We've Entered a New Era with GPT-5.6 Sol

Unless you've been living under a rock, GBT 5.6 Soul was released and it is becoming mighty close to the great Fable for a third of the price. Command AI Yeah, it seems like it's super close to Fable 5. Uh, it's a bit better than Grok 4.5. So, it's basically the second best and it's a bit strange for OpenAI to make a big announcement like we have something amazing and it's still second best. Uh, especially, you know, the underdog taking over so good that they can't even like announce an after and be better. Either [snorts] way, this is one of those, um, weeks where you say, "This week we have a new model and it's this week's best model." That's not true, not this week. Fable is still the king. Um, but yeah. Wow, I mean, we were talking about 5.6. Yeah. Yes. Sorry about that. What's the model called? But if you care about cost and cost per intelligence, especially, uh, it's a very solid model. I haven't used it myself, but from the reviews that we watched, everybody's like, "It's really good." I guess there is not too much to say about it. I think the general consensus is, what I'm hearing is that Fable is without a doubt the most intelligent model, right? It can do or it can figure out some very, very complicated, uh, tasks. However, uh, GBT 5.6 Soul, and we'll get into that, that this is more of a workhorse. This is more something that will just is the hard worker that will will get it done with brute force almost, Which is, arguably, you know, a better or worse way to get things done. So, either way, very, very good model. Um can run for days. People have been running things for days on this with by setting a goal. And yeah, I think it's gone I think gone are the days of like, you know, the these are these are new generations of models. These are new We've entered a new era of AI models, for sure. But, interesting, interesting. So, let's talk about that because I haven't used GPT-5.6, but from what I watched uh on Twitter, I've been like watching on YouTube, everybody's like Not everybody, but like from what I understand, GPT-5.6 Soul is still the same generation as the previous one, but Fable feels like it's the next gen. You literally just I mean, it's still it's well, yeah. It's not as good as Fable. Uh it's not as intelligent as Fable. But, you've you've described exactly what it is, basically. Like, the the fact that we've got a major version jump between uh Opus 4.8. I mean, to be fair, Fable didn't exist before. I guess it was Mythos, but like um it's uh four like Opus 4.8 is the same track Basically, you're talking about the training that happened to create Opus 4 or whatever. Four. It's Claude 4, let's call it, right? Yeah. And we're still on those those models are still on the minor versions of tweaking that training set data of the fourth model. Fable is a new set of training data, the fifth model. So, oh yeah, so of course, and we can't forget bloody Sonic 5 as well. That's on that's on a new architecture, let's call it. It's Yeah. Whereas, GPT 5.6, that's that's also that's on the fifth generation of GPT. It's still getting squeezing every bit of juice you can out of the fifth generation of GPT. So, yeah, you are quite right and that's why Fable does come out on top when I talk about intelligence, but it doesn't mean to say that 5.6 isn't impressive and is I mean, for a fraction of the price, you know, it's a it's a good it's it's good. This is like comparing cost to intelligence. Fable is the most expensive. Mhm. Um and GPT 5.6 has one it's like scoring at 59. Um Claude Fable 5 is at 60, but costs way more. It's more than double. Yeah, almost triple. That that's the thing like if the task is worth it to you, uh is it worth going with the the top best model? I think it's quickly becoming in a I hate to say this, but in a way a luxury thing where if you have the the money, you can get the better engineer and if you don't have the money for it, you get the one that is maybe not as good, but still can get done. Either way, we're we're moving in the direction of everything getting better, but not as fast as I personally would love to see it because I would love to see um GPT beating Fable 5 by a big margin cuz that would mean that access to better models. Like at the end of the day, I don't care who I pay the money to, if it's this one or the other company. All I want is like better and better models as a consumer, right? So, it's a bit of let down that it's not as good. Yes, it's cheaper, it's pretty good, but I I don't want it to be comparably good. I want it to be, you know, the difference between Fable how Fable 5 blew us all blew all the other models out of the water, I want it Soul to also do the same with um Fable 5. I I really want to see this because even though I I feel like everybody got a favorite and you root for, for example, GPT to not be better than your favorite Fable 5, which is stupid. Like just I'll just say it, it's stupid. Like it's in your best interest for the Chinese company tomorrow to come out with something that is better than Fable 5 because competition is good for us consumers at the end of the day paying uh to these token dealers. Wow. Saying that, so they've released three new models. They've kind of copied Anthropic here. Whereas it used to be GPT 5.5 and then you just choose the thinking level. Well, actually, you've got GPT 5.5. You had nano, mini, all that kind of stuff. I think they've removed that and now it's Terra, Luna, Terra, and Soul. So, moon earth and sun. The moon is the smallest one. Luna is the smallest model. That's your haiku. That's for haiku. Terra is your middleweight. That's your sonnet. That's the the medium-sized model and then Soul is the big boy, right? So, that's how they've broken it up. And the reason why I say about the Chinese model is if they actually do compare Luna, which is the smallest one, um matches or exceeds the intelligence of GLM 5.2. Imagine being a Chinese lab being compared to I know it's the smallest model. I know that. But even still, that's pretty impressive, isn't it really, that GLM 5.2 is up there. So, I am I am pretty stoked on that model. Um it's just about working it getting it all in your workflow and making it into something like when do you send when do you send things off to GLM 5.2, do you know? I I'm I'm not on that level quite yet, but I can imagine setups where you you have something very expensive like Fable 5 that is really good at giving prompts to other sub-agents and then sending those sub creating those sub other sub-agents not with Anthropic models, but let's say with um GLM. That Can you Can you do that actually? You would use um open code, probably. So, you can use your Claude Max subscription to go nicely. But you can set up sub-agents to use different models and whatever. If you're inside a Claude code, what you would need to do is get the get Fable to write up a detailed implementation plan in a markdown or HTML format and then you would get you know, you'd get GP GLM 5.2 to to work with that. If you're using APIs, you could build your own tool theoretically where Yeah. you you create this harness yourself where um on the top, you have the the best model and then you allow it to create sub-agents. Uh and you have like strict rules for them. That That's actually really interesting. Um that's an interesting approach. Maybe somebody with uh much more money can can do this or people are probably already doing this. Yeah. I mean like someone asked for a video on this and then it might be something that I put together, but like Um yeah, it's totally something you can do. I would probably go down the cheap ass route and be like coding plan. Just have Claude code open in one window and open code open with the other one with GLM and just have them on pointed to the same directory. So I'll just do it like that sort of thing, but Uh yeah, you could totally build your own harness to to be able to again, you can define sub agents in open code and then tell it what model that sub agent should use. So Yeah. Yeah. Yeah. Once upon a time you could have Claude code inside of open code. This is before all of this nonsense with the programmatic usage came to play. You could literally run Claude code max subscription alongside inside of Um open code. However, open code do have their own coding plan, so that could also be an option. Yeah. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI