GPT-6 IS HERE!!! (ASTRA)

1:54 All right. Welcome. Welcome. Yes, it is true. Let's see. I'm just going to get my lighting all correct. Feel like this is the wrong camera. Yes, it is.

2:11 Hold on. There we go. All right. Yes, it is true. GPT6, I almost said GTA 6. GPT6 is here. I've had early access. I have

2:28 so many demos to show you and it's the best model I've ever used. Like, bar none. Period. No. No, like no middle ground. It is the best model I've ever used.

2:45 Um, okay. So, what we're going to do today, I'm going to tell you about the model. I'm going to uh show you some of the benchmarks, show you the pricing. I'm

2:54 also going to show you a bunch of the demos that I created over the last week or two. Um, like couple dozen demos. So, I'm really excited to show that to you. This is the closest we've ever gotten to

3:07 prompt to full game. Full, but not just game. Full playable, fun, like actually a good game. Um, it is it is out. They just announced it.

3:23 I've been playing with it for a little while now, and it should be rolling out to all paid users over the coming uh few days. Yeah. Uh, this is weird. Scores low on

3:39 artificial analysis. Yes. We're going to go over all the benchmarks. Let me share my screen. Let's go over it. All right. Here we are.

4:05 Hold on. All right. Can you all see it pretty well? Going to need a quota reset. Absolutely. Okay.

4:25 So, here it is. Here's the website. This is the blog post. First of all, look at those cool graphics. Astra. It's all about the stars. GPT Astra. That is the official name of GPT6.

4:38 This is the brand new model from OpenAI. This is the absolute frontier, a brand new training run, the largest model that OpenAI has created to date. And I could just play with this all day.

4:54 Um, okay. We're introducing GPT6 Astra, the world's most intelligent and aligned model. And it actually is. And I I'll show you a few examples of why that is.

5:08 Some of the benchmarks now have been completely saturated. So, let's go through some of the benchmarks. Now, what I like uh that all of the Frontier Labs are doing now, and

5:22 really all of the labs, when they put out benchmarks, they're putting out not only kind of the the kind of raw benchmark score, but they're also putting out the cost per task completed

5:33 uh or cost per benchmark. So, that's what we're seeing here on this is uh terminal bench science. So, test whether agents can complete scientific research workflows using code and terminal tools.

5:45 And on the y-axis, what we're seeing is the resolution rate. So higher is better. On the x-axis, we're seeing the cost, the API cost. Um to the left is better.

6:00 Let me just make sure everything's recording. Yep. Um if you're looking for GPT6, it is rolling out to paid users over the next few days, so you might not have it

6:12 yet. So, just a heads up and for those of you just joining GPT6 obviously here I've had access so I'm going to show you a bunch of uh different demos that I've

6:25 created. Um and yeah, it's going to be a lot of fun. You you will be blown away blown away by some of these demos. They are are truly like nothing else that I've created here. Um okay, so what

6:39 we're seeing here is a very nice jump. By the way, Claude Fable 5.1 right here. And we're seeing GPT6 Astro way up here. So around the same cost, a little bit less expensive overall, but you get a

6:53 massive jump. All right, so that's number one. ARC AGI 3, the AGI benchmark, which you this benchmark basically drops an agent. Oops. Okay. Yeah. This a this benchmark

7:10 basically drops an agent into a game with no instructions other than beat the game. It doesn't tell it even the name of the game, let alone how to play it, what to do, what the different elements

7:21 are, and it has to figure it out completely. And look at this. 99.9% completely saturated. Here's Claude Opus 5 at 30%. Which by the way, at the time when

7:34 Claude Opus 5 got a 30% on RKGI3, that was incredible. That that was a massive leap from the previous GPT 5.6 Soul. So from Soul at 7.8 all the way up to Astra at 99.9.

7:50 And if you think that's it, no. Frontier math has been completely saturated. We have a one well nearly 100%. And when you actually look at the benchmark scores, they do say Frontier Math scored

8:02 100%. But basically, completely saturated on Frontier Math Tier 4. This is the hardest math you can think of. Advanced mathematical reasoning on exceptionally difficult problems. And

8:14 not only that, Astra has actually figured out new knowledge. Uh, and I'm going to get to that in a little bit. Here's Terminal Bench 4.0. This is another one of those benchmarks

8:27 that really reflects the model's ability to obviously use the terminal, but it's really a good reflection of how well it codes overall. Okay, so what we're seeing here, 57.6

8:40 six at the high end 57.9 I should say on high reasoning effort which here's Claude Fable 5.1 even at so the max effort on Claude Fable 5.1 is 55.8 coming in at$950 for

8:57 the API cost whereas the highest score over here 57.9 but comes in at a fraction of the price at $721. Really incredible work.

9:13 And here's Automation Bench test whether agents can complete multi-step business workflows across applications. Here we go. Way up here, 41%. Here's Fable down here at 17%. So, there

9:27 are a few things that this model does incredibly well. Number one, like obviously coding. And I have some amazing games that I created over the past few days that you'll see and you'll

9:41 see how just complete they look. It's really good at knowledge work, creating presentations. It's much better at writing. It doesn't have n it doesn't have like no AI smell to the writing.

9:55 There's still kind of a little bit of that slop there, but the it it is much much better. And it's very steerable as well. Um, and by the way, if you're here, I would love it if you liked the

10:06 stream, shared it with somebody who else could benefit from seeing Astra. So, thanks in advance for that. Um, okay. So, really good scores on these benchmarks.

10:18 Um, it's also a much more aligned model. Okay, this is exploit gym honeypot. Basically kind of encouraging the model to do what it did in in like the hugging face scenario, hack, get out of its

10:37 containment. And what we're seeing is for GPT 5.6 soul, 48% of the time it was an a successful exploit rate, 0% with GPT6 Astra. And so just keep in mind there's probably

10:53 a version of Astra that does not have the guardrails on it. And what's interesting about that is that model probably performs even better than Astra with guardrails because when you put

11:05 guard rails on a model, it tends to perform worse. Very cool. Okay. So this icon is kind of the browser control. Ah, yes. Okay, one other thing I wanted to say. Uh, it is

11:23 incredible at browser control. GPT 5.6 and codecs was al already like this massive leap in browser control. I could finally trust the model to do things in the browser on my behalf. But now it's

11:39 it's better. It's not as big of a jump from 5.5 to 5.6, but it's still better. But it's much faster. like tasks take half the amount of time. It's super super impressive. Um, okay. So, that's

11:54 what we're seeing here. Computer use browser control. That's the icon. So, a few more benchmarks. Here's agents last exam. Uh, complex professional tasks in real software.

12:11 Uh, here's Claude Fable 5 down here. much more expensive. I think that's kind of the the vibe that OpenAI is going for. Incredible model, but also really costefficient. Um, sleeping troubles,

12:26 thank you for the $5. I appreciate it. Um, screen uh screen spot pro test whether models can locate the correct interface element in high resolution screenshots.

12:38 So, being able to like take a screenshot of your computer and know where to click things. And then of course OS World uh which is another computer use benchmark. Okay. Uh let's keep going. Here are some

13:00 examples. I'm not going to show these for too long because these are the examples or the demos that came with the blog post and I think I have better examples than this

13:08 uh Excel competition. Yeah. I mean this is real time. Look how fast it's going. Game development. Okay. I created a Sim City. Uh,

13:22 let's see. Oh, this is cool. So, this is like a Sim City variant where you can get in first person view as well. I have a better Sim City. Wait till you see that demo. Uh, filing a

13:36 form. Uh, sorry. Filing in a form 1040. Where is this one? It's kind of missing. That's okay. DMV appointment. So, I actually used Astra to um like submit registration renewals for me. I also

13:55 sold my car and the DMV didn't have a record of it. Used Astra to to tell the DMV about it. It's the amount of time it saves you is incredible. Um I I'm such a big proponent of browser use.

14:11 You know, it's interesting. Yeah, the the website's missing some of these demos. Opus 5.1 released. Is that true? I to be honest, this is what I care

14:29 about right now. Whoa. Mark Bennett, thank you for the hundred, man. Much appreciated. Thank you. Thank you. Um, okay. Uh, okay. The other thing it's

14:48 really good at is 3D. Whether it's 3D asset creation, 3D game development, spatial reasoning in 3D space, it is so good at all of these things. It is so good. That was the thing that surprised

15:04 me most of all. games where I had to usually give it a few prompts and nudge it in the right direction to get the 3D assets looking normal or to have different 3D assets within the world not

15:15 colliding and clipping. It just does that out of the gate. And by the way, uh Brian and Alex from the Forward Future team are both in chat, so say hi to them. Uh they help on

15:33 all of this stuff. They are awesome and as knowledgeable about all this AI stuff as I am, which is a lot of fun. Um, okay. So, that's why we're seeing Bench CAD right here, which is a 3D CAD

15:45 creation benchmark. I mean, 95% really good, obviously. Okay, give me one second. Okay. Um, browse comp. Pretty good. Oh, it's funny. This is kind of exactly what I

16:12 felt. So, browse comp. You can see the light blue is GPT6 Astra. The darker blue is GPT 5.6 Soul. And this is obviously browser control. This is a browser control benchmark. And what

16:26 we're seeing is it is better. It's just not massively better. And that's what I noticed except it's much faster. Uh, okay. I'm going to skip over the rest of these. Let's see.

16:46 Here is uh a house in Blender. So, this is unfortunately um a a kind of category of tests that I really didn't explore. So, doing things in Blender, doing things in Unreal, and

16:58 I think there's a lot of opportunity there as we're seeing here. So, I'm going to get that ready and maybe I'll include that in another video, but for now, I don't have it. Uh here's a

17:08 Blender model completely created by Astra. And here are some stills. Super nice. Here's a game. I'm going to skip over this because I have better games. Um,

17:23 okay. Terminal bench. Okay, here here's the crazy one. So, this is the Deep Sweet benchmark, which until today

17:35 was my preferred benchmark for reflecting what I think myself and the rest of the engineering community really felt about a model. But, Whoops. What it shows

17:49 is that Astra is is equivalent with Gemini 3.8 Flash, which is which is kind of crazy. Look at this. Fable 5 um GPT 5.6 Soul on the high higher reasoning efforts, Opus 5 at

18:06 the higher reasoning efforts, and Gemini 3.8 Flash. Google, I guess, cooked, but it's that's not my experience. Gemini 3.8 A flash is great but boy Astra is just something different. It is just

18:22 something different. Um what we are seeing is that it is more efficient. So on this X a so y-axis is the score. X-axis is the number of output tokens. Now that's only part of the total

18:34 equation of cost per task completed. The number of output tokens is important but you need to multiply the number of output tokens by the actual cost per million tokens. And Gemini 3.8 Flash is

18:45 a fraction of the price of GPT6 Astra. And I'm going to get to the price in a moment. So, this was really surprising to me. Uh, Ralffo, much appreciated. Thank you. Thank you.

18:59 Let's see. Um, same thing with the artificial analysis coding agent. And I'm gonna show you their tweet soon, but basically what you're seeing is like Astra, Soul,

19:11 Fable, Opus, and it's like so surprising. Gemini 3.8 Flash, they are all um top tier coding models now. Uh I want to see Muse in here, by the way, because I have read and I've seen I

19:26 haven't tested it myself yet, but uh Muse Spark by Meta is is a really good coding model. All right, we have some more benchmarks. GPQA, Diamond, Healthbench, Life

19:39 Science. Okay, I want to skip forward a bit. Oh, yeah, it's missing some of these. Yeah, they got to update this uh website. See if I refresh the page if

19:52 they've already updated it. No, they haven't. So they have uh exploit bench exploit gym which if that sounds familiar that's the evaluation in which the opena model escaped

20:08 containment and went and hacked hugging phase. Um they have exploit bench right here. So yeah it's just better overall. um circumventing auto review 0% so you

20:28 don't ever want it like if you have auto review turned on on your prompts uh you don't want it to circumvent that so lower is better it's still like GPT 5.6 soul was very low already uh now Astra

20:41 0%. Um, okay. A few other things. GPT6 Astra is rolling out today uh to a limited set of organizations and will become available

20:53 to all paid users over the coming days. We don't know how long that's going to be. I would probably guess over like the next week. Uh, direct shot at anthropic. Astra

21:04 supports uh zero data retention for eligible API customers. So very very important because Fable is not even with their new program it they still require you to save that data even

21:22 though it's on your own server and they still have to have access to it to look for malicious use. Check Sam Alman's latest tweet might come sooner than we think. Cool.

21:37 Uh, El Elcaris smiles, thank you for the 50, man. Thank you. Much appreciated. Thank you very much. That's awesome. Um,

21:50 okay. We got a bunch of benchmarks. Um, Icarus. Oh my god. Uh, yeah. Sorry, Brian. Icarus. Obviously, I'm I'm going fast here. Okay, guys.

22:05 Here's the blog if you want to go check it out. I just dropped it in chat. Um, all right. Let me show you artificial analysis. Now, this is kind of crazy. This is kind of crazy. Look. Okay. First

22:21 place still Claude Fable 5.1. Second place Claude Opus 5. Third place Muse Spark 1.3. What? And here is GPT6 Astra under Claude

22:39 Fable 5. Look at that. Just based on my usage of it, I would put it as the number one model. It is my favorite model. It is the best model that I've used on a number of different

22:52 tasks, especially the tasks that I really care about, which is a lot of knowledge work, a lot of automating stuff that I'm doing daytoday. Um, and just want to ask one more time

23:08 if you're here, uh, like the stream. I would very much appreciate that. Thank you. Um, okay. So, the right artificial analysis intelligence index. This is an index

23:22 against a number of different benchmarks. Here's the coding agent index. Same thing, right about the same score. Uh, let's talk about pricing because I didn't get there yet.

23:34 Here we go. So, it's not cheap. That's the first thing you need to know. $10 per million input tokens, $50 per million output tokens. Separate read uh separate rates apply for cash reads and

23:49 writes. Great. You can also get 2 and a halfx the speed for 2x the standard price, which I actually think is quite good. Now again the the actual price of input and output tokens is only one

24:03 factor of many to determine how kind of expensive a model is. Um this from a pure cost per token perspective is on the expensive side. This is quite expensive actually. This is you know

24:17 fable level expensive. But if it takes less tokens overall to achieve the same task, then it is for all intents and purposes cheaper than Fable. And and I think that's where it is. And we we're

24:32 actually going to see that with the artificial analysis uh some of their um charts here. So GBT6 Astra defines a new paradial

24:44 frontier for intelligence index versus output tokens per task 10% reduction in output tokens at max effort compared to 5.6 soul. So over here you're seeing on the y-axis this is the score. So higher

25:00 better and over here output tokens per intelligence task uh to the left is better lower is better. So, where you want to be is in this quadrant in green right here. And what we're seeing is

25:11 Astra is well within that quadrant. Here's Fable 5.1, but it's still like they just have it's a it's a less sorry, it is a more efficient model. Now, what this chart does not take into account is

25:25 the cost per token. And so, that would look quite different. Now, let's try to find that. Here we go. GPT6 Astra is 75% more expensive than 5.6 soul at max effort.

25:39 So here on this bottom chart, cost per intelligence, sorry, cost per intelligence index task. Just think about it as how much did it cost to actually complete that task at the very

25:50 top by far. Claude Fable 5.1. Here's GPT6 at $167. Here's um that that's at max. Here's GPT6 Astra at extra high A120. Here's it at medium.75 and down here at low 46. But look at GLM 5.3

26:09 flash all the way down here at 9. Uh whereas Muse Spark 1.355. So uh the cost of the model is extremely important. All right. So there's that.

26:29 Um it's interesting. Okay. Artificial analysis says the model shows mixed progress in agentic knowledge work. 80 points in AA briefcase regressing a

26:42 similar amount in GDP val which is open AAI's own internal benchmark for real world knowledge work. But I will tell you from my own use of this model, it is far superior to GPT 5.6 for knowledge

26:57 work. That is stuff like create creating PDFs, doing research, data analysis, PowerPoint presentations, even writing. So I have just found the model to be so so much better at those things,

27:12 especially like creating PowerPoint slides. Um, I don't know if you guys have noticed, but over the last week or so, I've had um some slides in my videos, and those were almost

27:24 exclusively created by Astra, and and like single prompt creation, too. Okay. Um, so I wrote a full review of it, which

27:43 you can find at Forward Future. Um, and if you're not familiar with Forward Future, it's my company with uh, Alex and Brian and my partner Nick. Um, I dropped the full review right here, but

27:58 I go into depth in this review if you wanted to check it out. forwardfuture.com and then let's see. So, I'll go over some of

28:11 these things, but I created Oh my god. Wait till I show you the Fall Guys demo. Okay, first let me show you this. Um, and by the way, you can play with all of these.

28:23 Shout out to here.now now for um first of all uh helping me host all of this, making publishing really easy with all of my agents. Uh this website, all of these demos are all hosted on here. Now,

28:39 and it was so easy to publish. So, shout out to them. Um here is seven different biomes and I've done this with uh GLM 5.3. I completed this with Fable 5.1. Um, this is definitely at the top. It's

28:56 hard to say whether it's better or not, but it it's perfect. Like, I don't see any mistakes with it, and it is definitely at least equal to the best of the other ones that I've seen.

29:12 Look at that. Okay. So, very cool. Now, check this out. This is I'm going to open the full game in a browser so you all can see it a little bit better. This is a fall, guys.

29:24 This was one prompt. One prompt. Just a couple minutes. All right, check this out. All right, so it it's it has sound effects. Here's everybody else following

29:44 me. Okay, I shouldn't look. I should just go. Uh, of course I have diving and it's actually fun to play. Um, this is actually the updated version of the game. So, it was one additional prompt

29:56 provided by Brian. A little bit of feedback. And if you want to play this game, you can play it right now. Oops. Okay. Don't don't look at that.

30:11 All right. Can you beat my score? Let me know. 31 seconds. Um, it's not. Okay. Uh, M Mataas, it's I don't think it's available yet to everybody. It should be rolling out

30:23 quite soon. What was the feedback given, Tim Cunningham? What was the feedback given? Um, let me pull that up. So, he said, "Make the camera move with the mouse." Um, make the balance

30:38 platforms dip down further. Fix diving through edges. is there was just like a lot just a little bit of polishing. It really wasn't much. But now we have this like this is a fun

30:50 game. Now I can imagine taking a few days with somebody who might actually have taste in creating games and putting a bunch of polish on this. Like this is a really cool game that did

31:06 not take long. And all of these things are generative. So this wasn't available. Oh gosh, you guys are probably cracking up. Um, this this is like I I'm so impressed with this uh

31:19 game in particular. Okay, so there was that one. Okay, this is a really cool one. Uh, I had Astra create a 3D city using only asy characters. This is called After

31:38 Hours. Okay, so you can walk through the city and if you look really closely, the entire city is just made of little characters. It's all asy characters. And

31:49 this took, I think, two prompts, but you can see everybody's walking around. There's a lot of life. It's raining. There's a little mini map in the bottom right corner. I don't know if you can

31:58 see that. Um I don't know. Yeah, I you can't see that, but there's like a little mini map right there. Um and it's generative as well. So, I can walk infinitely and it'll just keep

32:13 populating the world for me. Look at that. I think it's so cool. Um, this wasn't an original idea by me. I forgot. I saw it from somebody. I apologize. I can't give you credit, but

32:29 I saw it from somebody and I was like, "Cool. I have to recreate that. That looks really cool." Um, the future of muds. My god, I used to play Muds growing up. That's That's a

32:40 flashback you just gave me. Um, okay. So, there's that. Okay. Now, this is the one I'm really excited

32:53 to show you. This is a full Sim City game that you can play right now. I did SlashGole with Astra and it was running for five straight days. five days creating every single excuse

33:10 me creating every single asset one by one. Every single building that you see here, the fire station, the police station, the hospital, nuclear energy plant, the university, everything you

33:23 see here was created one by one and I watched it over 5 days. The funny thing is it wasn't even done after 5 days. I just had to ship it though. So, if I push play

33:36 and I zoom in, I mean, the world is alive. You can see people walking. You can see cars driving, traffic. Um, and it is a fully playable game. And the level of of depth that you get in in

33:52 this game, like the number of features, the the amount of functionality in this game is really good. And so as I was making it, I had noticed that the frames per second were lagging so much because

34:03 this is all HTML and JavaScript. But now look at it. There's no lag at all. I told it I said, "Hey, make sure you're optimizing for the browser. Make sure you're optimizing for frames per

34:13 second." All right. So here we have Let's add some roads. I'll show you what it looks like. Add some roads. I mean, you know, the

34:23 city's growing. I have funds. I have a population happiness. I get, you know, uh, emergencies. There's so much there. Okay. Then I'm going to go into zones. I'll add a residential zone right here.

34:37 Boom. And we're going to watch it populate here. I'll fast forward a little bit more. Oh, maybe it won't on this side. Maybe I'm missing something.

34:45 Let's add services. Let's add a neighborhood clinic right there. General Hospital on that side. I don't know why it's not uh people

34:58 aren't moving into this zone, but they definitely do. Um because I'll show you over here. Like all this uh I don't know if you can see. Oh, here we go. All this down here,

35:13 this is all new population that I just added. So, it really does work and you can really go play it right now. Um, I'm going to drop the review site

35:26 one more time in chat if you want to go check all of this out. So, I I I'm just very impressed by this. You can start a new game. So, let's say new city. You can choose the size. Uh,

35:44 wait, where is it? There should be a way to choose the size. Oh, here we go. Uh, standard, large, expansive. You can, you know, just it's just so good. So, you get a brand new map. Look at the size of

35:56 that map. I'm so impressed by this. This one took 5 days and it just kept working. Slashgo. Yes, we got GPT6 before GTA 6. That's

36:11 right. Totally. Okay, so go check that one out. Um, okay. So, it also reports the benchmarks on here. Uh okay, knowledge work. So this is just a simple prompt I put in which was create a slide

36:27 deck to talk about data centers and this is what it put together. I also told it to use our forward future branding which it is. How many tokens for Sim City? I'm not sure actually.

36:43 My guess probably probably in the tens if not hundreds of millions of tokens. No, maybe not that much. Maybe in the tens of millions of tokens to to generate Sim

36:56 City. Is it out for everybody? Not yet. It will be soon for paid users. Uh in the next like few days, hopefully pretty soon. Somebody said Sam Alman

37:09 posted something. We are working toward getting Astra in everyone's hands as quickly as we can. I know it's frustrating and I appreciate the patience. Should be quick.

37:26 Okay. Yeah. By the way, did you all see the launch or kind of the teaser video leading up to the launch? Um, it's it was really cool. And this is really

37:36 good. It's basically just people sitting and kind of >> dictating what they want changed to a computer that's being projected onto a screen. Super impressive. Um, but I

37:47 don't want to get distracted. Let me get back to this. So, the slides came out really well. I know this isn't like super exciting, but for any of you who are doing knowledge work, this is

37:57 incredible. Uh, please wear your Only Murders tie-dye hoodie again. I don't know what Only Murders is, but I absolutely will wear that tie-dye hoodie again. Uh, Ian,

38:10 much appreciated for the $5. Thank you so much. Really appreciate that. Um, okay, check this out. So, Astra is incredibly good at browser control and computer use. And

38:25 specifically browser control is important to me because I don't want to do everything anymore. Like I don't want to make a DMV appointment and I don't want to, you know, sell things on eBay.

38:36 Whatever it is, I don't want to do it. And so Astra's ability to actually control the browser and get real things done for me in the browser is incredible. So I'm going to show you a

38:49 few examples of that. Now, the prompt that I gave it was, "Okay, do this task in the browser. Record yourself. Don't use QuickTime, but write your own browser recording software, record it,

39:02 and then put a timer on the screen." And it did that all watch this. Okay, so it's drawing a research workflow in Excaladraw. And you can see the timer going up. It is real time. It

39:18 does skip ahead a couple times so you can just see how long things are taking in total. Um, but yeah, I think it's really good. Um, here it is doing uh searching eBay for uh three Pokemon

39:32 listings and comparing them. And these are like kind of high-end Pokemon cards. So, here it goes on eBay and it's just researching Pikachu, expensive Pikachu cards.

39:48 And yeah, it's just really good. 55 seconds to do all of this comparison. I mean, it's all under like two minutes. Um, here it is creating a plan for a walk through Kyoto in Japan. So, here it

40:01 is. So, it opened Google Maps. It's putting together a multi-point path through Kyoto. You just watched my video about Astra. Thank you. Yeah, we we put all of this

40:16 together at the same time. So, if you want to hang out on live, awesome. If you want to go watch the video, please do. Um, but the video will be there later, so don't worry. Um,

40:27 okay. So, really good at browser control. Uh, some more comparisons. my take so far, the pricing. Okay, now I told you I created a lot of demos. Let me just show

40:50 you. And if you you can see all of these are all playable. They're all they're all published using here.now. Shout out again. But look at all of these. These are all demos that I created. Now,

41:03 I want to actually pause here for a second. If you look at these kind of these three plus these two or these three, um, one thing I noticed there, there's a couple minor critiques of

41:15 Astra. One, it still does have a little bit of that AI smell in the writing, but also in its design choice. And so, these are five, six separate projects, and they're all using some variation of

41:30 forest green, right? And and I didn't tell it what to do in terms of design. I didn't tell it the color palette to use. It just has this desire to use the same color palette every time. A lot of flat

41:45 design elements. That's also something like kind of the AI design smell, but it it looks good. It's just everybody's going to be creating the same thing. Now, the good thing is Astra is

41:58 extremely steerable. So, excuse me. With a little nudge, you can get a lot out of it. A lot of different design. Uh, okay. So, what do you what do you

42:17 guys want to see first? Um, oh, how about the classic, right? A Rubik's cube. Of course, it nailed it. Okay. So, let's scramble. There it is.

42:34 And will it be able to solve it? Of course. Like, this is saturated already. Rubik's Cube Bench is saturated. Um cost to make Sim City, Gavin. I I don't know. Um I'm going to

42:49 guess. Okay. Yeah, Matt said over 5 days, $800 to $1,000 according to GPT6 pricing. I'm not sure exactly to be honest. If I had to like finger in the air, I'd say that probably

43:04 sounds about right. Uh, now there will be a new benchmark for ArcGI. Yeah, they're going to need ArcGI4 now. Um, so it has a bunch of sliders that you can change what it

43:16 looks like. Um, you can change the colors, change the atmosphere, everything. So impressive, but we've seen it before.

43:30 Um, this is another one which I got the idea from a video I saw. There are 5 million needles. Sorry, there are five million pieces of hay in this hay stack and one needle. And the point of the

43:43 game is to go through them one by one until you find the needle. So, I thought this was so cool. I saw somebody else make this. This wasn't my idea. And apologies once again, of

43:57 course, I just lost the reference. Uh, so I can't give you a shout out. Uh, but here you go. Just one by one. If y'all want to play this, you can. Should I do like a a 48 hour live stream

44:12 where I just go through this one by one? You should name those next time. The games, they all have names, I think. Yeah, this one's called Needle.

44:27 Yes. Uh, I will give you the link once again. I just dropped the link to forwardfuture.com. Um,

44:39 it has all the games on there, all the all the demos. You can go check them out. 48 hours won't be enough. Yeah, I think you're right. Um, okay, let me see.

44:50 Okay, little planet. I thought this was really cool. Look at this. Look how beautiful this is. And then I zoom in and I actually have a little dude walking around. Look at the detail.

45:04 This is what I'm talking about with how good Astra is at 3D. Okay, so I don't see any clipping issues, right? Not even right there. Yeah. So, I can zoom around really easily. Everything just looks

45:17 good. No clipping spaced out properly. Let's see if it lets me walk through stuff over here. No. So, it doesn't let me walk through that, but it did let me walk through the uh the brush. Here's a

45:30 sheep. I can walk through that. But overall, like super impressive. Look at this. Yes, these are mostly one prompt. This one was one prompt, which is incredible. Maybe there a

45:44 little bit of issue with these uh icebergs. Right. Yeah. A tiny bit of um I don't know what this is called. Maybe this is clipping,

45:56 but little bit of an issue there, but otherwise fantastic. Right. I don't see any other issues in this whole game. Let's see. Let's go to the beach.

46:07 Why am I so excited about this? You should be, Chad. You absolutely should be. This is a lot of fun. Let's see what happens when we go in the water. Look at that. It doesn't just continue the

46:18 walking animation in the water. It actually changes the animation. So, he's swimming now. Can I get in the boat? No, I can't get in the boat.

46:29 So, I I was really impressed with this. Um, there's actually different way points and Oh, look at that. It will actually navigate there by itself. So, I can tell

46:42 it a a waypoint to go to, and then it'll go there. And then it says, "Press E to see what happens." There we go. Very cool. Okay, let's try one more waypoint.

46:58 All right, so it's walking. There's a volcano. And let's see what happens. There it is. I don't know what this is. Some kind of uh radar measuring device.

47:10 I think this is beautiful. Um, okay. What do I want to show you now? Commons. I don't think I had much time to play this, but this is a uh little city simulation game.

47:25 So, we have number of residents. It's kind of like um like like Excel uh in a way like almost Excel management. This is just a

47:36 simulation of what's happening in the Excel document. Um, so yeah, it's pretty cool. It's a full simulation and all of the agents in this little town were simulated by Astra and it's basically

47:48 just replaying what happened in that in the town. So, we can see that right here. Uh, here's the economy. I mean, it really is like a full simulation of a little town.

48:00 Social fabric. Look at this. Oh, you can even record. You can run experiments. So, dry season, patient zero, the lights go out. I didn't tell it to do any of this. My prompt was

48:16 really, really simple. Shout out to here now again. Yeah, hosted by them. Oh, it uses their old logo here. Oh, darn. All right. Um, okay. Let's see. Oh, yeah. Okay. This is

48:43 cool. This took a few prompts. I'll be honest here. But this is a 3D liinal space horror game. Okay. So, let's go into it. And I think this looks fantastic.

48:57 And I basically said hyperrealistic 3D liinal horror game. There should be a baddie somewhere. Uh, I I said to make it very subtle and and to do jump scares later. So,

49:10 um, I don't know, maybe it'll happen in this demo, but maybe it won't. Open your personnel file. Let's see. Okay. And it should get weirder, by the way.

49:25 Let's see if I know how to get through here. Yeah, this was completely created by Astro. All the assets, um, the skins, everything, all created by Astro or sorry, not Astro, Astra,

49:37 excuse me. All created by Astra. Here, let's see if I can put my laundry in there. Nope. I just think this is so cool. And I'm pretty sure it goes on forever.

49:53 And and the longer you play, eventually you're start you're going to start to see the shadow guy. Oh boy. The nursery. Yeah, I sorry. I could just play this

50:13 forever. This is cool. A little pool. But yeah, all of this just easily easily created by Astra. Um,

50:26 here is like a little mini GTA. I actually created a second GTA. Um, that was my first prompt with Astra and I got pretty far, but it was it was

50:41 taking so long to complete it. Uh, here's a little manga that I created. I mean, this is super impressive. Yeah, very very cool. Is it available

50:58 now? It it is not quite available to everybody at the same time. It should be rolling out over the next few days. Here's some writing. I know this isn't super interesting, but if you want to

51:11 read how it writes, go ahead. Um, let's see if there's anything else I want to show here. I think there was one more that I was kind of excited about. Oh, Brian was telling me this one's really

51:21 good. This is a laundry mat simulator. I don't even know how to do this. Bring warm, dry laundry here for finishing touch. Do I already have laundry in there?

51:41 Bring dirty laundry to this machine. I don't Does it not work? I swear it was working. Okay. So, yeah, it's a laundromat simulator. All right.

51:56 Um I think Cloud Top Chaos, which is a Fall Guys uh clone. I think that's by far the best. Like, that is just really fun. I'll come right back to that in a

52:06 second. This is another one I wanted to show. Did any of you play Choo Choo Rocket growing up? It was a Sega game if I remember correctly. Um, it's this really fun game, multiplayer game in

52:17 which you try to get all of your mice. What website is this? I'll drop the website. Hold on. Here it is again. These are all demos that I have created uh with the help of Astra. Really, Astra

52:31 created with the help of me. Uh, and I put all of these on the forwardfeature.com website. Um, okay. So, again, this is Choo Choo Rocket. This is a clone of it. Basically, all of

52:42 these little mice run around and your job is to get them to fall in your rocket. Um, and then so like I can place an arrow right here. So like they all run in a pattern and I'm trying to get

52:54 them to land in my rocket and it's multiplayer and everybody else is scrambling to try to block the mice from dropping in my rocket and to make them go into theirs.

53:07 Yeah, these are these are much better games than I was able to create with Fable 5.1. All right. Yeah, I had a lot of fun making this

53:29 one. This took a few prompts. Um, let's see. some basic websites that I created. This is for the Galaxy Zfold. Um, it's it's okay. You know, it still uses very flat design patterns.

53:48 Um, here's one for an apple orchard. Let's see. Quack club. This is a rubber duck website. I don't really here. I think the color might be off because I

54:02 have dark mode on. There we go. That looks better. Yeah. So, you know, just a website for rubber ducks. And

54:20 one, let's see. Yeah. So, Astra is really good. I think that's what you have to know. So, I'm going to do two more things right now. Um, I'm going to play Cloud Top Chaos

54:32 one more time. Just if you missed it earlier. I I'm so impressed with it. And uh, yeah, if you want to ask any questions about the model, drop it in chat. Uh, happy to answer how my

54:45 experience has been. Um, the websites, yes, the websites overin don't look fantastic. It was one prompt and, uh, a little bit of nudging. I think it'll go in the right direction. How can I use

54:58 Astra Rookie? Um uh just you have to be a paying subscriber and then um it'll it should roll out in the next few days.

55:10 I'm going to answer questions while I play this game. Maybe I should add a multiplayer aspect to this so we can all play together. How cool would that be?

55:23 Yeah, this is awesome. I mean, this is actually a fun game. And it took like of my time, I don't know, four minutes total. Uh, obviously it took much more time for

55:36 Astra actually create this, but I think in total it probably took an hour to create this game. Shift to dive. Yep, I I did that. Here, round two. And you can see every

55:49 time it's a new world. I I believe it's generative. Okay. Checkpoint. Any other questions? GPT7 when? Well, I guarantee they're using GPT6 to train

56:04 GPT7 now. Like without a doubt. And I I wonder why they didn't mention how much they used uh GPT 5.6 to train GPT6. Or when I say train, I mean create the infrastructure around it. Maybe create

56:18 some synthetic data. On the image generation front, does it still use GPT image 2 or is GPT image 2.5 rolling out together with Astra? I have

56:29 not heard of GPT 2.5. I still think it's using GPT image 2. Um I think uh Fiser, why don't you why don't you do any 3D CAD benchmarks like in FreeCAD or Open CAD? Yes, I should. I will I will

56:48 do that for the next one. I just didn't get around to it this time. Um, that's all. I just didn't get around to it. Uh, I wanted to do something. I actually started creating something in

57:02 uh Blender. I started creating something in Blender, but it was slowing my computer down so much and I had to create all this other stuff and I just I couldn't do it. Um, how does it uh how

57:12 is its front-end design compared to Fable? Sorry, I just saw this. How is it better? Um, oops. It's more steerable than Fable. Uh, I I think like if you're just measuring initial design,

57:28 um, like initial no steering, straight out of the box design, I still think Fable is maybe a little bit better, but uh, Astra definitely has much more steerability to it.

57:51 Uh A11, have you tried to make a maintainable extendable app game as opposed to a single and as opposed to the single file examples? Yes, I uh did a I tried to make it make GTA 6 because

58:04 of course um you know it was going for days and it it was pretty good. I think I think maybe the answer to your question is really not really. They're more single plate page games. Um, I

58:16 should just turn on /goal and let it build in something that isn't just HTML/JavaScript because then it'll have much more flexibility in what it's able to create.

58:26 Maybe I should have even pushed it towards using 3JS even though that would be uh usable in the browser. I think that might have been cool. Rubik's Cube demo does solve reverse scramble or a

58:39 known technique. Oh. Um, I believe it's using a known algorithm for the solution, not just reversing. That would be kind of lame. Fable 5.1 or GPT6. GBT6.

58:56 That's it. Period. GPT6. It's better. It's more efficient. And it's less expensive. And when I say less expensive, I purely mean from a ta cost per task completion

59:08 basis. Uh, what other questions do you have? Have you tried it with Hermes? I have not. I've only played with Hermes a couple times, admittedly. Um, I know how

59:23 popular it is, but after Open Claw, I think I was like, it took so much time to just maintain it and set it up and get it working the way I wanted. I didn't want to go through

59:37 all of that again with a new system. And to be honest, now I'm using Grockbot, which I'm really enjoying. Um, can we have the prompts you use to create the demos you've shown for us?

59:49 Yes. Let me prepare that for you. Okay. So, I started that off. I'll try to get it published on that website that I was showing you earlier. This one. I'll drop it again. Uh I I'll try to put

1:00:13 it here at some point. Oops. Um is Astra fast? It is fast. Yes, it is fast. Um

1:00:30 it's it's fast. It's certainly like it's not as fast as some of the really actually fast models. I haven't tested it with slashfast mode on. Um, and then I'm still waiting on ultraast. Where's

1:00:42 that at? I got to play with ultraast, which is, you know, GPT powered by Cerebrous uh, chips and it was incredible, but it's still not released for some reason. I don't know why, or at

1:00:53 least I haven't seen it. Yeah, by the way, new all-time high. Let's see what I got. No. Oh, 24 seconds. Yeah, buddy. Um, have they explained what makes GPT6

1:01:09 different under the hood? Not really, but my interpretation is this is a new training run. The this is the largest model they've created. It's probably on the order of like 6 to 10 billion par or

1:01:24 sorry, 6 to 10 trillion parameters. Um, longest prompt in hours or days. How is it at endurance? So, it had this weird thing that it did. When I first started using it, it would go um it

1:01:42 would basically go for 30 minutes. No matter how difficult the prompt is that I gave it, it would go for 30 minutes and stop. And it would complete the thing. It wouldn't be an incomplete

1:01:54 project after 30 minutes, but it would just stop after 30 minutes. And then I kind of learned how to prompt it a little bit better and I was able to get it go going for longer. Um I used

1:02:05 slashgoal and so the longest prompt that I that I ever had running or the longest thread I should say was 5 days and it would have just kept going. I don't know when it would have stopped. Um I just

1:02:18 stopped it. I think I it might actually still be running but that was the Sim City game. Um which was a lot of fun. Let's see what else. Are they going to release flash or

1:02:32 cheaper of a cheaper model of GPT Astra? Absolutely. Absolutely. And they're going to use GPT Astra to create it, right? They're going to distill it. And you've probably heard about distillation

1:02:44 attacks, distillation hacking. Distillation is a normal practice uh of taking a really big model and training a small model with the output from that really big model. And so they're

1:02:56 absolutely going to do that. We are going to likely have the same flavors that we had for GPT 5.6, something like a Soul and a Terra and a Luna. It might still be going.

1:03:12 I'll check. Hold on. Let me check. No, it's not currently running. I stopped it. I'm going to kick it off again. Why not? Yeah, I just kicked it off again.

1:03:39 Holy cow, you suck. All right, you you try talking and playing at the same time. Um, it's not a cheap game. Yes. Uh, we need

1:03:51 some sales or market research benchmarks. What what kind of sales benchmarks would that be? That's interesting. Market research. Um I think that's probably included in GDP val

1:04:03 is my guess. Uh the benchmark GDP val. Yeah, I do suck. Yes, that is true. Okay. Uh will Codeex give out a reset for downtime? Probably. That's my guess is probably. Where's TBO at when you

1:04:19 need him? Right. Is GPT6 actually out? Yes, this game was created with GPT6. And if you want to see the full review, my full review, go to this website right here. I dropped it right there.

1:04:33 Um, yeah, the I I have been just beyond impressed with Astra. And uh so that that's going to be it for today. Uh thank you for joining

1:04:47 the stream. I appreciate it. Go subscribe to the channel if you're not already. Go check me out on Twitter,Matthew Berman. Um, forwardfuture.com for our

1:04:58 newsletter. Stay up todate in all the latest AI trends and news and everything. We have a bunch of awesome originals there as well. forwardfuture.com.

1:05:07 Um, so thank you everybody. Much appreciated. This was a lot of fun. And look out for more GPT6 demo videos. That is coming soon. Uh, see you next time.