GPT 6 Astra Makes New Things Possible transcript

GPT-6 Astra released on Friday, and I have been blown away by this model. It solved a bug in Bridgemind that was unsolvable for six months. Fable 5 couldn't solve it, Fable 5.1 couldn't solve it, Grok 4.6 couldn't solve it, GPT 5.6 couldn't solve it. But GPT-6 Astra solved this bug in one shot. I'm gonna be covering that and more in this video. If we do get to 1,000 likes on this video, I will be giving one of you a $200 ChatGPT Pro subscription.

All you have to do is like the video, subscribe to Bridgemind, and turn on post notifications and join our Discord community of over 17,000 builders. Bridgemind is the home of the vibe coding movement. But first, let me get started by just breaking down one of the craziest weeks that I have ever experienced in the world of AI. We got four new frontier models this week: Fable 5.1, GPT-6 Astra, Muse Spark 1.3, and Gemini 3.8 Flash.

And I'm gonna be specifically covering GPT-6 Astra in this video. I will be covering these other models, but just briefly here, I tested each of these models on the same exact prompt, and this is what Fable 5.1 created. It uses Three.js to basically create a scenic, overview of the ocean with a sunset and clouds, and this is what Fable 5.1 created. And this actually took Fable 5.1, over two hours to create this scenic sunset.

And GPT-6, with the same prompt, created this in 13 minutes. And this is definitely one huge aspect that we're gonna be talking about in this video of GPT-6. It's an incredible strength. But this is what GPT-6 created. You guys can let me know in the comments down below what you think is better. This here is what Muse Spark 1.3 created Holy benchmarks. And this is what Gemini 3.8 Flash created. So you guys can be the judges of this.

They all used the exact same prompt, and these were the results. But that's the frontier this week that shifted, and just a quick results that, I think tells q- you quite a bit. Now, GPT-6 Astra does come at a price of $10 and $50 per million tokens, which is actually quite expensive. When we saw that price initially when it released, we were all a little bit disappointed because we were hoping that it would stay the same as GPT-5.6 Sol, which is priced at $5 and $30.

So we are seeing a pretty substantial price increase from GPT models here, but the speeds of it are quite good. Like I said, that prompt that I just showed you, it only took 13 minutes. But if you look here on OpenRouter, check this out. OpenAI is currently serving GPT-6 at 30 tokens per second. And if you look at Fable 5.1, it's actually serving it faster at 38 tokens per second. So the question is: why would the prompt that I used with Fable 5.1 take two hours and GPT-6 Astra only take 13 minutes?

The answer is token efficiency, and this chart tells a very compelling story about why GPT-6 Astra is such an incredible model. If you take a look at this, you can see that per task, GPT-6 Astra was taking 21,000 tokens, while Fable 5.1 was taking three times as much, 64,000 tokens. So even though they're priced the same, basically what you're seeing is that GPT-6 Astra is an incredibly token efficient model, which means that it's gonna get tasks done cheaper and faster, while Fable 5.1 is the exact opposite.

It's gonna take longer. It's gonna output more tokens. So this is one of the craziest stories with GPT-6 Astra, is that this is the most efficient model that has ever been created by OpenAI. Another thing that I did with Astra is I had it create a blender model of my office by passing in pictures of my office, and then I just had it literally rebuild my office in blender. And you guys can let me know what you think in the comments down below, but, I mean, this is Insane.

One thing that totally has shocked me is the difference in cost per task between Fable 5.1 here and GPT-6 Astra. Now, Fable 5.1 is an incredible model. I've spent $10,000 on this model this past week, and I will be making a review video on that tomorrow. But just to see this in a chart, and if you're a user of Codex or Claude, this is something that you need to take note of. You know, if you are a Claude user, on the $200 max plan, you get about 20 to 45 minutes of usage with Claude Fable 5.1 if you're using it w- on, like, multiple tasks at once.

But with GPT-6 Astra, it's less than half the price per task. So on a $200 ChatGPT Pro plan, I have not hit my limits at all. Now, we have been getting a lot of resets, which does, you know, taint the waters a little bit because we've gotten, like, two or three resets in the past couple days for OpenAI subscriptions. But this is incredible to see in a chart because remember, these are priced at the same amount, $10 and $50 per input and output.

So to see this, it really is an incredible reflection of those token efficiencies and how they end up playing out in real world scenarios in terms of cost savings. Here's another Blender model that was made of a rocket ship. So I just used the Blender MCP, and, you know, this is incredibly high quality I've never seen a model that is able to do 3D modeling like GPT-6 Astra can. And, you know, when we were testing this on live, we couldn't even believe how fast we were getting these results.

You know, sometimes when we test a new model, we'll give it a prompt, like this one here, you know, to Fable or to Opus or, you know, it's a, a model like that, and we'll expect that it'll take, you know, maybe 30 minutes to an hour. But GPT-6 Astra literally built this, I think, in, like, five minutes. And, you know, whether it be the token efficiencies or what, for whatever reason, Astra is able to get tasks done incredibly efficiently and fast.

And you can see the quality of the output here is … I've never seen output like this. I also tested Astra on computer use, and what you're seeing here is I gave Codex a reference image from the start of the Vibe Coding with GPT-6 Astra stream where I was smoking the cigar, and it used computer use, and it used, it's called Krita, but, or Krita, I may be saying that wrong, but it used this software with computer use, and it created this painting from this reference image.

And what I will say is that you can see, like, the quality of this is not perfect, and there's a lot of inconsistencies, like you'll see here in a second. The Bridgemind logo that it draws at the top here isn't great, and even, like, this painting isn't perfect. But we are seeing the closest thing to AGI, or the start, the beginning of our path towards AGI, because to see a model use a computer like this is We haven't seen this before.

Anthropic has computer use for Claude, but it's, like, super slow, and it's not very reliable, whereas with OpenAI now, you know, this took about 25 minutes to do this, but it's using software like a human would on a computer. So it's, it's, it's shocking to me. You guys can let me know what you think. It's not perfect in the design, and that's one thing that I also do wanna talk about. I wanna talk about design with GPT-6 Astra, and I'm gonna show you an example of a F1 racing game.

So this is the racing game that I created with Astra, and, I'll just play this a little bit, and you guys can, you know, drop a comment down below what you think So the controls of the game are very good. I'm able to actually play it, which is great. So you can see that the controls are fine. I can actually race around the map. the design is-- I mean, you could even see if you stop the video, like the stadium was like on the track.

Like, I don't know, just like some visual bugs and stuff, and just overall the graphics aren't that good. It's, it's good, but I'm now gonna show you what the same exact prompt with Fable 5.1 created, okay? And I think it really is gonna paint a picture for you guys for the differences between Fable 5.1 and GPT-6 Astra are definitely still what is a weakness, in my opinion, of GPT-6 Astra. So this is what Fable 5.1 created Yeah.

So you guys can, can see the difference But then the controls, like, s- are terrible. I promise you I'm not this bad at the game Yeah. So when it comes to design, I think that GPT-6 still struggles in design, but it is significantly better than GPT-5.6 Sol. So I would say it's not better than Fable 5.1, and I still do think that having a Claude subscription is worth it. Now, for example, look at this. This was the lava lamp that I created using GPT-6 Sol, okay?

And for those of you that have seen this test in the past, you know, this is a very good result. Look at the physics of the lava lamp. This is incredible. You know, this is a very solid result. maybe they saw Bridgembench and they benchmarked it for the lava lamp test, but it's, it's very impressive. So from what I've seen from the model, another example would be on-- if I actually go to the Bridgemin website, I did update this right here.

If you see this in-app demo, I rebuilt this using GPT-6 Astra, and you can see that this demo, I can now open it up, and it opens it up in a modal. And then I can interact with it like this, just like I would if I was using the app. And this literally is what G- Bridgemin 1 looks like, and GPT-6 Astra was able to create this. So the design here I think is very good. So in some areas, I think that Fable 5.1 is gonna significantly outperform GPT-6 Astra in design.

But what I'm seeing is that Astra is a leap in design for GPT models. So I would say not as good as Claude models, but a leap for GPT models when it comes to design. So now I do wanna show you guys how I'm using Astra in my workflow. So currently this is linear, and this has been phenomenal for real world work in Bridgemin 1. All of these linear tickets, I've just been using Codex, and I've been solving bugs. I've been resolving support tickets.

I've been working with real users. Over the past seven days, Bridgemin did do $14,500 in gross volume. So I've been using GPT-6 Astra. You know, this is not a fun side project that I'm doing for fun. This is a real business. And what I've found as a business owner and a developer is that I can give GPT-6 Astra pretty much any back-end task, and it's able to solve it. Here is how I have been using GPT-6 Astra in my workflow.

So right now you can see that it's opening stuff up. Some of these agents are literally using computer use to test the app. So it'll make a change. It'll use computer use. It'll test the change. It'll-- Like, it's incredible. But these agents, I've been using GPT-6 Astra in extra high, and the usage, you know, if you are ripping it like this, you will use your usage on a ChatGPT Pro subscription. it won't last like the whole week if you do work like this But this is how I work.

The usage resets are incredible. I currently have two ChatGPT Pro accounts, and this is by far the best model. One thing that it one-shotted, which is absolutely incredible, I still remember using Fable 5 and Opus 5 to try and get this working. Watch this. I think it's… Hold on. I think it's actually this one right here. So this is, this is Bridgemind, and in Bridgemind, you use credits to be able to orchestrate the app.

So literally using the app, I can do this now. Watch this. Hey, Bridge Would you please go over to agent mode and help me create a new agent that is called Bridgemind Merch Agent that will help me create the Bridgemind merch? Heading to agent mode Done Bridgemind Merch Agent is saved and its pane is open Okay. So the Hey Bridge functionality is a wake word. So it used to be I would have to go up to this orb right here and click this, right?

And then it activates, right? And it connects. But now I can literally just say, "Hey, Bridge." I still remember using Opus 5 and Fable 5 and working with that particular task for, like, two to three hours, maybe, like, a month or two ago, and we couldn't build it, and GPT-6 Astra just one-shots it. One of the biggest problems with OpenAI models prior to Astra has been hallucination. Literally, this is a model that deletes production code bases, deletes users' C drives.

It canceled all the Stripe subscriptions in Vibe Academy by writing a migration script that was faulty, and GPT-6 Astra has finally fixed the hallucination problem with OpenAI models. This is something that we were talking about prior to release, and we knew that this needed to change. If you look at this, GPT-5.6 Sol had a hallucination rate of ninety-two percent. That has almost been cut in half. This is the single biggest improvement from any lab ever.

It dropped all the way to fifty-one percent with GPT-6 Astra. This is huge for OpenAI, and it makes it so that you can now trust OpenAI models. You can trust GPT-6, which makes me feel so much better with it writing, and being able to, like, ma- like, confidently be able to merge code. Now, with Fable 5.1, what's crazy is that it actually went the other direction. Fable 5 was sixty-four percent, and Fable 5.1 increased to seventy-three percent.

So for the first time ever, we actually have a GPT model that is more trustworthy than a Fable model or even Opus 5. It's more trustworthy than an Opus model. This is incredible. GPT-6 Astra is completely in a league of its own. When I think about the current frontier, this is what I see. I think that GPT-6 Astra is literally in S tier. It is a tier above Fable 5.1, and I will make a review video on Fable 5.1 tomorrow because that also is an incredible model, and I need to show you guys what I built with that model.

But these models are different, okay? You use these models for different things, and what I see from GPT-6 Astra is that this is going to be my go-to model moving forward. It is way better than Fable 5.1. I'll kind of, like, go back to my point that I made earlier, where when we were on stream, we were doing some of these tests with GPT-6, and we couldn't even believe how fast some of the results were coming back, and that's what I'm seeing from GPT-6 Astra.

Like, even right now, if I go back to my current workspace that I just showed you guys a second ago, like, look at this. A lot of the agents that were running literally have already finished. Some- Sometimes when I give agents tasks, they will run for so long, and sometimes, like, they just don't have what it takes to get the task completed. But with GPT-6 Astra, I've just been able to one-shot practically everything that I've given to it.

You know, that wake word with, "Hey, Bridge," I was working on it, and I just woke it up right now, so now she's gonna start talking, or I have it to a woman. But you know, it, it's incredible because I was trying to build this for six months, and it was a bug. It wouldn't work. I had the functionality built, but it was buggy. It wouldn't actually work. And GPT-6 just one-shots it. Bridge, how do you feel about this?

That's wild. Six months of a stubborn bug, then one shot from GPT-6. So you guys get it. Interesting. This, this is an incredible model. I think that this is a massive leap for people that are on the frontier of AI. If you do not already have an OpenAI subscription and you haven't tested this out, this is legit. This is now the best model in the world, and I think that, yes, there are use cases for Fable 5.1, and there are use cases for additional models, and this month is going to be insane.

I don't see how Anthropic is even going to survive this, because you guys saw the chart. Fable 5.1 is way more expensive and way slower than GPT-6, so I just don't even know what's going to happen. But Anthropic is gonna have to respond with something or it's over, and that's why we love competition. But with that being said, if we do get to 1,000 likes on this video, I will be giving one of you a $200 ChatGPT Pro subscription.

All you have to do is like the video, subscribe to Bridge Mind, and turn on post notifications, and join our Discord community of over 17,000 builders. Bridge Mind is the home of the vibe coding movement. But with that being said, guys, let's keep locking in, making progress, and building the future together. And with that being said, I will see you guys in the future