it kept the overall aesthetic as well. Okay, this was kind of kept here, but I guess I didn't specifically ask it to change that. So overall, this was a very, very nice improvement over what we had seen right here. And we've used so far $1.3. So don't get confused by the erroneous 6 cents there that was included from a different date in a different test. Next up, I'm going to be giving this a more difficult test that I did with the Deep Seek Flash model as well. This is to do the Chrono City test where it has to show a city over a bunch of different time periods and have good-looking transitions between the time periods. However, it is being instructed to do this using C++ without RayLib. So, this is going to add definitely a level of complexity to this test. But I have to say, I had a lot of fun seeing what Deep Seek Flash did as that was the first time I'd run this test telling it to use C++. So, I want to do it with this model as well as they are somewhat comparable as well. All right, we're getting our first look at this. So, I've just been like playing guitar while I wait for these results. But overall, that's actually not bad. This test tends to get kind of spoiled when we do it with C++ because oftentimes they're gonna like look at their results and make sure they run and things like this. So, it seems like this would probably be 2055. Now, I don't want to touch it, but normally they time out by now. Okay, good. Right as I said that, but it's we're just going to leave it and see what it does. All right, let's take a look at how much the Chrono City in C++ cost. Okay, so we went up to A120. I think we were at A$19. It is all set. Let's take a peek at it. And we also do have sound as well. And keep in mind that this is entirely a self-contained. Well, not self-contained, but this is entirely done in C++. Okay, it starts in 2055. We can move around. There are people there. We have WD to move around. Okay, the people are just like floating. This looks like a very eco-friendly and sterile future. Hey, that's kind of cool. When we zoom out, there are also like buildings in the background. Okay. I mean, it's pretty simple. Are the street lights changing color? Yeah, they are. Check that out. They're cycling quite rapidly, but do we have like hover tool tips on anything? No. Okay. It's the symbiotic district. Bioceramic and lumen autonomous pods, wearable and airrated. Okay. Interesting. Let's now go because it started us here. Let's just That was an awful awful sound. Okay, we have the same street lights. They're also working here. Pretty much like very similar. The color palette changed a bit, but unfortunately, one of the biggest disappointments here is that none of the people are moving. Although I almost swear I see different sized people. So that's cool. I guess that it's paid some mind to differences in um size. And then also just a basically just like the same thing but green. Let's check 2005. I'll turn the speaker back up. Okay, it's still got that awful sound. So for now, we'll keep that off. And then it's also changing what we see right here. Glass and concrete city bandwidth access. The buildings have changed. Again, it just kind of like is changing the color palette mainly. Let's go to 1985. This should be a drastic visual difference. Not much. The buildings got a bit shorter. It looks like everyone's wearing like sunglasses or I don't know what those would be specifically. More purplish. Definitely the effect of the 80s is seen outside the main city center here. The cars are moving around. Got like a bit of neon and things. Let's see how 65 is. Okay, much simpler. I think it attempted like pastel colors and things of the sort, but they're so washed out it's it's a bit difficult to tell. But again, we have to keep in mind and something I have to continuously remind myself is this was done in C++. So this is not just like the 3JS version of this which would produce better and more aesthetically pleasing results. It's interesting that the tips of all the buildings have these weird plates or circles or whatever you'd like to call those. Then finally, we have 1945. Okay, war bonds. That's what that sign says right there. So, the signs are a little harder to make sense of. These are some very modern street lights for 1945, I would assume. Not having personally been there to check, but I don't think I don't know. And the folks though, like they have the hats on perhaps in ode to more formal dressear, I think. So, okay. This was interesting. I'm going to assume the same horrible sound is playing. It is. So, I'm just going to turn the speaker off. But interesting to see how these do when given the same task. but in a completely different and more difficult language. So interesting. All right, for the final test, we're going to be giving it the Street Eat game. However, this is a bit different. So, this is almost going to be morphed with that very intricate Steve the PC repairman cinematic game prompt. This is Street Eat where you walk around and punch things. They go hilariously long distance. However, I've given it an OpenAI API key to generate some speech assets. And I've also told it there need to be some cutscenes in the game. So, here's the prompt for this game. at least a portion I can show before the audio and API keys. There make a low poly 3D game called Street Eat built in 3JS thirdp person camera and behind the main character. The setup Gary is a normal guy having a normal day. His right arm is not normal. It's the size of a chest freezer. Nobody in the game ever explains this. It is never mentioned. The game opens with a short cut scene at a pizza truck parked on a city block. Gary is second in line, one slice left in the tray. Pepperoni his favorite. A guy in sunglasses walks up, steps in front of him, and takes it. Cutscene dialogue. This is the only dialogue in the game. Gary, that was the last one. Guy in sunglasses, was it? Pizza guy, that was the last one. Gary, long pause, flex his right hand. Okay. And then basically, street eat begins. And you just go around and eat everything on the city block. So, we're having it make this. And it will generate these dialogue bits itself just using the API key I've given it. First, let's fix that. That was the last. >> Was it? >> That was the last one. >> Okay. >> All right. This is actually kind of good. The biggest issue I'm noticing is for some reason it doesn't properly give time for the audio clips to stop. I'm also noticing a sad lack of movement basically of like anything else in this block. Even though, to be honest, it's not badly drawn, it's definitely a very cartoon-like sound. Everything is just static, which is really kind of frustrating. So, and then we get like some errors where they just keep like flailing around there. You know what? I think this is deserving of a second try or an improvement round. So, I've given it a bit of feedback here. basically just asking it to improve everything relating to this game. All right, in a little over 4 minutes, we got our hypothetically improved result. So, let's see if it starts off by one, making the intro scene look better, and two, the dialogue should have enough time to finish without more dialogue starting. >> That was the last one. >> Okay. >> Was it? >> That was the last one. >> Okay. Uh >> oh, he's raised his fist. Okay, something's not quite right here, but the movement is working. So, let's just um not bad. Oh, is hyper yeet. Did it make it so they bounce off things in the environment? Let's do a small yeet. I do believe so. All right, let's Yeah, the hyper yeet didn't happen there because it wasn't bouncing off things. All right. Why does this thing consistently make vehicles way too small? Okay, the sound effects are weird. And we can't unfortunately move the camera. So, let's just try the pigeon. Yep. All right. The dog is wagging its tail, which is just not right. All right. Street. Um, not street ye the light pole street lights. Can we eat buildings? I don't know because there's a collision detection there. Okay. Overall, it definitely improved it. I actually have to say the way it's drawn the characters in this interesting low poly style is actually kind of neat. And it's also neat how everything is like kind of bounces off the environment. We'll do one more hyper yeet. Okay. Interesting. Oh, there's a bus. All right. We'll do That's so weird. I did not ask it to make everything so bouncy. And why are these cars so tiny? All right. Overall, let's take a look at our pricing. All right. Overall, so the entirety of our testing today, including a few speech bits like getting generated here as well, we spent $1.34 on the entirety of everything we tested here today, which really is not bad at all. I mean, this model is incredibly cheap. And I think that's the main takeaway here. If I really had to think very hard to find like an analogy for what I feel about this model, it's like a fast food place where it's cheap and it's fast. It might not be the best thing out there, but for the convenience and the low cost and the speed. I suppose it's definitely a prudent model to use for maybe cheaper, more automated, repetitive things. It does have chopsaw. I mean, look at this right here. This 3JS game scene is something that 6 months ago, maybe 8 months ago, would have been like, "Oh my, this is like state-of-the-art top tier model. This is incredible." So, the improvements in both the cheap models and the big models have been pretty impressive. And I wanted to specifically test this, especially after they dropped the pricing by 80%. Because I know a lot of folks this may have now been placed on their radar with that massive reduction in price. So I wanted to see what it was really capable of. Keep in mind, we ran everything on max effort. So the highest possible reasoning effort for this model, which will inevitably have made the results more expensive than they could have been, but I also wanted to see full like potential. The browser OS was just by far the weirdest, most out there one I've ever seen in my life. It just looked like a dashboard. It did not look like a browser OS. I Oh, okay. That was interesting. Personally, didn't much care for it. Our Subway FPS was much better. It was clean. It didn't have a lot of wow factor to it or polish, but it worked properly and it was cheap. The skate game was actually quite good. I was happy with it. The folks had walking animations. The tricks looked good. The only real bug that we had here was when we were able to manage getting onto a rail to grind. It would kind of just get stuck at the end. And that's something that I've seen multiple times before. So, it's not an totally unexpected thing, especially for a cheaper, smaller model like this. Additionally, we did also do the Chrono City test from within C++ as well. Aside from a really like disgusting soundtrack that it decided to include with this, it was interesting. There wasn't a lot of interactivity, meaning like all of the pedestrians were just static in their places, which was a little disappointing. And overall, it wasn't the best, but I also have to remind myself that we're not doing this in 3JS. We're doing this in C++, and they can't use Ray Lib. So, this is definitely a more difficult test. And I did do the same specific one. Maybe the prompt was slightly different with Deepseek V4 Flash Light, whatever the smaller model was that we tested a couple days ago. So, if anyone's interested, a lot of these tests will have crossed over because these two are definitely pretty interesting to compare. The Jerry's Apartment result, which has now been replaced with Street Heat because that's the newest one that's being served on that address. The Jerry's Apartment one was actually pretty good-looking, even if all the pieces weren't properly put together. I just liked the aesthetic of that one. Same with this watch website. Even when I had mistakenly not denoted it needs to be a 3D watch model, this front end here just looks refreshingly like different. It's not the traditional like purple vibe coded look like we had in the browser OS test. So, I was pretty pleased with this overall. Then when we had it turn them into 3D, it really did a quality job, especially considering a model of this price. It's This is one that still trips up a lot of models unless they're fully on their game or benchmarked to like properly do well in this test. Then finally, I don't know what that was. We had Street Heat, which we just saw, so I'm not going to go over it again entirely. The only other thing that I neglected to specifically mention, I believe, would be the 3D model of that engine. And unfortunately, this just was not 100% together. It may very well have fit the N20 motor, and it did have some semblance of like an inline 6, which was cool. here. The biggest problem was it was not realistic to print because I said little to no supports and it needed far more than little to no would be indicative of. So overall, this was not as good-looking as I had expected it to be. However, it may very well be geometrically competent. So that's kind of just maybe a limitation of this in terms of like the intricate detail. And then overall, finally, our entire spend for this again was $1.34. So, I wanted to specifically use this through the API and not through a GPT subscription, just so we could really get a proper breakdown of each task and what the individual cost was. So, that's probably going to wrap up our first look and test of GPT 5.6 Luna on max effort mode. If you have any questions, please feel free to leave them in the comments and thanks for watching.

Frontier News · by Hyperjump Technology