This content from an AI expert discusses a strategic approach to leveraging large language models by combining high-performance frontier models like GPT-6 Astra and Claude Fable 5.1 with more cost-efficient alternatives such as OpenAI's Luna and Terra. It emphasizes selecting models based on task complexity and budget, and introduces tools like ClaudeX Loop for adversarial evaluation and ClaudeX Route for model selection to optimize outcomes and minimize costs.
0:00 I'm going to teach you how to combine the powers of GPT-6 Astra and Claude Fable 5.1 because the question you should be asking is not which of these two models is better, it's how can we get the most out of both of them? Furthermore, how can we get the most out of all the models under the Open AI and Anthropic umbrella because often times we are working on features and problems that don't even require the power of Astra or require the power of Fable. What about things like Luna and Tara? When do we bring those into the fold when we need cheap models to execute simple tasks even if we're working inside of something like Claude Code? Well, luckily this is a riddle that is pretty simple to solve and I'm going to walk you through it in today's video. Now, I will be giving you a set of skills that make this entire process super easy to execute. I have an updated version of ClaudeX Loop which includes Astra and a brand new ClaudeX Route skill which makes it really easy to figure out which of these models I should be using. But before we go into the skill section, let's kind of talk about why we even need this in the first place. Now, at this point we know there's kind of two big players when it comes to the best frontier model in the game. We have GPT-6 Astra and Claude Fable 5.1. When we are coming up with huge tasks, we're planning something from scratch, it's a big project and we want to know who's going to come up with it, these are the two models we are going to lean on. But another area that doesn't get talked about as much is what do we do when we have some sort of simpler feature or simpler task that just doesn't require Fable 5.1, just doesn't require Astra because these are extremely expensive models. They both cost the same, we're looking at $10 per input and $50 per output. So, if you're someone who's on like API pricing for example, this can be prohibitive.
1:36 On the other end of the spectrum, we have models like Claude Sonnet 5 as well as Tara and Luna. If you're someone who has used Claude Sonnet 5 lately, I think you can agree that this model leaves something to be desired especially when we compare it to the supposed benchmarks and the performance that we've seen. Like it just really doesn't feel great. And so if you're someone who has been using primarily Anthropic models over the last few months, there is this gap where if I want a model that's going to be cheap and effective, I just can't really find it in the Anthropic library. But, if we look at OpenAI, they have multiple models that satisfy this requirement when it comes to Terra and Luna. Now, Terra is slightly more expensive than Sonnet 5 when we talk about token to token, but as I'll show you in a minute, it's cheaper overall because it's more token efficient. And what really is a standout is Luna. This is like what Haiku should be if they hadn't completely abandoned Haiku and just left it in the dumpster months ago. We're looking at 20 cents per input and cash is 2 cents and output is $1.20. Like, this is basically free compared to everything else. And like I'm going to show you, this has outputs and this has performance that matches or exceeds Sonnet 5. So, this idea of being able to call on OpenAI models is really, really important if you are someone who's been mostly an Anthropic user.
3:03 Especially if you use APIs. And even if you're someone who's on, say, the 20X plan with Anthropic and isn't on any plan with OpenAI, well, maybe it makes sense to just have sort of a $20 a month plan specifically to make these sort of Luna calls and save on tokens. Let me quickly illustrate this idea. We're taking a look at the Deep Sweet benchmark. We have Sonnet here in the orange and then Terra and Luna. So, straight up in terms of performance, Terra and Luna perform better than Sonnet. At max, Sonnet is hitting 54% max on Terra is 70% max on Luna is 67%. But, what it really what really is the difference though is the cost. You know, when we look at max, it's Sonnet, $26 for average cost per task versus Luna, 60 cents. So, and And for Terra, $4. Now, this becomes a little less pronounced as we go down sort of the effort level. But even on low, you know, $2.19 for Sonnet versus 1 cent for Luna. Granted, it's not doing anything for you. But, with Tara, 34 cents. And so, there is this issue with the Anthropic models where there is no great bang for your buck model, and Sonnet is just like always at risk of just going nuts in terms of the cost, and then just like taking way too many tokens to complete these things. We just don't see that with Tara and Luna. In fact, with Luna in particular, going from low all the way to max barely increases the cost and just like completely spikes in terms of its effectiveness. So, again, real value to be had here in being able to call down these models. Now, the second reason you want to be able to combine the power of Open AI and Anthropic is something I've talked about at length in some of my previous videos dealing with Claude's Loop. And that is the idea that if I have one model execute, I should have a different model take a look and evaluate its work. So, for example, if I have Claude Fable 5.1 come up with a plan for some sort of project, then I probably want Astra to come in there with completely blank context without any of the idiosyncrasies that all the Anthropic models have, and I want it to take a look at that plan Fable created and say, "All right, this is good, this is bad, here's what we should fix."
5:10 And ideally, and this is what the Claude's Loop skill does, it gives that feedback back to Fable 5.1, and then Fable says, "Mhm, agree, disagree." Sends it back to Astra. Astra says, "Agree, disagree." And you have this continuous cycle, this loop, of these models going back and forth adversarially till they finally reach a point where they're like, "Okay, this is good to go." Now, there's safeguards in place to make sure that this doesn't go on and on forever and burn all your tokens, but even though on the surface that might seem kind of expensive, it's going to save you tokens in the long run versus having to iterate after we've already spent a bunch of time building things. And this is something we can extend to a lower level, right? We could have Opus build something and have Terra take a look or have Luna take a look, so on and so forth. The idea is we never want the model that builds to be the model that evaluates because frankly, it's just not going to do a great job and these models in general tend to grade their own work poorly. I mean, in like in a very favorable light. Fable's always going to think Fable's work is great. Astro's always going to think Astro's work is great, so on and so forth. So, that's why we're doing it and now we'll jump into the skill, but first a quick word from today's sponsor, me. So, inside of Chase AI Plus, I have just released both a Claude code and a Codex Masterclass. So, if you're someone who's trying to figure out how to use either of these tools and you don't come from a technical background, this is the perfect place for you. I update this every single week. We focus on real examples. So, if this sounds like something you really want to master, definitely check us out there. There will be a link in the pinned comment.
6:42 Hope to see you there. So, all the skills you need are found in this GitHub repo, the Claudex Loop GitHub repo. I'll put a link to it in the pinned comment. Now, the two big ones we're going to talk about today is Claudex Loop, which I've done videos on the past on, and the new Claudex Route. So, let's talk about Claudex Route first cuz this is the new one. Big picture, if you run Claudex Route and you can run this inside of Claude code or inside of Codex and you give it a brief description of what you're trying to do like, "Hey, I'm trying to do this particular task. Can you use Claudex Route to figure out what model I should use?" Well, the skill isn't going to tell you "Here's the model that makes the most sense for you." Maybe it's a very simple so it's going to use Luna. Or maybe it's something that, "Hey, requires a lot of ambiguous work. It's pretty difficult. Let's use Astra." So, it takes away any sort of thought you need to do in terms of what model to use because sometimes it's kind of like a gray area and this is basing it off the actual model information that has been posted for all the models by the Frontier Labs including some of the usage guidelines. And it takes doing into account like the actual cost. And what that looks like in reality is let's say I'm inside of Claude Code and I do {slash} Claud X route and I say, "Hey, I'm starting this new project. I think it's going to be pretty complicated.
7:49 What model should I use?" And it says, "Go use Astra." Well, it's just going to send all the relevant information to a headless instance of Codex. So, think of it pulling up Codex in the CLI, but it's invisible. Codex is going to do its thing, do all of its work, and then it will send it all the information back to Claude Code saying, "Here's what I did. Here's where that sort of code is." Now, this also works in the reverse. Like I said, you can start this from Claude Code or Codex. Doesn't really matter. Now, the next skill you need to pay attention to is Claud X Loop. I've done full videos on Claud X Loop taking you from the very beginning to the end. So, I will link that above if you want to watch that. For now, I'll kind of just go over some of the changes I've made to it and what it sort of does at a high level. So, if Claud X route is for like one-off features like, "Hey, what model should I use?" Go send it to do it. Claud X Loop is sort of for really big implementations that walks you through that kind of holds your hand from planning stage all the way to execution. And it's all about that idea I talked about earlier of hey, if model A executes, then model B checks its work.
8:51 So, first things first, if I'm inside of Claude Code or inside of Codex, it's going to use the opposite model to sort of be the inspector. So, hey, if I'm inside of Codex, right? Who's going to be doing the plan? Well, it's going to be Astra. Who's going to take a look at their plan? It's going to be Claude Code and vice versa depending on what platform you're on. From there, it goes through four stages. So, first things first, it's going to do some reconnaissance, which just means it's going to do deep research. And let's say you're saying, "Hey, I want to clone something like WhisperFlow." Well, it's going to spawn a bunch of sub agents, figure out what WhisperFlow actually is, what we need to think about, has anyone already cloned this, and like really sort of validate your assumptions. From there, it's going to ask you a few questions. Be like, "Okay, what do you want to change about it? What do you care about? What is your vision for this project? Before finally it executes the plan. Well, it doesn't execute the plan, it builds out the plan. So, Astra builds out the plan, sends it to Claude code. They have their back and forth for several rounds until they finally reach an approved verdict. From there, we kind of repeat that process. And it's going to ask you if you want Astra or Fable to actually execute it.
9:58 One of them executes the plan, and then the opposite one takes a look at the executed plan and says, "Hey, here's what you missed. Here's, you know, well, looks good." So, at the end of the day, the Claudex loop is really just meant to have these opposite models sort of check for blind spots that you're not going to be able to figure out. Because let's be honest, for many of us, what we're building inside of Claude code and Codex goes well beyond the scope of our coding abilities. So, are you going to double-check what Codex comes up with and Astra comes up with? Are you going to double-check what Fable comes up with? Probably not. But, it will give you something of a like a warm and fuzzy feeling knowing that Astra looked at Fable's work, or Fable looked at Astra's work. And that is sort of the whole point of the Claudex loop. And I think it's the easiest way to kind of get the best out of both of these models without some like super convoluted setup. So, if you want to get this working for yourself, just go ahead, copy the URL here, point Claude code at it, point Claude Codex at it, and it will install for you. So, that's where I'm going to leave you guys. Remember, if you want to see the Claudex loop actually executed from beginning to end, check out that video I linked earlier. I go in depth showing here's what GPT found, here's all the errors, here's what it looks like when they have their back and forth, and here's sort of what the final setup sort of looks like. But, besides that, I think we're in such an awesome place where we have these dueling models.
11:17 Like, Fable 5.1 is great. GPT-6 Astra is also great. So, I don't really think there should be an either and thing. And I think the sort of pricing of all these can can kind of scare you off. Like, should I be on a 20x here and a 20x there? Should I be paying 400 bucks a month? I don't think so. I think for a lot of people who have kind of been, you know, Claude-filled the last 6 9 12 months, you've been probably on a $200 plan. I would suggest going half and half. Do the 5x with GPT, do the 5x with Anthropic, and really see what you like. You know, maybe you like using them in tandem, maybe you love GPT-6. I think in general though, when it comes to all these AI tools and how fast everything is moving, we need to be tool agnostic. And it's really hard to be tool agnostic if you don't play around with all of them. So, that's my suggestion. Definitely check out the skill. Let me know what you think. Let me know if you have any improvements for it. Always trying to keep it updated. And besides that, I'll see you around.