GPT-5.6 is INSANE!

Today, we have three new models from GPT 5.6, massive step up. I can tell you now, I'll show you some of the tests that we've gone into, but we have a full solar system of new models directly with GPT 5.6. If you don't have access to it right now, by the way, it is rolling out slowly. So some people around the world are getting it sort of step by step, if that makes sense.

Now, OpenAI's new GPT 5.6 has three new layers, the Sol, Terra, and Luna. Luna is a fast, cheap one. Terra is the everyday model. Sol is the beast.

If you want to see what Sol can achieve, let me show you some of the latest creations from GPT 5.6 that we've tested on GoldieBench on the leaderboard. So let me show you some of the stuff it's created. It looks really, really cool. We've tested out relentlessly, already built out 33 different projects with it, and you can see these on GoldieBench.

So let's check them out. For example, we have this game, and it's looking really, really good. You might say, is it Fable 5 level? I would say at this point, it's on par.

Like, it's really, really good. It can create some amazing stuff. I have a little feeling that if I was to choose one model and only one model that I could use forever, it would definitely be Fable 5 over GPT 5.6, but Sol is awesome, as you can see, and creates some amazing stuff. Let's have a look at another one.

So this is like an open world game, as you can see right here. It feels really nice to use. I was testing out Grok 4.5, and that impressed me yesterday, but this is way, way better. It's just more fun.

The gameplay is sharper. It feels just a lot nicer visually. It's a lot smoother and a lot more interesting in these open worlds as well. So it looks fantastic, as you can see.

Let's have a look at another one. This is like a Doom-style game, as you can see, and one of the things that I'll say here is like, quite often, models struggle to really build out like a smooth game like this, and this has worked perfectly, and it just feels really nice to play. It feels smooth when I'm looking at the creations it's made. It's like a racer game as well, so Sol is the absolute beast.

In fact, I probably wouldn't use any other model apart from Sol if I had the choice. I think that's what I'll try and stick to for now, but if you're wondering about the other two models, I'll come on to those in a second, but just to show you like the creations here are absolutely unbelievable. Here's a full video that it created, so it created a full video animation here. It looks super nice.

This is, you can see here, for example, as a progress bar at the bottom, it's created all of this. Like, I haven't touched it. I didn't create any of this. It just, from one single prompt, does a test, and it looks fantastic, as you can see right here.

So this is fully, like a video fully created with GPT 5.6 Sol. So it can create videos, it can create games, it can create landing pages, it can create awesome stuff, as you can see right here, and yeah, unbelievable. All of it was really, really good. There was nothing I looked at where I was like, ah, not sure about that.

Like, all of it that I checked out was just fantastic, and it's a really, really good model, which is the first time that I've said that about GPT this year, apart from the image model. Their image model was goated as well, to be fair. So let's run through it. What is GPT 5.6?

How does it work, et cetera? So this just dropped today. OpenAI made GPT 5.6 generally available, not as one model, but as three. So same architecture, three sizes.

All three share 1 million context window, 128K max output, and February 2026 knowledge cutoff. You can see the prices here. So for example, of course, like Sol was a lot more expensive to use than Luna, which is a lot cheaper. One thing to bear in mind here is when you're using it, like, you can just use it with OAuth.

So that means you don't have to use it with an API. So for example, if I was using my AgentOS, and I'm using Codecs inside here with GPT 5.6, so we can switch between the models. I can use a free model as well, like OmniRoot inside Codecs if I want to. But if we went to the workspace here, what you can see is that we can flick through between everything that we've built.

It's inside one place. And this was all created with OAuth. Like, I didn't pay for the API to build this stuff, which is absolutely awesome. And it's the same, for example, inside Hermes.

We can use GPT 5.6 as a separate profile. And the great thing about that is it doesn't cost us any extra. Whereas, for example, if you're using Fable 5 or any form of Claude inside Hermes agent or any sort of agent like OpenCLAW, you'd have to pay. So there's a big difference between these.

So how would you use these? Well, if you look at the position here, Sol is more for, like, flagship, hard multi-step agentic work. Terra is the everyday production model. And then Lunar is, like, high volume, latency-sensitive work.

So Sun for moon, biggest to smallest. And this was actually delayed. So there was a government review first. It dropped in preview about a week or two ago.

And then from there, they've finally released it to the public, which is fantastic. So you can see here, if we look at the benchmarks on intelligence index, Fable 5 is still outperforming Sol by one point. So it's 60 versus 59 versus 55 versus 51. This is from the artificial analysis index.

And so Sol is literally right up there with Fable 5. Cost per tasks. So you can see here, Sol is a lot more expensive, like almost twice the price of Terra. Lunar is 80% less than Sol.

And then we've got Terminal Bench as well. So Sol Ultra is right at the top there. I don't think I've even tested Sol Ultra. That could be fun as well.

So Sol itself is really, really good. And then if we have a look here as well on Terminal Bench 2.1, Fable 5 is 83.4. Sol Ultra is 91.9. Sol is 88.8.

Now, from what I've tested so far, it does look really good. And the thing I would say here is like, don't believe all the benchmarks. Don't listen to all the benchmarks. Just test it yourself.

That's why we created GoldieBench so that you can have a look at what I've built with it and then test it out. So, you know, on our actual creations, it's right up there. Like it's one of the best models you can use undeniably with GPT 5.6 Sol. So if you're wondering like how long does each one, so we actually gave each model the same task with one shot.

This was for a 3D space game. And from what we saw, Sol takes a lot longer to build. So six minutes versus one minute 37 for Luna. The total cost was $1.18 versus 0.12 from Luna and 0.43 for Terra.

Code written was much bigger from Sol. Measured through was a lot slower with Sol as well. So you can see 198 tokens per second versus Sol's at 109 tokens per second and builds it ran error free. They all ran error free, which was fantastic as well.

If you're looking, if you want to see like what the actual creations look like. So this is Sol versus Terra versus Luna. So Sol's version here, as you can see, like 3D game, pretty fun to play, et cetera. Let's have a look.

I mean, there's a huge difference between, it's not even comparable at that point. Look at Terra's version versus Sol. Like Sol is just a much better model. And then you've got Luna, which is kind of the same level of bad as Terra.

And at the same time has less going on, less interesting, right? So Sol is literally like the best model by a long, long way. Also built out some landing pages. So if we test these live, and by the way, if you want to see this full guide, it's available at asianosk.guide.

So this is the page that we built with Sol. Looks super nice. I look at the colors. It doesn't feel like, you know, AI slop to me.

Like it looks super nice. So we built out a full landing page for the aircraft for boarding. It looks beautiful, super interesting, looks fantastic. If we have a look at the version from Terra, the medium model, this is not bad still.

Like it's kind of, it kind of feels like Opus 4.8, honestly, when I have a look at this, when I'm testing out, it still looks good, but it just, it doesn't have that nice UI that Sol does. Sol's output is way nicer. And then if we have a look at Terra, actually it did a very good job. For such a cheap model, that did a great job at building a landing page.

So fair play to Luna on that job. And then we built out an analytics dashboard as well, as Tessie's out. So this is the one from Sol, looks great. The buttons on the left-hand side don't actually work, just to be 100% honest with you.

This looks very similar to be fair. I mean, they all look very similar at that point. I can't really see a big difference between them. Now you might also say, well, you know, Sol most expensive model, like, isn't that a waste?

But actually if the outputs are like 10 times better, especially for the first test, then it's 100% worth it, right? That's the way that I look at it. It's kind of like, would you use Haiku or Sonnet or Opus or Fable 5 if you had a choice? Of course you're going to go with Fable 5.

So here's the way that I would recommend using it to get the most out of this stuff. You know, the old way would be like one model for everything, like just a flagship model for every request, big or small, summaries and drafts at frontier level, slow responses and tasks that need to know that depth, and a build that gets more expensive with usage, not with difficulty, right? And then with this method, I would recommend that actually you switch. So you change the model depending on what tier you need it.

As you saw, for example, like Terra and Luna did a pretty good job on the website. So if it's like a high volume work where you can use Luna, if it's just day-to-day work, you might use something like Terra, but on the harder genetic stuff, like building out big projects, coding something really interesting, then you can use Sol. And that way, like you keep the costs down or you keep the token usage down without things getting expensive. Now you also might say, okay, the cheap tier, for example, Terra or Luna, it's not going to create something nice, but I've shown you three different tests and on two of them, the outputs were very similar.

You might also say, I should just wait and use whatever's in ChatGPT. I would switch the models depending on what you want. And I also don't think the best way to get the most out of AI is using ChatGPT. So if you just go back and forth inside the chat of ChatGPT here, you're not going to get anywhere near the level of output that you would get from something like this, right?

So we've got the AgentOS here. And if you want to use GPT 5.6, I wouldn't use ChatGPT. I'd go straight into Codex. I would create some amazing stuff like this.

Everything's saved inside my workspace. We can check it out anytime. Easy to use, easy to navigate between everything. And it's all there ready to go inside one place.

And then if we want to switch, we can go to Hermes. We can set up a profile with GPT 5.6. It's ready over there. And everything's linked inside my memory system, my memory galaxy.

That is a way to get maybe 10 times more out of your AI, whatever AI you use. Whereas for example, if you're just going back and forth inside the chat or using Codex, I don't think you're going to get that much out of this. So Luna for volume. Terra as a default.

Solve for the hard 20%. Longer gentic sessions. You can actually run Solve for hours or days, and especially with the forward slash goal feature and get a lot out of it. So if you go over to Codex over here, and then you go to the goal mode, you can run the goal mode for like literally days.

You can say, don't stop until you get the job done. And that's a really powerful way of just using autonomous agents. So yeah, really, really good stuff. All of them very impressive.

Definitely the best release that I've seen from chat cheapity this year, apart from the image update. You might also think, okay, when would you not use this stuff? I mean, if you have a choice between Fable 5 and Sol, I would still go with Fable 5 personally, but I think you can get a lot of them by just having them inside an agent operating system as well. So what have you learned today?

You've learned about the three sizes of chat cheapity. Sol, Terra, Luna. You've learned how to get the most out of each one and when to use each one, which one is the gentic leader. And also you've seen how Sol performs in terms of creations and everything else.

You might also be wondering, okay, how does it perform versus something like Fable 5? So we can actually have a look here and compare Fable 5 versus cheapity 5.6 Sol side by side. So if we have a look, let's just check the arcade game, for example. So this is the output from cheapity 5.6 Sol.

It's pretty nice. This is the output from Fable 5. It's actually very similar. I would say this just feels slightly smoother from Fable 5, but not a big difference.

We have a look at this game here. So this is the Crypt game. I like the way that, you know, the enemies and the characters appear. The controls are a little bit weird on this game, but apart from that, pretty nice.

And lots of characters, lots of interesting stuff going on there. If we have a look at Fable 5, feels a little bit more linear, but again, I would say the graphics are nice. It looks cool. It's easy to use.

One thing I will say is like, you see how the staff doesn't move when we use it? That seems a bit weird. But apart from that, it's pretty good. It's a little bit buggy there as well.

So, I mean, very, very similar outputs. Like, I can't see a big difference. This is the only one where I see a huge difference between them. Like, I would say that Soul absolutely crushed it on this example.

Like, it looks really good, super smooth. If we have a look at the version from Fable 5, we can't really turn around. So it's kind of strange. It's like, you can only really move backwards, forwards and sideways, but you can't change like the camera angle, which is kind of weird.

Pretty hard to use. Yeah. But yeah, very, very compatible. So if you want to get our systems for using all of this, you can get that inside the AI Profit Boarding.

We've got the complete agent operating system with the model routing playbooks. We have the agent OS. You get four coaching calls a week. There's 1000 prebuilt agents inside there.

You get new model briefings all the time. 4000 members across many different countries. And all of the systems I've shown you right here are inside the AI Profit Boarding. So we have a codec section here.

And if you want to use free models with codecs as well, you can choose that inside the dropdown. You can also switch to goal mode. You can see all of your sessions. You can check out your workspace here.

You can see everything you've created. You've also got, for example, the system here so that you can have Paperclip orchestrating a team of agents that could be working with Claude and GPT 5.6 and everything else. And then you can also plug in Hermes with Sol over here too. And then we additionally have a image generator, which is powered by GPT 5.6.

So if you want to get all of this, it's inside the AI Profit Boarding. Link in the comments description, or just go to theaiprofitboarding.com. This is my community for helping you save time, grow, and scale with AI automation. Inside the classroom, you can get access to all of my best trainings.

If you want the agent OS system that we've talked about today, you can get that over here. Token minimization playbook, so you can use GPT 5.6 for more without paying more. You can check out the token minimization section here. And we have loads of great guides and new tutorials all the time.

Inside the community, you can ask questions, get help and support. I personally answer all the questions inside here daily. And then inside the calendar, you can jump on weekly coaching calls, get help and support in real time, ask questions, et cetera. And inside the map, you can meet people in your local area.

So feel free to get all of this inside the AI Profit Boarding. Link in the comments description, or just go to theaiprofitboarding.com. Thanks for watching.