Transcript of Claude Fable 5.1 - лучшая нейросеть для вайбкодинга игр?
Студия Игор
0:07Today we will compare Claude Fable 5.10:10 and 5.0 and build two games: a 3D0:12 shooter and a strategy game. It will be0:15 interesting. Hello everyone and welcome0:21 to the Game Studio channel. My name is0:23 Igor, and today we will test the new0:25 model from Anthropic, Claude Fable 5.1.0:30 Just a few days ago, Anthropic released0:32 a new version of its most powerful0:34 model, available to all users. Claude0:37 Fable. It is a minor version, 5.1. And0:40 yet, according to benchmarks, it now0:42 outperforms all of the company's other0:44 models. And today, of course, we are0:46 going to test how well this model0:48 builds games. I spent a long time0:51 thinking about which game to make: a 3D0:53 shooter or a strategy game? A 3D0:55 shooter or a strategy game? A 3D0:57 shooter or a strategy game. And I1:00 decided that today we will do both. And1:03 to make it interesting, we will compare1:05 them with Claude Fable 5 face-to-face,1:07 one-on-one. The same prompts and the1:09 same settings. And I have prepared the1:12 following prompts. Today we are going1:14 to make an FPS: "Oil Platform." It is a1:17 3D shooter where we are attacked by1:19 waves of zombies, and we must survive1:21 several waves. The key here is the1:24 atmosphere. Imagine a rusty oil1:27 platform. At night, in the middle of a1:30 dark sea: a storm, lightning, rain, and1:32 the screech of metal. It should be1:34 interesting. Right away, in all the1:36 prompts, I outlined the main rules. No1:39 using skills, no using memory, and no1:41 looking into neighboring folders. This1:45 is important, and we must have a fair1:47 test today. And to be even fairer, I1:50 will start both sessions simultaneously1:53. On the left, I will have Fable 5 open1:56. Here I am selecting the model. And on1:59 the right, Fable 5.1 will be open. The2:03 reasoning level in both sessions will2:05 be set to max. We will build both games2:08 in planning mode. First, the models2:10 will write a detailed plan, and then2:12 they will implement it. It takes Fable2:14 5 about 2 minutes to write a plan,2:17 while it takes Fable 5.1 about 4. And2:21 it was like that for both games. The2:23 plans themselves are very similar to2:25 each other, but the plan from the Fable2:28 5.1 model is noticeably more detailed.2:32 And, by the way, the Fable 5.1 model,2:34 unlike Fable 5, even went online to2:36 check for available post-processing2:38 add-ons to make the graphics even2:40 better. That was not forbidden in my2:44 original prompt, and both games took2:46 about the same amount of time. For2:50 Fable 5.0, the game implementation took2:52 about 40 minutes for both, while for2:55 Fable 5.1, it took about an hour for2:57 the first and second game. And, of3:00 course, we will be testing blindly3:02 today. First, I will play both games in3:05 turn, and then I will try to guess3:07 which model made which game. You can3:11 try to guess along with me; just pause3:13 the video and write your choices in the3:15 comments. So, you can already see some3:18 data here. In total, we have four3:20 playable builds that we will test today3:23. And about 18,000 lines of code were3:25 written by both versions of the model.3:29 If you calculate the API costs, about3:31 90 dollars were spent on four games. I3:35 remind you that I don't use pay-per-use3:37, I use subscriptions. Round one. FPS3:41 atmospheric horror. Night on the oil3:43 rig. And let's play version A. For the3:46 first game, I asked both models to add3:49 a cutscene at the very beginning. And3:52 now we will look at the cutscene of the3:54 first game. Well, the sea, actually,3:56 looks really cool. And I also like how3:58 the lightning strikes in the background4:00, illuminating the space around and4:02 creating an atmosphere. We are starting4:04. And, by the way, it's quite dark. We4:06 don't have a flashlight. And as for4:07 lighting. Only some shadows on the4:09 platform itself and periodically4:11 appearing lightning. Enemies, by the4:13 way, are visible from afar. Their eyes4:15 glow red. The first wave is quite easy4:18 to pass. There are literally four or4:19 five zombies that don't even really4:21 attack. By the way, I checked. The4:24 barrels, unfortunately, do not explode,4:26 although that was in the initial prompt4:28. Everything looks pretty good,4:30 although the floor reflection is4:32 perhaps a bit, let's say, too smooth.4:35 You don't feel the roughness. of the4:37 rusty surface. Meanwhile, I'm already4:40 fighting off the second wave, and we4:42 have new zombies appearing that crawl4:44 on the floor. They look quite funny, by4:46 the way, but when they run up close,4:48 it's quite hard to shoot them.4:50 Meanwhile, the fourth wave is starting.4:52 Zombies are climbing from all sides.4:54 And overall, we ended up with this kind4:56 of game and this kind of gameplay. I4:59 never managed to pass the fourth wave,5:00 by the way, because of these zombies5:02 that crawl on the floor. It's5:04 practically impossible to shoot them. I5:06 don't know where their colliders are or5:08 exactly where you need to shoot. So the5:10 signal is lost, and let's look at the5:12 second game. And the second game also5:15 starts with a cutscene. And the sea5:17 looks significantly better. And there5:20 is even a camera switch in the cutscene5:22. I can already see that the materials5:24 on the platform itself are also better5:26 crafted. There are some rusty puddles,5:29 you can see the texture of the metal5:31 floor. And overall, it feels and plays5:34 significantly better. You can blow up5:37 barrels. That was in the initial prompt5:39. We didn't have that in the first game5:41. Here, it is implemented. It looks5:43 quite spectacular, by the way. I can5:44 say right away that the gameplay turned5:46 out to be more dynamic. And, in general5:49, the level itself looks much more5:51 atmospheric. By the way, we also have a5:53 flashlight, so it's a bit easier to see5:55 in the dark when we run into some dark5:57 areas of the map. Meanwhile, I am6:00 already fending off the second wave,6:02 and with each wave, the number of6:03 monsters keeps growing. By the way,6:05 they even behave differently. These6:08 zombies that are larger, they start out6:10 walking slowly, and then they seem to6:12 start running, and the whole thing6:14 looks dangerous. There are also smaller6:16 zombies; they don't crawl on the floor,6:18 but they run very quickly. And at6:21 around the third or fourth wave, big6:23 zombies appear, and it's better not to6:25 get too close to them. Well, and in6:28 some way like this, you can even beat6:30 this game. But, in my opinion, the6:32 number of waves here is unlimited. I6:35 mean, I tried to get through four, five6:36, and six waves, and in the end, the6:38 game just didn't end. And that is the6:40 kind of game we got from the second6:42 model. Well then, let's reveal the6:44 results. I remind you that you can6:47 pause and write your guess in the6:49 comments about which model made option6:51 A and which model made option B. But6:54 actually, I think it's all obvious here6:56. And option A is Claude 3.5, and7:00 option B is Claude 3.5 Sonnet. We also7:04 have information available on exactly7:06 what was done in the game, how many7:08 lines of code were written, the number7:10 of modules, how many kilobytes the7:11 final bundle weighs, and so on and so7:13 forth. In addition, detailed costs for7:17 input tokens, output tokens, cache7:19 reads, and cache writes are presented7:21 here, as well as their quantities and7:23 prices. And at the very end, there is7:26 the prompt with which all of this was7:28 implemented. You can either pause if7:30 you are interested in reading it, or7:32 follow the link in the description and7:34 try playing the games. But we continue,7:38 and let's see how the models handled7:40 the second game. By the way, I have a7:43 channel on Boosty. If you like what I7:46 do, you can subscribe or donate once to7:49 support the channel. And we continue.7:52 The second game is a Viking strategy7:55 game in a 2.5 D style. The entire7:58 environment will be 3D, and the units8:00 themselves will be flat sprites. And8:02 the models will have to implement both8:04 of these procedurally. The main thing8:06 in this game is the visuals and the fun8:08, not seriousness. Actually, just from8:11 the screenshots, I can guess which8:13 version of the model made which of8:16 these variants. And we will probably8:19 start our testing with option B. It is8:21 already clear that the ships are8:22 sailing in from somewhere off the map.8:25 And in general, they look quite simple.8:28 These are the kinds of Vikings that8:29 tumble out onto the shore from the8:31 ships. The graphics look quite8:33 primitive, both the 2D art, I mean the8:35 characters themselves, and the 3D8:37 environment. I would actually even say8:40 that this result was made by some GPT8:42 5.6 Sol, but nonetheless, it was8:44 created by Clot Fable 5 or 5.1. Let's8:48 see. We select all the Vikings and8:50 simply send them into the village. Most8:52 likely, there is nothing special to do8:55 here. Next, the Vikings just tear down8:57 the walls. By the way, I didn't8:59 understand at first how exactly this9:00 needed to be done. Then I seemed to9:02 figure it out. You need to click on the9:04 tower, and then the Vikings will attack9:06 the structure. It was already getting9:07 dark. And finally, the gates fell. We9:09 were able to get inside the village in9:12 order to, well, of course, loot it.9:14 Then the units just stand there like9:16 this, exchanging arrows and slowly,9:18 slowly attacking the houses. Actually,9:21 I didn't really like this result. It9:24 doesn't play very dynamically, not very9:26 interestingly. The graphics are quite9:28 simple, but it is what it is. I remind9:32 you that we have a blind test today, so9:34 we aren't drawing any conclusions yet.9:36 And let's move on to testing the next9:39 game. The next game is option A. Let's9:42 watch. And it is immediately obvious9:44 that the graphics are much better. The9:46 environment is much better, the water9:48 is much better. The ships are more9:50 detailed, the units land on the shore,9:53 the units themselves are somehow cooler9:55. And overall, the village also looks9:57 very good, as does the rest of the9:59 environment. We select the squad, burst10:02 into the village, attack the units, and10:04 attack the houses. To win, you need to10:07 destroy all enemy units in the village.10:10 An interesting thing, by the way, when10:11 units attack, they shout something10:13 while doing so. If you pause, you can10:16 take a closer look and read something10:18 like "For Odin, Valhalla,""We won't10:20 give up the chickens," the defenders10:23 shout, "For the village," and some10:25 other phrases. Overall, it's done quite10:28 funnily. I didn't see any chickens, by10:30 the way, but pigs are running around in10:32 a pen inside the village. And you can10:34 even see a couple outside of it. I will10:37 say again that I really liked how the10:39 environment is implemented. And in10:42 general, as a playable prototype, this10:44 is a very, very decent result. And in10:47 my opinion, it is done much better than10:50 the previous version of this same game10:52 that we just watched. So, our village10:55 is looted, and let's move on to the10:57 results. You can pause and write your11:01 guesses, but I think that option A is11:04 Clot Fable 5.1, and option B is Clot11:07 Fable 5. Again, if you click the "more11:12 details" button, you can see11:13 information on what was done, how much11:15 code was written, how many tokens and11:17 how much time were spent, and the total11:19 cost in AI expenses. By the way, both11:23 the first and second games cost both11:26 versions of the model roughly the same11:29 amount: 20 or 25 dollars. But, in my11:32 opinion, the Fable 5.1 model produced a11:36 much better result. And as for the11:39 second game, it is a perfectly viable11:42 prototype. You could ask some other,11:45 cheaper model, like Claude Opus 5 or, I11:48 don't know, GPT 5.6 SOL, to port the11:51 source code and graphics to a game11:53 engine and make this one of the11:55 mechanics, where ships land near a11:58 village and you have to attack, loot,12:01 or capture that village, or whatever12:03 else you want to do. And the core12:07 mechanic could be, for example, a12:09 global map like this, where a main12:12 strategy layer similar to Total War or12:14 other such games is available. By the12:17 way, I generated these concept images12:20 in GPT to show you roughly what it12:22 could look like and what our takeaways12:25 are today. And the takeaway is very12:27 simple. Despite the upgrade being minor12:31, version 5.1 is a much better version12:33 of the model than the previous Claude12:36 Fable 5; it shows a significantly12:38 better result. It spends a little more12:43 time on all of this, but roughly the12:45 same amount of money if you count by12:46 the number of tokens. And this means—12:50and Anthropic themselves actually wrote12:52 about this, too—that this model is a12:55 bit cheaper due to being better12:57 optimized, and I believe reading from12:59 the cache is cheaper there. If you like13:02 this kind of content, hit the like13:04 button and subscribe to the channel. It13:06 really helps it grow. Leave comments,13:09 play games, and make games. And bye13:12 everyone.
2,247 words · 323 lines
