Transcript of Opus 5.5, Sonnet 5.5, Astra 6 и Sol 6.1 - кто лучше делает игры?
Студия Игор
0:00Today we will test the newest and0:02 top-tier models currently available. We0:04 will perform five very different0:06 gamedev tests and finish by building a0:08 game in Unity. It’s going to be0:10 interesting. Hello everyone and welcome0:16 to the Game Studio channel. My name is0:19 Igor, and today we are going to make a0:21 game. Two new models were released just0:23 this week. They are Sonnet 3.5 from0:27 Anthropic and its direct price-point0:29 competitor from OpenAI, GPT-4o. And0:33 today, of course, we are going to test0:35 them. To make things interesting, we0:37 will also test Claude 3 Opus. GPT-4o.0:41 In total, we will be doing five very0:43 different tests today. The neural0:45 networks will be making 3D models in0:47 Blender, generating music and sounds0:50 for the game, animating our characters,0:52 building a level based on a reference,0:55 and finally, each network will assemble0:57 a full game in Unity. Now, a bit about1:00 the game that each neural network will1:02 be creating today. I have prepared this1:04 concept. In the image, you can see that1:07 we control a robot and fight other1:09 robots on a level that looks like an1:12 arena. So, it will be a Battle Royale1:15 featuring combat robots. To make it1:18 even more interesting, we will make the1:20 environment fully destructible, along1:22 with the robots themselves. We will1:25 destroy them using several types of1:27 weapons: lasers and rockets. I have1:30 prepared detailed descriptions of all1:32 the tests that each neural network must1:34 complete, and I’ve created five files1:35 which I placed in the project folder.1:38 You will be able to see the text for1:40 each test on the screen. You can pause1:42 the video if you want to read them in1:44 more detail. The first test is to1:47 create 3D models for all three robot1:50 types: large and heavy, medium, and1:52 small and fast. In addition to this, it1:55 is necessary to immediately account for1:57 full robot destructibility—well, as1:59 much as possible—and prepare the2:01 models for import into Unity. We will2:03 work with Anthropic models in Claude,2:06 and with OpenAI models in ChatGPT. I2:09 will be using the maximum reasoning2:11 level for all models participating in2:13 our test today. Let’s move on to the2:17 results of the first test. And we will2:19 start with the result from Opus 3.5.2:22 Right now on the screen, you can see2:24 the image of the robots. On the left is2:27 the reference, and on the right is the2:29 result that Opus 3.5 created in Blender2:31. Getting ahead of myself, I will say2:34 that this is the best result in this2:36 test among all the models we will be2:38 testing today. It is also worth noting2:41 that Opus didn't stop there; it2:43 additionally created a Unity project2:46 and set up a scene where all three2:48 robots stand in a row so they can be2:51 examined in more detail. This is how2:53 the large robot looks, the one that is2:56 the biggest and heaviest. This is what2:58 the medium robot looks like, and3:00 finally, the smaller robot. Furthermore3:02, a separate mode is available in this3:04 scene where you can shoot at these3:07 robots and watch them get destroyed.3:09 This is also done quite nicely. You can3:12 immediately see that Opus really tried3:14 hard to complete this task. Perhaps3:17 some might think it did too much or3:19 things it wasn't really asked to do.3:23 But if you read the task it was given3:26 in detail, there is a line like this.3:29 You can read other tasks in the folder3:31 for context to understand what we will3:34 do next. So, it turns out that Opus 5.53:36 is simply playing ahead of the curve.3:38 Next, let’s look at the result from3:41 Sonnet 5.5. Now on the screen, you can3:44 also see the image of the robots,3:46 references, and the results that Sonnet3:49 modeled in Blender. It is worth noting3:51 that Sonnet did not prepare the scene3:53 in Unity. Nevertheless, there are3:56 renders like these where all three3:58 robot models stand in a row. And there4:01 are separate renders showing the4:03 destruction of these robots. If you4:06 look closely, the result seems a bit4:09 worse than what Opus did. This is4:12 especially noticeable on the middle4:14 robot model. Its legs look like they4:16 are slightly twisted at the knees.4:18 Although I can say that, overall, the4:21 result is quite similar to what Opus4:23 5.5 did. If we compare prices, Sonnet4:26 5.5 is cheaper than. Opus, but the4:31 quality of work is lower, and the4:34 volume of work performed is also much4:36 less. As soon as we look at all four4:39 models, I will show how much it cost in4:41 terms of AI expenses. And at the very4:44 end of the video, we will see how many4:46 limits were consumed for each model to4:49 complete all these five tests and build4:51 the game. And now let’s take a look4:53 at the result from GPT-6 Astra. Now you4:56 can also see the references on the left4:58 and the modeling results from GPT in5:00 Blender on the right. I can’t say5:03 it’s very bad, but I wouldn't call5:06 this result very good either. Overall,5:09 it seems that other models not only5:11 handle Blender modeling better but also5:14 maintain a consistent style better.5:17 Whatever I model with GPT-6 Astra in5:19 Blender, there’s a certain plasticky5:22 feel. I don’t know how to put it into5:24 words. So, you can see the modeling5:26 results for yourselves. Astra did not5:31 prepare the scene in Unity, nor did it5:34 prepare any renders or scenes where one5:36 could see the destructibility of the5:39 robots themselves. In terms of money,5:41 it turned out a little cheaper than5:43 what Sonnet did. And now let’s see5:46 what GPT-6.1 Soul prepared for us.5:49 According to OpenAI's claims, GPT-6.15:51 Soul is five times cheaper than GPT-65:54 Astra, but in quality, it's almost at5:57 the Astra level. On the whole, one6:00 could probably say that’s true. You6:02 can see the results for yourselves on6:04 the screen right now. And indeed, it6:09 all looks similar to what GPT-6 Astra6:11 made for us. It also has that certain6:14 plasticky feel. The general GPT style6:17 is recognizable, but additionally, GPT6:19 6.1 Sol prepared a web page for us with6:21 a preview of all three robots, where6:23 you can look around and rotate the6:25 result, meaning you can view each robot6:28 from the front, back, and take them6:30 apart piece by piece. Besides this, it6:34 was intended that one could also6:35 bombard these robots here to see their6:37 destructibility. But, unfortunately,6:40 this function did not work properly.6:42 And as for the money, it turned out not6:44 five times cheaper than GPT6 Astra, but6:46 about 15 times cheaper. And right now6:50 on the screen, you can see the6:52 statistics for the first test: how much6:54 time and how many tokens were spent on6:57 this task for each model, and how much6:59 it costs based on AI pricing. And I7:03 will show how many weekly subscription7:05 limits I spent at the end of this video7:07. Another important point worth7:09 mentioning. Everything you will see7:12 today was generated procedurally,7:15 without using any assets, music, sounds7:17, textures, ready-made models, or7:19 anything else downloaded from the7:21 internet. That is, literally everything7:25 the models did, they did using code.7:27 But we continue and move on to the next7:29 test. And the next test consists of7:31 creating sounds and music for the game.7:34 In the task description, I provided a7:36 hint. It is necessary to find7:39 references for game sounds and music,7:41 and specifically what can be used to7:42 generate these sounds. For example,7:46 noise reduction for servos, wave7:48 dispersion for lasers, modal synthesis7:50 for metal parts, and so on. Literally7:54 with a description of mathematical7:56 formulas. The number of sounds is left7:58 to the discretion of the model itself.8:00 But for the music, we need to generate8:02 two tracks. One for the main menu of8:05 the game and one track for combat. All8:07 of this the models will do procedurally8:09 via code. And now let's listen to the8:12 result. And first, let's see what OPUS8:14 5.5 made for us. Like the other models,8:17 it prepared a preview website like this8:19 for us. And here you can listen to the8:22 sounds of footsteps, servos, lasers,8:24 rockets, explosions, and so on, as well8:26 as the music. The most interesting8:29 thing that I liked, which Opus has and8:32 other models don't, is that for each of8:34 these sound categories, it offers a8:36 reference to listen to on the left, and8:38 on the right, it lets you listen to the8:41 result it generated with code. Besides8:44 this, it shows exactly which references8:47 from the internet were used, and8:50 exactly how it reproduced this sound8:52 wave. I want to note once again that no8:56 samples were used here. All the models8:58 made the sounds procedurally via code.9:01 I can't say whether I like or dislike9:03 this result. Some sounds turned out to9:06 be really quite good. For example, the9:09 sounds of explosions and rockets. Well,9:11 probably all the models produced decent9:13 ones. But for Opus, for example, I also9:15 like the hydraulics sounds. That turned9:17 out pretty cool too. But unfortunately,9:20 some sounds are unsuccessful. They9:22 appear in Opus as well as in other9:24 models. But what we can definitely9:26 evaluate is the quality of the composed9:28 music. And now, let's listen to the9:30 music for the main menu. And this is9:45 what the battle music sounds like.9:58 Write in the comments if you liked it10:00 or not, and which model's music you10:01 liked the most. Now let's look at the10:05 result from Sonnet 5.5, or rather,10:07 listen to it. Visually, the webpage is10:10 designed in the same style as the one10:12 made by OPUS 5.5. Don't think the10:15 models were peeking at each other. I10:17 performed these tests sequentially, not10:20 in parallel. And before running each10:22 new test for each new model, I10:24 completely deleted the project folder10:26 from the previous model and cleared the10:28 caches. So, now we can also listen to10:31 the sounds that Sonnet 5.5 prepared for10:33 us. Overall, it looks and sounds very10:43 similar. And probably the only thing we10:46 can really evaluate is the quality of10:48 the composed music. And let's listen to10:51 it. This is how the main menu music10:54 sounds. And I would like to11:08 specifically note the music composed11:09 for the battle. I would probably set11:23 that as my ringtone or alarm clock. Now11:26 let's listen to the result from GPT-611:28 Astra. The sound of a heavy machine.11:34 The model also prepared a preview11:36 webpage for us in its own style. You11:39 can listen to different sounds. There11:41 is a description of how these sounds11:44 were generated, as well as formulas and11:46 references. And now let's listen to the11:49 music. And this is what the main menu11:51 music sounds like. That reminds me of12:04 something very strongly. And this is12:10 what the battle music sounds like. And12:23 let's see what GPT-6.1 Soul did for us.12:27 And Soul also prepared a worthy result12:29 for us that doesn't look like the12:31 results of other models. Hear the12:34 weight of the machine. What I liked12:36 here was this "battle scene" button.12:40 When you click it, you can hear a12:41 composition of sounds and music,12:43 showing roughly how the game will sound12:45. In my opinion, that's very cool. And12:57 further, similar to how GPT-6 Astra did12:59 it, you can listen to the sounds. There13:02 is a description of how these sounds13:04 were created, which formulas were used,13:06 and which synthesis method. And now13:08 let's listen to the music from GPT-6.113:11 Soul. And this is how the main menu13:13 theme sounds. And this is how the13:25 battle theme sounds. I cannot single13:38 out any one winner here. Overall,13:40 everyone did a pretty good job. As I13:41 said, everyone managed to make the13:43 sounds of explosions and rocket fire13:45 quite well. But everything else is a13:47 matter of taste. Write in the comments13:49 which result you liked the most. Right13:53 now on the screen, you can see the13:55 token count, time, and money spent on13:57 this task for each model, calculated13:59 based on the AI cost. We are continuing14:03, so let's move on to the next test. If14:05 you like what I do, you can support the14:08 channel on Boosty. You can subscribe or14:10 make a one-time donation. You can see14:13 the QR codes on the screen. This really14:15 helps the channel grow. The next test14:17 is quite complex and extensive. The14:20 models will need to create a level for14:22 a game in Unity, as close as possible14:24 to the provided reference. All models14:27 for the level will also need to be14:29 created in Blender. To do this, we will14:31 need to bring in sub-agents, as with14:34 all other tasks. You can also see the14:36 full text of the test on the screen14:38 right now. As a result, the neural14:40 networks need to produce a video14:41 demonstrating the level in Unity, which14:43 they will record directly in the engine14:45. And we will watch the first level14:48 video from Opus 3.5. I want to note14:52 that Opus shows not just the level, but14:55 also immediately places mechs in it14:58 that shoot at houses and various15:00 objects, showing some interactivity. So15:04 you can immediately see the level's15:06 destructibility and how it might look15:08 in the game later. Overall, it's done15:11 quite well. I can't say that Opus 3.515:14 made it look exactly like the reference15:17, but the result was truly decent for a15:20 single run. Let's watch a bit of this15:23 video. And now, let's look at the15:34 result from Sonnet 3.5. This is15:37 probably one of those tests where15:39 Sonnet loses to Opus both in the15:41 quality of the final result and in cost15:43. Sonnet tried very hard and did15:46 everything in its own unique style,15:49 although, in my opinion, it turned out15:52 a bit too acidic. The level also has15:55 mechs shooting, and Sonnet invested15:57 heavily in special effects. When the16:01 robot fires missiles, you can barely16:03 see anything through the clouds of16:05 smoke, although the effects themselves16:07 don't look all that impressive. Sonnet16:10 spent 270 dollars on this task in terms16:13 of API costs, compared to 190 for Opus.16:16 And now, let's watch a little bit of16:18 the video result from Sonnet. And now16:32 let's take a look at what GPT-4o has16:34 prepared for us. Here is the video of16:38 our level result, how it will look in16:41 Unity according to GPT-4o. And the16:54 result turned out so bad, in my opinion16:56, that I asked it to improve it. I17:00 literally asked the model to work for a17:02 while longer and do better, but the17:04 second iteration wasn't much better.17:08 And you can see the final result of the17:10 level on the screen right now, how it17:12 will look in the game. The model spent17:22 about 110 dollars on all of this. In my17:25 opinion, something like Claude or Qwen17:27 would have handled it better. But we17:28 aren't giving up, so let's take a look17:31 at the results from GPT 6.1 Sol. And I17:43 have to say, this result is truly on17:46 par with GPT6 Astra and maybe even17:48 slightly better than what GPT6 Astra17:51 produced. And, by the way, it's not17:54 five times as expensive. Again, GPT617:56 Astra did this level for 110 dollars,17:59 while GPT 6.1 Sol completed the same18:02 task for 2.5 dollars. I don't even know18:06 how. I just can't fathom how GPT 6.118:10 Sol manages to operate so cheaply. You18:13 can see the results on the screen, but18:16 it definitely didn't turn out much18:18 worse than GPT6 Astra. By the way, on18:21 this task, I stopped tracking how many18:23 of my weekly limits I was spending when18:26 working with GPT 6.1 Sol. And I think18:30 for completing all the tasks and18:32 generating the entire game from scratch18:34, I used, well, maybe one, at most 2%of18:37 my weekly limits. Most likely one,18:41 because after this task I turned on the18:43 high-speed mode, that "fast mode" in18:45 the IDE, which is one and a half times18:47 faster and uses significantly more18:49 limits, but I still didn't see any18:52 noticeable consumption. And now on the18:55 screen, you can see detailed statistics18:57 for the third test. How much time was18:59 spent, how much money was spent19:01 calculated by API costs, and how many19:04 tokens all the neural networks used to19:06 complete this task. And we are moving19:09 on to the next test. And now for the19:11 next test. Now our neural networks will19:13 create animations for our robots. And19:16 we will use an approach called inverse19:20 kinematics. The essence of this19:23 approach is that, first, the animation19:25 is calculated procedurally in the code,19:27 and second, it is calculated from the19:29 foot, not from the hip. That is, the19:32 robot first places its foot on19:34 something, on the ground or maybe on an19:37 obstacle, and then, from that point and19:40 position, the angle of all other joints19:42 is calculated. And this way, the foot19:45 lands precisely on a piece of debris or19:48 perhaps a slope, and the knee bends19:50 exactly as it should. We will also make19:53 this a small part of the gameplay,19:55 where you can destroy objects by19:58 stepping on them or crushing cars or20:00 something else. You can see the task on20:03 the screen, as usual. The essence is20:05 simple. Create inverse kinematics for20:08 me in Unity. Also, implement WASD20:10 movement and have the body rotate to20:12 follow the mouse cursor. And at the end20:15, I want to be able to walk around as20:18 the robot in the Unity scene. First, of20:20 course, we look at the result from20:21 Claude Opus 5. I launch the project in20:26 Unity and start our game. And I20:29 immediately see that everything is done20:32 more than well. So, the robot walks,20:36 moving its legs properly depending on20:39 whether it’s moving fast or slow,20:42 starting to walk, or perhaps stopping;20:44 it has some kind of acceleration or20:47 deceleration as it moves. That also20:50 looks cool. I also want to note that20:53 OPC is the only neural network that did20:56 the controls properly. I mean, the21:00 camera and the robot's body really turn21:03 independently of how I’m moving,21:05 using the mouse cursor. And that is21:09 very convenient. There will be various21:11 other options. Today, of course,21:13 we’ll take a look at all of them.21:15 I’ll talk about that later. Besides21:17 movement, in principle, one could21:19 probably say that OPUS 5.5, as usual,21:22 jumped ahead and did almost everything.21:26 You can walk, you can shoot, you can21:28 use a laser, you can fire rockets at21:30 houses. The houses are destroyed.21:33 Overall, it all looks almost like a21:35 game already. I even got a bit hooked21:38 on this whole thing for a while. But21:43 let's not get distracted; let's look at21:45 how the inverse kinematics were21:47 implemented. And yes, indeed, the21:53 robot's leg steps onto the cars. The21:57 cars get crushed, and the legs also21:59 handle various types of obstacles and22:02 slopes properly. And in general, it all22:05 looks very, very good. And today, this22:08 is probably the best result we will see22:12. And next in line is Sonet 5. I’m22:16 launching the scene in Unity. It all22:18 looks something like this. There is22:20 already a menu where you can choose one22:23 of the scenes. We choose the city to22:25 walk around with the mech. And I want22:28 to note the controls right away. Unlike22:31 Opus, all the other neural networks22:34 added camera rotation on the Q and E22:37 keys, even though it wasn't in the22:39 original prompt regarding how to handle22:42 controls. Literally, in the task, I22:46 wrote: "I want to move with WASD and22:49 rotate the camera with the mouse."22:52 That's it. Nevertheless, Sonet, as well22:55 as Astra and Sol, for some reason,22:58 added camera rotation to Q and E, with23:01 aiming on the mouse, and that is not23:03 super convenient. But let's look at the23:07 result. There are both pros and cons23:13 here. Among the cons, I want to note23:15 something that has been trailing along23:17 since the very first task. These are23:20 the crooked legs on the robots. They23:22 look as if they are pointing in23:23 different directions. And the leg23:26 movement, the animation itself,23:27 wasn’t done, well, not exactly very23:30 well. They move somehow, well, not as23:34 naturally, let’s say, as Opus did,23:36 but as if they are sticking a bit to23:38 the ground. Sonet also allows you to23:42 shoot. Use a laser and rockets to23:45 destroy houses and trample surrounding23:48 obstacles. And it all looks like some23:51 kind of bacchanalia of explosions,23:54 smoke, splashes, fountains, and dust.23:58 And in general, it is very difficult to23:59 make out what is happening on the24:01 screen. But we are interested in24:03 inverse kinematics. Of course, Sanet24:05 can also be said to have coped with24:08 this task. The legs step onto24:11 elevations and various obstacles. It's24:15 okay. This result can be counted as24:19 successful, but I can't say that I24:21 really like how it was all implemented.24:24 And now let's see what GPT6 Astra has24:27 prepared for us. Well, Astra prepared a24:31 scene in Unity for us where you can24:33 walk with a mech. I asked to fix these24:36 awkward controls, remove the camera24:38 rotation on the Q and E keys, and use24:41 the mouse instead. And it probably only24:45 got worse, because the crosshair was24:47 moved somewhere behind the mech's back.24:50 But overall, it's fine; I didn't ask to24:53 change anything else in the end and24:55 left it as is. We can evaluate how the24:58 walking works, how the animation works,25:00 and how the kinematics work. You can25:06 step on cars, and you can also step on25:08 buses like this. Nothing happens to the25:11 buses, of course. And similarly, at the25:14 end, I stepped on a fountain and ended25:16 up getting stuck. But overall, the25:19 kinematics were implemented. We25:21 consider that Astra successfully25:23 handled the task. Nevertheless, not25:25 without bugs. That is, there are indeed25:28 moments that literally ruin the entire25:32 gameplay. You can simply get stuck in a25:34 fountain. And let's look now at what25:37 GPT 6.1 Soul did for us. GPT 6.1 Soul25:41 gave us a result five times cheaper25:44 than GPT6 Astra. And, in my opinion,25:48 even better. You can see how the mech25:51 moves its legs. It can actually step on25:55 cars, although, unfortunately, not all25:57 leg positions are processed correctly.26:01 That is, sometimes when stepping on a26:03 car, the mech flies up into the air.26:05 This, of course, should not be26:06 happening. That is, one leg should26:09 remain on the ground, and the other leg26:11 should remain on the car. Nevertheless,26:15 GPT 6.1 Soul had fewer bugs, and I26:18 never managed to get stuck in the26:20 fountain. And now on the screen, you26:23 can see detailed statistics for the26:25 fourth test. How many tokens were spent26:28, how much money, and how much time.26:30 But we continue. My channel recently26:33 started a Discord community. You can26:35 see the QR code on the screen. The link26:37 will also be in the description. There,26:39 you can discuss neural networks, game26:41 coding, and various tools and26:43 approaches. Join us, it's interesting26:46 there. And we are moving on to the next26:48 test. And the next test is the final26:51 one. Now each neural network will26:53 assemble a full-fledged game in Unity26:56 using everything it did in the previous26:58 tests. This will be a mech arena in27:01 Battle Royale mode. If you want to read27:04 the full text of this test, it is27:06 available on the screen right now. And27:09 we will begin testing in reverse order27:11 and see what GPT 6.1 S made for us. I27:15 am launching the game in Unity. And we27:18 are greeted by this starting menu. You27:20 can choose a mech here. And I will27:22 start playing with the heaviest one,27:24 which is called Bastion. Next, we have27:27 a 3-second countdown. And after that,27:29 the battle begins. As I already said,27:41 aiming is quite difficult. I lose27:44 almost immediately. I tried all sorts27:46 of different tactics. The most27:49 effective one is simply to stand aside27:50 and wait for all the robots to shoot27:52 each other. And that way, you can take27:56 second place. I also tried rushing into27:59 battle, taking third or fourth place,28:05 or looking for obstacles to hide behind28:07 buildings. And, perhaps, the best28:25 result can be seen in my last run, when28:28 I switched to a medium mech, which is a28:31 bit faster, although it has less health28:34 and less armor, but I managed to fully28:44 participate in this battle and even win28:47. So that is the kind of game GPT 6.128:50 Soul produced. It is very far from what28:54 was in the reference, but it is very28:57 cheap. I mean, even including the cost29:02 of the AI, the price of this, let's say29:06, prototype is very small. And now29:09 let's see what GPT6 Astra made for us.29:13 Next, we test the result from GPT629:16 Astra. Similarly, I start the game in29:19 Unity. And in the starting menu, you29:21 can also choose one of three mechs. I29:23 choose the heaviest one, and the battle29:26 begins. What can I say about the game29:29 from GPT6 Astra? No matter how many29:41 times I tried to play this game with29:44 various mechs and tactics, I always29:47 died first. Well, perhaps with rare29:50 exceptions. It seems to me that the29:52 Battle Royale wasn't really implemented29:55, because as soon as I appeared on the29:57 level, it didn't matter if I was in the29:59 field of view of other mechs or not,30:01 they all started attacking me. In the30:16 end, only at the very end, I think with30:18 the medium mech, I managed to take30:20 second place. And that is probably only30:31 because this tactic, where you hide to30:34 the side while they fight each other,30:36 is the most effective one. But as soon30:39 as you enter their line of sight, the30:41 enemies always switch to you. And let's30:45 not forget that, on top of everything30:47 else, you can get stuck in a fountain.30:54 So that is the result, that is the game30:57 that GPT6 Astra produced. And next up31:01 is Sonet 5.5. I start the game in Unity31:04, and we are greeted by this31:06 neon-colored start screen. Here you can31:09 also choose a robot. I choose the31:11 heaviest one. And we start the battle31:14 from the roof of a building. Similarly,31:17 there is a countdown timer before the31:19 fight begins, and then some kind of31:22 chaos starts. Rocket fire, smoke,31:25 lasers, splashes, dirt, dust—you31:28 can't see anything at all. I tried31:31 playing this for a while. You can see31:33 it for yourselves on the screen. It’s31:35 unclear what is happening. Plus, the31:37 controls are really clunky. In general,31:39 in my opinion, SNEET went way too far.31:42 It also spent way more on limits than31:45 OPUS did. But more on that a little31:47 later. For now, I suggest you just take31:50 a look at the result for a bit. And32:07 then we will move on to what OPUS 5.532:10 has done. To stay updated with the32:13 latest news and announcements,32:15 subscribe to my Telegram channel. You32:17 can see the QR code on the screen, and32:19 the link will be in the description.32:20 Well, let’s continue. So, let’s32:23 look at the result from OPUS 5.5. And32:27 here, in my opinion, it turned out,32:29 well, really quite good. Starting from32:32 the main menu, we can change a large32:35 number of settings and individually32:37 select a robot in the hangar to play32:40 with. By the way, the previews during32:42 the robot selection are also made32:44 nicely. The camera shifts to the32:46 selected robot, and they fire into the32:49 air. Then we start the battle. And the32:52 first thing we need to do is choose a32:54 landing point. We tap a random spot on32:57 the map, and our robot arrives in the33:00 city from above. And that already looks33:03 quite spectacular. The graphics are33:06 implemented at a very decent level.33:09 This result reminds me more of the33:12 original concept than any of the others33:15. And I remind you that we did this,33:17 well, literally in one shot. If we give33:20 this process a little more control, we33:24 can get an even better result. Moving33:27 on to the effects and controls,33:29 everything is done much better than33:31 with any other participant. There is a33:34 moderate amount of effects, and the33:36 controls are convenient. It was33:38 actually fun to play. Just like in the33:40 previous test, I got a bit hooked while33:42 testing the movement. And here, too, I33:45 spent several gaming sessions. And in33:48 the end, I even managed to win. And it33:51 was a fair fight, not just standing33:54 aside behind an obstacle and shooting33:58 back. Every time, I had to think34:11 through the landing spot to roughly34:13 plan where the other robots would land.34:17 And then, before the zone closes in,34:19 you need to manage to defeat all the34:21 other enemies on the level. Besides34:24 that, the enemies fight each other, and34:27 it's not like what happened, for34:29 example, with GPT6 Astra. As soon as I34:32 get into the enemies 'line of sight,34:34 they don't just start attacking me. No,34:38 there really is a battle going on here,34:40 and there's a real sense of, well, some34:42 kind of living process to a certain34:45 extent, as if you're playing with real34:47 players. Once again, I want to note34:52 that all of this looks very cool. I34:54 really like the effects. The graphics34:57 turned out fun and are quite34:59 well-developed. And I suggest we just35:02 take a look at the result that Opus 3.535:05 created. And now on the screen, you can35:35 see detailed statistics on how much35:37 time, money, and tokens were spent on35:39 this final fifth task. And now you can35:43 see the statistics on the total time,35:45 money, and tokens spent on the entire35:47 game by all participants. You can pause35:52 the video to take a closer look. But I35:56 didn't spend all this money via API on35:58 creating these four games. I have two36:01 subscriptions for 200 dollars each. One36:04 for Claude, the other for GPT. And how36:08 many of my weekly limits did I use up?36:10 Claude Opus 3. Sonnet 3.5 used 19%of36:18 the weekly limit. GPT-4o Astra used 13%36:23of the weekly limit, and GPT-4o mini36:27 used 1%. Well, maybe a maximum of two.36:30 At some point, I stopped counting36:32 because I switched to Fast Mode. And36:34 all this on a 200-dollar subscription.36:38 If you want to see the numbers36:39 converted for other types of36:41 subscriptions, they will also be shown36:43 on the screen now. You can pause and36:46 see what this looks like in limits for36:48 the 100-dollar and 20-dollar plans.36:51 There is one important nuance.36:55 Anthropic does not disclose how many36:57 weekly limits they actually provide for36:59 200-dollar subscriptions relative to37:02 other plans. But there is official37:06 documentation they published a year ago37:08, where they wrote that the limits on37:11 the 200-dollar subscription are 1.737:13 times higher than on the 100-dollar37:15 subscription. Then they stopped37:18 publishing that documentation.37:20 Independent measurements today confirm37:22 this information, that the limits have37:24 remained at approximately the same37:26 level. What conclusions can be drawn37:28 from this experiment? If you just want37:31 to one-shot a game and make it look as37:34 close as possible to the references,37:36 documentation, and description you37:39 provided to the AI, use Opus 3.5. If37:44 you consider yourself an AI tuning37:47 expert and can find a Sonnet37:49 configuration where, for example, it37:52 works faster and more efficiently on "37:54high" than Opus on "medium," you can do37:57 that. But the result will not always be38:01 predictable. If you are using GPT-4o38:04 Astra, keep using it. It is a good38:07 model. It handles tasks quite decently.38:13 Some of you might notice, well, like,38:15 it worked faster or it spent less money38:18 via API. If it had spent the same38:21 amount of money via API—well, in38:23 terms of API costs, I mean—then the38:26 result would 100%be on par with Opus or38:28 Sonnet. I have to disagree with you38:31 there, because you saw the weekly limit38:34 usage yourself. If we make Astra work38:38 four times as much, or two or three38:40 times as much, the amount of the weekly38:43 limit it uses up on the $ 20038:46 subscription will also be higher. Than38:50 the limits consumed by the ORC.38:52 Therefore, in my view, it is not very38:54 cost-effective. And now, GPT 6.1 Sol.38:58 What can be said about it? Overall, it39:00 can be used for everything. For39:02 everything when you need to get results39:04 quickly and cheaply. It is an excellent39:07 model, and as OpenAI calls it, a39:10 workhorse. The results were indeed39:13 comparable to the results shown by GPT39:16 6 Astra. And perhaps Sol has gotten39:19 smarter, or maybe Astra has gotten39:20 dumber. But today, I would rather use39:23 Sol on a regular basis. And that is all39:26 for today. Leave comments on which game39:29 you liked most and whether you agree39:32 with my conclusions. Hit the like39:34 button. That way, I know you enjoy this39:36 type of content. Subscribe to the39:38 channel; it really helps it grow. And39:41 goodbye, everyone.
5,544 words · 801 lines
