NEW GPT 5.6 is INSANE!

**Julian Goldie** (0:00)
So, there's three new models from GPT 5.6 today. There's a limited preview. There's a lot of controversy. And I want to walk you through everything that we know about GPT 5.6 today, what it means for you, how it works, how people can get access right now, who's got access, and also what the three new models are. So let's kick it off right now. What does this all mean? So GPT 5.6 has dropped, but it's in a limited preview. And it's only available for government-approved people who have access. I think there's about 100 people who have access right now, something like that. So it's very, very limited, and it's not open for applications or anything like that. Now, if you want to indicate when does it come out for you, I'll come on to that in a second. The second thing to note here is that there are three new models. So there's Sol, there's Terra, and there's Luna. Now, what do each of one of these do? So Sol is the frontier model, and there's actually two versions of that, which I'll come on to in a second. And then you've got GPT 5.6 Terra, which is the balance model, right? So Sol is like the super powered version, Terra is the balanced version that most people use day to day. And then you've got GPT 5.6 Luna, which is a faster, more affordable model for high volume work. So think, for example, like SubAgeants or ChatGPT Instant style process, right? So that is the setup. You've got an announcement itself, which is Sol for the hard long horizon work, Terra for everyday production, Luna for cheap, high volume jobs and three options, which I think is a lot clearer than what they've currently got, because, for example, right now they have GPT 5.5, then GPT 5.5 Instant, and then they have Pro, and it's just like, it's super messy. I think this is easier to understand. Now, here's the bad news, my friends. So that is, as exciting as that is, you cannot actually use it yet.
Read that one twice. GPT 5.6 is a limited preview. So I said about 100 before, no, it's actually about 20 partner organizations via Codex and the API only at the request of the US government. Now, apparently, and this is not confirmed because it sounds like they're just hoping, general access is in the coming weeks. So the most capable model on earth right now, and it does beat Fable 5 on benchmarks, is one that you literally cannot touch. This seems to be a trend, which is all these preview models come out, we hear about them, we see the benchmarks, we actually can't use them. And I'll come on to how you can get around that in a second and what the good news is.
But you can see, for example, OpenAI have announced, we believe in broad access and plan to make GPT 5.6 Sol, Terra and Luna generally available in the coming weeks. Now, one thing to note here is it might be limited to country by country basis. So if you look at these situations, this is interesting because it might be a case where it's only available for people in the US. You know, we're seeing that with, for example, Fable 5, it was banned for non US residents and non US nationals.
It might be the same with GPT 5.6. Who knows? Who knows where this is going? Also, it tops Terminal Bench and that's the headline number. So Sol sets a new state of the art on Terminal Bench 2.1, which is the test for real multi-step command line and agent work. So that's a genuine win, but at the same time, you can't even access it. So, you know, it's all good, it's all good fun, but you know, we can talk about this all day, but it is all theory. We don't know until we've tested it. I always like to test stuff first before I actually see the, you know, before I even think about the benchmarks because the benchmarks are totally different. So let's have a look at this.
Sol is a new flagship and a step function better than GPT 5.5. Terra delivers performance competitive to GPT 5.5 at 2x lower cost. And Luna is the most cost efficient model delivering strong capability at lower cost. So together, this is for developers and people, they say. Together the GPT 5.6 family gives people and developers. That must be cut off.
Yeah, more choice in how they balance intelligence, speed and cost.
So if we look at the benchmarks right here, let's see where we're up to.
So Sol Ultra from GPT 5.6 is scoring 91.9 on Terminal Bench 2.1. GPT 5.6 Sol is scoring 88.8. Bear in mind, there's Sol Ultra and the Sol. So there's two different versions right there. And then you have MiFos 5, which is 88%. Bear in mind, the public never got access to MiFos 5 So something to bear in mind there. It was only available for Project Glasswing. And now Fable 5 and MiFos 5 are both completely gone. GPT 5.6 Terra is at 84.3%.