GPT-6 Astra Is Finally Here (And It’s REALLY Good) transcript

We just got another huge model launched today in GPT-6 Astra. And I was going to save this for my news video and just talk about it tomorrow. But if I'm being honest, this launch feels too big to just be like a footnote in an AI news video. When I open up my X feed, it is literally the only thing anybody is talking about on X. It is purely GPT-6 and what early access testers were able to build with it. So, let's talk about what GPT-6 is, do a few tests with it ourself, and take a look at what some other people have managed to build with it.

Now, full disclosure, I did get slightly early access to GPT-6, so I've had a little bit of time to play with it. Also, in another disclosure, OpenAI has sponsored my channel before, but this video is not sponsored. They don't even know I'm making a video about GPT-6, and I can say whatever I want about it because, well, it's not sponsored by OpenAI. So, first thing to note is, as of today, the day I'm recording this, September 3rd, it is not fully publicly accessible yet.

They made the announcement that GPT-6 Astra is rolling out today to a limited set of organizations, and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users. Although the day this video is going live, you may not have access to it yet. Within the next, you know, 2 to 3 days, you probably will have access to it. They have a handful of demos on their blog post here, like this interactive circuit board, this Excel competition, developing like what looks like a SimCity type thing, and a bunch of other demos.

I'm not going to really go through those demos cuz it's easy enough for you to just [laughter] go read the blog post and click through them yourself. But in a minute, I will show you some of the things that I tested, as well as some of the other things I've seen others test. But let's scroll down to the benchmarks really quickly. Although I know benchmarks mean less and less to most people these days, there are some benchmarks that I still think are impressive and interesting to take a peek at.

Things like automation bench, where the jump went from 18.1% to 41% and the benchmark I always like to talk about, the one that feels most correlated with how you'll actually feel when coding with it is Deep Sue. And this one, I'm going to be honest, it's interesting to me cuz I thought this would actually be a bigger number. From my experience, it feels really, really good at coding. And don't get me wrong, 74.1% is a great benchmark on Deep Sue.

It puts it at pretty much the top of the list. It's about a 2% jump over 5.6 Soul. However, the reason I thought it would kind of be a bigger jump is if we take a look at Deep Sue here, Gemini 3.8 Flash is at 74%. Claude Opus 5 is saying 74%. Meta's Muse Spark 1.3, which came out yesterday, according to their website, is 75.4 on Deep Sue. So, this new Astra model is underperforming on the coding benchmark that most people are paying attention to compared to Muse Spark 1.3, which is just really surprising to me.

Now, I do say according to the Meta website because, again, if I look at the Deep Sue website, Muse Spark 1.3 isn't actually listed on their website yet. We have 1.2 down here, but 1.3 is not showing on their ranking yet. But then again, neither is GPT-6 Astra. Now, where the benchmarks start to get really interesting are down here, Terminal Bench Science. It went from a 22% to a 64.6% massive, massive leap. GPQA, Google Proof Question and Answers, I mean, it's basically saturated benchmark at this point, same with Frontier Math Tier 4.

And then the Arc AGI benchmark, it went from 7.8% to 99.9%. I mean, it's basically solved this benchmark. But for additional context, the Arc AGI 3 tests how well agents learn as they solve unfamiliar interactive tasks. GPT-6 Astra saturates the evaluation, scoring 99.9%. The average human tester scored 48%. Now, the other benchmark that I often refer to is artificial analysis because this is like a sort of aggregated benchmark.

They take the scores of a bunch of different tests, put different weights on those different tests, and then rank what they believe to be the objectively smartest models. And this one I find kind of interesting because GPT-6 lands in fifth place. It pretty much tied with GPT-5.6, which if you've used 5.6 and then you use six, it definitely feels like a big leap, which makes me sort of question the legitimacy of this benchmark just a little bit.

Now, I still like the aggregated weighted take that artificial analysis does, but my tests, I've been way more impressed with GPT-6 Astra than I have been with Opus 5 and Fable 5. I haven't personally tested MuSpark yet on anything but BUS bench, but it's on my list to play with next. And when we come down here and look at cost per task, like I talked about in yesterday's video, Fable 5.1 is the most expensive per task to run.

And GPT-6 falls down here at about a dollar 67 per task, which is, you know, a little bit more expensive than GPT-5.6. So, if you're using this through the API or if you have a ChatGPT account and have usage credits, well, you might find you use them a little bit quicker or it's a little bit more expensive through the API if you're using six versus 5.6. But let's take a look at BUS bench. Here's what GPT-6 generated, and according to my AI as a judge here, GPT-6 falls into first place, just above Claude Fable 5.1 and the two Gemini flash models here.

Now, MuSpark 1.3, which actually performs better on both Deep Sue and the artificial analysis benchmark created this SVG of Gary Busey. Now, if you're not familiar with SVGs, they're basically images that were created from code. So, the code is basically telling it where to draw the lines and the circles and the shading and things like that to generate the image. So, it is pretty much a coding test to see how well the code is at drawing images with code.

And again, here's Muse Spark 1.3, which scored 5.7 on my AI as a judge. And here's GPT-6. Like, I don't even think they're close, which is again why I was slightly confused by the difference on Deep Swat and artificial [laughter] analysis. But again, artificial analysis is taking a bunch of different benchmarks into account and not just coding. Now, digging in a little bit deeper here on the GPT-6, it does say NA for cost.

That's because I had early access. I didn't run it directly through the API, so it didn't get the cost specifically. However, the AI did estimate it would have cost about $1.94 to generate this. But again, that's just an estimate. It used 63,858 tokens and it took 9 minutes to generate. And when I do my generations on Busey bench, I always use the highest mode I have available. So, this was actually using the max mode.

Now, if you've been watching my videos for a while, another thing I like to test just for fun is to see its kind of design taste for making a game. So, I always like to tell it to go and clone Mega Bonk for me. If you're curious what the actual prompt was, it was literally create a clone of the popular Mega Bonk 3D game. I always like to add 3D game because in the past it's made like a 2D top-down version. I just like to specify that.

And honestly, this is one of the most visually appealing versions of Mega Bonk I've gotten so far. Here's what it looks like. I can be a knight, I can be a ranger, or I can be a mage, each with different starting weapons. So, let's start as a mage this time and click let's bonk and you can see it's pretty good looking. I mean the character is one of my favorite character generations from any of these mega bonk tests that I've done so far.

Like it's just a cool looking character and it's all done with 3js. It didn't actually build the models. So it it coded all of this up and yeah, it looks really good. I'm going to quit out of this. I want to try the ranger, see what their weapon is. So they have more of like a bow and arrow where they're shooting just one arrow at a time. And then if I use the knight, they have more of this like bonk which is like this wave.

I think it's supposed to be hitting with a hammer. It didn't actually animate the hammer moving very much. I guess it kind of did a little bit. But yeah, the hammer animation is kind of leaving a little bit to be desired. I just really like the design aesthetic. I think it did a great great job and the game works beautifully. It works really smooth. And one thing I found really interesting was I prompted this game and it built the whole thing in 8 minutes.

In the past when I've done this mega bonk test, it would take an hour and a half, 2 hours to make something like this. This one, it did the whole thing in 8 minutes. Now to be honest, I'm not sure if it was super quick because less people were using it. It wasn't in public access yet. And so because there was only a handful of people that had early access, it was loading a lot faster or if this new model is just way faster.

But in everything I've tested so far, it worked really really quickly. Now another prompt that I like to test just to see what the AI comes up with is create a beautifully designed interactive website that shows off your capabilities as a model. The site can focus on whatever you want but design it to really impress me with your capabilities. I've gotten some pretty fun websites out of that prompt in the past. And here's what it generated.

I'm going to hide my face real quick so I can show it off a little better but it created this like world simulator. So I can control the Earth. And look at the details. You can see like this is the dark side of the Earth and all of the city lights are on over here. It's not Earth, I guess. It's Orbis, a little world uh that it created. So, it's it's like a simulated world similar to Earth. But, it gave me all of these sliders.

So, I can see what this world would look like if it got a lot more sunlight. You can see how it dried the whole thing out. Or a lot less sunlight. You can see how the whole thing freezes over. I could raise the sea level and just make it a water planet. Or lower the sea level and make the whole thing just completely solid. I can change the level of rainfall, which affects how much clouds there are and how green it is.

But, here's what I found the most cool about this is that I can change what this world looks like. I can raise land. So, any of these areas where there's no land, I can click around and actually add more land to them. I can carve an ocean. So, I can go through here and actually add oceans into it. I could add more uh seed life and make areas greener, add more plants. I could leave craters. So, there's like an impact thing.

So, if I wanted to put a crater here, boom, an asteroid just hit that spot. Or an asteroid just hit right [laughter] there. And you can see what would happen and how it would affect the the world. But, check this out. Down here, it shows the habitability. 85% habitability, surface temp 16.5° C, population 1.35 billion and rising. Now, if I crank up the sunlight so it gets drier, now 10% habitability. And we can watch the population just drop like a rock.

Bring this back to 50% habitability goes back up, population goes up. Reduce the sunlight. Once again, we see habitability start to drop. We'll watch the population drop. And it's just this really cool like little world simulator. And there's some like beginning templates that you can start with here that changes everything for you to see what these other worlds would look like. And once you build a world that you really like, it'll build a little postcard of the planet that you made.

But I think probably the thing I've had the most fun doing is getting this new model to use its computer use functionality to go and play around in Blender and Unreal Engine for me. So first I prompted it with I want you to take control of Blender and create a 3D figure of a humanoid wolf. So it took about 8 minutes and it created this little figure. Now this is just a screenshot of it, but if I pop open Blender here, you can actually see the character it created.

It's not the best there's definitely quite a bit of wonkiness to it, but this was a one prompt thing. I could have gone back and forth prompted it over and over again and got it more dialed in, but I wanted to see what it would do with one prompt and this is what it designed for me. And then I went, "Hey, can this like rig it up and actually animate it for me?" So I said, "Can you rig this up, remove the stand and then animate it?" The original version had that little stand on the bottom.

It took 6 minutes and it says, "Done." Removed it, added a 50 bone rig, and created an 8-second looping animation and it created this animation which left a little bit to be desired cuz the head's just kind of moving and the tail's wagging, but there's not much else going on. So I said, "Now animate it so it looks like it's running." And it made it so it looked like it's running. I mean, little little wonkiness with the hips, but it made it so it looked like it was running.

And if I open it up in Blender here, you can see that's actually what this animation is right now. It's actually the wolf running. But it did this by taking control of my computer, opening Blender, drawing out the wolf for me, and then after that it rigged it up with 50 different bones so that it knew where to move and then it created the little animations for me all through prompts. Like I have no idea how to use Blender.

I'm not good at Blender at all. So that was pretty impressive. But then I was like, "Let's take this to the next level. Can I make this as a character in a game that I can run around as?" So I gave it the prompt, "Create a beautiful forest world in Unreal Engine and then add this wolf as a playable character that I can use and run around as within the world. That was supposed to say run around as. I said round around, uh, but it knew what I was talking about, luckily.

It took 35 minutes, and it said it created Whisperwood, a forest with winding trails, a pond, wildflowers, etc., and it added my character in. And here is Whisperwood. Again, this is all computer use. It took control, opened Unreal Engine, and made this. Not the greatest running animation, but it built this whole world. I can jump around. I can play as this wolf. So, I can use my arrow keys to look around. I can jump.

Shift is to run. And it built all of this for me. I know people who know what they're doing with Blender and Unreal Engine are going to be very, very unimpressed, but as somebody who doesn't know how to use [laughter] either of these tools, just telling ChatGPT to prompt this into existence for me, yeah, that for me it's wild. I'm I'm super impressed with what this was capable of using computer use. Oh, but I just found the edge of the world.

And again, this was just like one prompt at a time. Like, I wasn't trying to optimize or dial it in or really, really improve the look of the wolf or anything like that. A few more prompts, and this would have looked like really good. And quite honestly, most of what I created actually feels kind of janky compared to some of the stuff I've seen people posting on X that are way more creative than I am. So, I figured I would quickly share some of the stuff that I saw other people share that they made with GPT-6.

I don't have to look very far on X, cuz again, that's my whole feed right now. But, here's something that Matt Berman created, like these little planet worlds where you're this character that can move around in them. It looks really good. I'm pretty sure this is all 3JS. Like, this didn't go and create separate assets. This is all written with code, and it looks really, really good. Matt Schumer here, GPT-6 Astra built this Manhattan world in Unreal Engine over the course of a week.

So again, somebody who managed to get it to just take control of Unreal Engine and build inside of that world. And yeah, it looks really, really impressive. Matt also created this 3D horror game here that looks like some sort of first-person shooter just without the shooter part of it. [laughter] But it looks pretty good. And another one here from Matt Berman, which is like a Fall Guys clone where you're, you know, running through these obstacle courses and trying not to fall off.

Looks pretty decent again. Here's another one from Matt Schumer. He claims he asked it to create a world in Unreal Engine and fill it with humans and each one have an Astra-powered agent who all work together to survive. A day later, he was in his bedroom, heard some voices coming from the other room, came out of his room scared, and it was the Astra agent. They actually started talking to each other. So here's the clip of what he shared of that. >> Thanks for helping, everyone.

Good to have a place in my home. >> You're welcome. >> Now the video here looks ultra compressed. I'm sure it looks better on his original system, but pretty interesting. Pietro here has some demos. He made this 3D model and the animation all from a image that he gave it. He made this Atlantis underwater simulator, which is one of the more impressive things I've seen for sure. A Beyblade game that you can play. A game that kind of looks like Wave Race from N64, but maybe even like slightly better than the N64 version graphics-wise.

Ethan Mollick here had it create the Library of Alexandria here. Now, I have the audio turned off, but it will actually explain things to you. >> Around 250 years before our era. >> And we're going to see more and more demos of what other people have built with it and what it's capable of over the coming days. But man, it is it is one of the more impressive models I've played with lately. I just did a whole video yesterday about how this release schedule of new models like every other day coming out is sort of getting frustrating because they all feel like small leaps, but this one does to me feel like a bigger leap.

It's one of the models I've had more fun with than others, especially getting it to go and use Unreal Engine and Blender and things like that for me. Really, really fun. I'm excited to use it to go and overhaul like my dashboard, find ways to improve it and make me even more productive on my daily routines, things like that. So, really, really exciting. Again, I'll talk about it a little bit more in tomorrow's Friday news breakdown where I break down all the news for the week, but I wanted to make a couple extra videos this week because a lot of models dropped and I didn't want tomorrow's video to be like an hour and a half-long video talking about all this stuff.

Plus, the GPT-6 launch felt like a big enough leap in model capabilities that it warranted its own video. If you like videos like this and you want to stay looped in on all of the latest AI news, again, I also make an end-of-week video that comes out every Friday that breaks down all of the AI news stories for the week. So, if you only want one video a week that breaks down everything you need to know in the world of AI, maybe consider liking this video and subscribing to this channel.

That'll make sure that those news videos show up in your YouTube feed. I drink from the firehose all week. I keep up with all of it so that ideally I'll be overwhelmed on your behalf and you can just watch one or two videos a week and be fully looped in on what's going on. That's my goal with this channel. Again, consider liking and subscribing if you're into that kind of thing. I also have a free AI newsletter that comes out twice a week that will also keep you looped in.

I'll put that link down in the description below, but thank you so much to everybody that tunes into these videos and nerds out with me about all of the latest cool tech stuff. I am upping the amount of videos I'm doing outside of the Friday news video because I feel like there's a lot to talk about right now. So, I'll make a few more videos than I used to make. I've got a really cool project coming out next week and some really cool collaboration videos coming up as well with some other creators that I'm super excited to share with you.

So, again, I really really appreciate everybody that tunes into this channel and nerds out with me and obsesses over tech like I do. It's a lot of fun and the community that's been growing around this channel has been awesome to see. So, thank you again. I really really appreciate you. But, that's what I got. Hopefully, I'll see you in the next video, which will be tomorrow's news video. See you later. Bye-bye.