I built the same game with Astra and Fable 5.1... only one was fun Transcript, AI Summary & Key Points

00:00 Yesterday, Nvidia CEO Jensen Huang logged into X to make an announcement. AGI has arrived. Congratulations, OpenAI team. It was a sobering and totally independent assessment from a man who in the very same post pointed out how OpenAI's new Astra model was trained on more than 100,000 of his Grace Blackwell GPUs, and that another 400,000 of his GPUs are coming online soon.

00:20 Obviously, he has nothing to gain here, but is Jensen right, or is he wrong about Astra being AGI? Because if he's not wrong, that means every human being working in front of a computer right now is officially obsolete. And that's the fifth time that's happened this year. Well, in today's video, we'll learn the truth because yesterday Astra rolled out to the common man, or at least the man willing to pay $100 for a GPT Pro membership.

00:42 I compared Astra to Fable 5.1, and the results were not what I expected. It is September 9th, 2026, and you're watching The Code Report. There was one group of developers who were supposed to be safe from AI, game developers. Because for a game to be fun to play, it needs to have a very specific feel and soul. Not to mention, the ability to draw and design good-looking 3D graphics.

01:02 Unfortunately, though, it looks like the game developer and 3D artist jobs are about to become just as rare as the Angular JS developer jobs. Because GPT-6 Astra has been called a turbo AI machine god for 3D games. Not only can it slop out amazing games in 3JS, but it can also build extremely detailed graphics in Blender and work with other game engines like Unity and Unreal.

01:24 But what might be even more terrifying is that Astra was the first model to reach the top of the Arc benchmark, which was designed specifically as a benchmark that can't be memorized, which forces a model to demonstrate how well it generalizes intelligence. It completely mogs Anthropic with Opus sitting at 30%, but what they don't tell you is that a team at Berkeley already scored 99% on the Arc AGI benchmark several months ago.

01:48 They did it with Opus 4.8 and Fable 5, but the real difference maker was the harness. But today we're not going to focus on the trust me bro benchmarks, and instead, we'll handcraft some of our own slob to compare Fable 5.1 and GPT-6 Astra side by side. So, to do that in the most organic way possible, I asked my kids, "What kind of game do you want to play today?"

02:07 And they requested a rocket launch simulator where you customize a rocket and launch it into orbit. I sent both models the exact same prompt at the exact same time, and ChatGPT finished first in about 26 minutes. When I first launched the ChatGPT game, I was incredibly impressed. The 3D graphics are highly detailed and beautiful with very few mistakes, and the UI looks beautiful as well.

02:27 One thing that was kind of interesting, though, is that the UI looks almost identical to a lot of other games that I've seen on Twitter. There's a very obvious formula being used here that makes it easy to spot ChatGPT generated games. And now, as far as the actual gameplay goes, I was able to launch a rocket into orbit, but I wouldn't exactly describe this game as fun to play.

02:44 I think the most impressive thing was this graphic of the rocket actually orbiting the globe. Now, after playing that for a while, I went back to check on my Fable prompt, which still wasn't done, but eventually finish at some point in the far future. And when it did, the graphics were embarrassing by comparison. The UI looks like it was built with Twitter Bootstrap, which is a library that was written before most of you were born, and the 3D graphics aren't nearly as detailed as what Astra generated.

03:07 However, after actually playing the game for a few minutes, I quickly realized that the gameplay was far superior to what Astra had generated. For one, the options for rocket customization are way more complex. Claude provided way more knobs to turn and made scientific calculations that make this feel more like a legit rocket simulator. But then, when it came to gameplay, the game was also just more fun with a bunch of different possibilities for success or failure and cooler animations when your rocket explodes.

03:33 But the big test was which game do my kids prefer? Well, this actually ended in a split decision with the 3-year-old preferring the more simplistic Astra game with the 8-year-old preferring the more complex Fable game. It's mostly subjective, but one thing's for sure, Astra's ability to generate 3D graphics has improved dramatically. It just a couple months ago back in July, I had GPT Soul generate an exploded view of a mechanical watch, and the end result was a pile of useless garbage.

03:58 But when I re-ran the prompt again yesterday, the end result was this beautiful diagram that likely would have taken a 3D designer hundreds of hours to produce. And this actually terrifies me because it's going to lead to a new wave of slop on YouTube of people copying channels like Jared Owen and other people who make 3D explainer videos. Videos like this used to be what we would call high effort and might require hundreds of hours of 3D design work in Blender.

04:22 Unfortunately, that moat no longer exists and the rise of 3D slop is going to be the next big thing that makes the internet suck even more. But we still haven't answered the question, is Astra actually AGI? Well, as you know, the only way to find out is to see if it can build Horse Tender without making any mistakes. It took 21 minutes to run the prompt and this was the final result.

04:39 A design that looks surprisingly similar to one made by Fable a few months ago. However, this design is noticeably cleaner and more elegant. And after spending many, many hours going through every single line of code looking for a mistake, I wasn't able to find a single one. I felt like I was Antonio Salieri reading the flawless music of Mozart. >> Here again was the very voice of God.

05:02 >> And then I realized that can only mean one thing. Jensen was not wrong. AGI has been achieved. Oh wait, actually hold on. Wait a minute. This carrot is a few pixels off. The false alarm. This is just another dumb AI model. If you want to build a seriously good UI, you need to use Mobbin, the sponsor of today's video. It lets your coding agent study more than 600,000 real UI screens before it designs anything.

05:22 I've already been using Mobbin for over 5 years because it shows you complete user flows from thousands of applications. And its MCP server now puts that entire library inside Claude, Code, Cursor, Codex, and all the others. So, instead of asking for a settings page and getting the same purple gradient vibe slop, you start by telling your agent to study how the best apps build that complete user flow.

05:44 It'll use Mobbin to gather the references, build a visual report of the patterns they share, then use those findings to recreate it and adapt it to your app. Every result points back to its original screen, so you can check the source before trusting a design hallucination. Try Mobbin today and get a 20% discount at the link below. This has been the Code Report. Thanks for watching and I will see you in the next one.