This SIMPLE Hack Combines GPT 6 Astra With Fable 5.1

English
CChase AI
Computing/SoftwareInternet Technology

Transcript

00:00:00I'm going to teach you how to combine the powers of GPT-6 Astra and Claude Fable 5.1
00:00:05because the question you should be asking is not which of these two models is better,
00:00:09it's how can we get the most out of both of them.
00:00:12Furthermore, how can we get the most out of all the models under the open AI and anthropic umbrella?
00:00:17Because oftentimes, we are working on features and problems that don't even require the power of Astra
00:00:22or require the power of Fable.
00:00:24What about things like Luna and Terra?
00:00:26When do we bring those into the fold when we need cheap models to execute simple tasks,
00:00:31even if we're working inside of something like Claude Code?
00:00:33Well, luckily, this is a riddle that is pretty simple to solve,
00:00:36and I'm going to walk you through it in today's video.
00:00:38Now, I will be giving you a set of skills that make this entire process super easy to execute.
00:00:44I have an updated version of Claudex Loop, which includes Astra,
00:00:47and a brand new Claudex Route skill,
00:00:49which makes it really easy to figure out which of these models I should be using.
00:00:52But before we go into the skills section, let's kind of talk about why we even need this in the first place.
00:00:58Now, at this point, we know there's kind of two big players when it comes to the best frontier model in the game.
00:01:02We have GPT-6 Astra and Claude Fable 5.1.
00:01:05We are coming up with huge tasks.
00:01:08We're planning something from scratch.
00:01:09It's a big project, and we want to know who's going to come up with it.
00:01:12These are the two models we are going to lean on.
00:01:15But 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
00:01:22that just doesn't require Fable 5.1, just doesn't require Astra,
00:01:25because these are extremely expensive models.
00:01:27They both cost the same.
00:01:28We're looking at $10 per input and $50 per output.
00:01:31So if you're someone who's on like API pricing, for example, this can be prohibitive.
00:01:36On the other end of the spectrum, we have models like Claude Sonnet 5, as well as Terra and Luna.
00:01:43If you're someone who has used Claude Sonnet 5 lately,
00:01:46I think you can agree that this model leaves something to be desired,
00:01:50especially when we compare it to the supposed benchmarks and the performance that we've seen.
00:01:54Like, it just really doesn't feel great.
00:01:56And so if you're someone who has been using primarily Anthropic models over the last few months,
00:02:01there is this gap where if I want a model that's going to be cheap and effective,
00:02:06I just can't really find it in the Anthropic library.
00:02:10But if we look at OpenAI, they have multiple models that satisfy this requirement
00:02:16when it comes to Terra and Luna.
00:02:18Now, Terra is slightly more expensive than Sonnet 5 when we talk about token to token,
00:02:24but as I'll show you in a minute, it's cheaper overall because it's more token efficient.
00:02:28And what really is a standout is Luna.
00:02:32This is like what Haiku should be if they hadn't completely abandoned Haiku
00:02:35and just left it in the dumpster months ago.
00:02:37We're looking at $0.20 per input and the cash is $0.02 and output is $1.20.
00:02:44Like, this is basically free compared to everything else.
00:02:48And like I'm going to show you, this has outputs and this has performance
00:02:51that matches or exceeds Sonnet 5.
00:02:54So this idea of being able to call on OpenAI models is really, really important
00:02:59if you are someone who's been mostly an Anthropic user, especially if you use APIs.
00:03:04And even if you're someone who's on, say, the 20x plan with Anthropic
00:03:08and isn't on any plan with OpenAI,
00:03:11well, maybe it makes sense to just have sort of a $20 a month plan
00:03:13specifically to make these sort of Luna calls and save on tokens.
00:03:17Now to quickly illustrate this idea, we're taking a look at the DeepSuite benchmark.
00:03:20We have Sonnet here in the orange and then Terra and Luna.
00:03:24So straight up in terms of performance, Terra and Luna perform better than Sonnet.
00:03:30At max, Sonnet is hitting 54%.
00:03:33Max on Terra is 70%.
00:03:35Max on Luna is 67%.
00:03:37But what really is the difference, though, is the cost.
00:03:41You know, when we look at max at Sonnet, $26 for average cost per task versus Luna, 60 cents.
00:03:48So, and then for Terra, $4.
00:03:52Now this becomes a little less pronounced as we go down sort of the effort level.
00:03:56But even on low, you know, $2.19 for Sonnet versus one cent for Luna,
00:04:02granted it's not doing anything for you.
00:04:04But with Terra, 34 cents.
00:04:07And so there is this issue with the Anthropic models where there is no great bang for your buck model.
00:04:13And Sonnet is just like always at risk of just going nuts in terms of the cost
00:04:18and then just like taking way too many tokens to complete these things.
00:04:21We just don't see that with Terra and Luna.
00:04:23In fact, with Luna in particular, going from low all the way to max barely increases the cost
00:04:27and just like completely spikes in terms of its effectiveness.
00:04:30So again, real value to be had here and being able to call down these models.
00:04:34Now the second reason you want to be able to combine the power of open AI and Anthropic
00:04:37is something I've talked about at length in some of my previous videos dealing with CloudX Loop.
00:04:42And that is the idea that if I have one model execute, I should have a different model take a look and evaluate its work.
00:04:49So for example, if I have Claude Fable 5.1 come up with a plan for some sort of project,
00:04:56then I probably want Astra to come in there with completely blank context without any of the idiosyncrasies
00:05:02that all the Anthropic models have.
00:05:04And I want it to take a look at that plan, Fable created and say, all right, this is good.
00:05:08This is bad.
00:05:09Here's what we should fix.
00:05:10And ideally, and this is what the CloudX Loop skill does.
00:05:14It gives that feedback back to Fable 5.1.
00:05:17And then Fable says, mm, agree, disagree, sets it back to Astra.
00:05:21Astra says, agree, disagree.
00:05:22And you have this continuous cycle, this loop of these models going back and forth adversarily
00:05:27till they finally reach a point where like, okay, this is good to go.
00:05:31Now there's safeguards in place to make sure that it just doesn't go on and on forever and burn all your tokens.
00:05:36But even though on the surface that might seem kind of expensive, it's going to save you tokens in the long run
00:05:41versus having to iterate after we've already spent a bunch of time building things.
00:05:46And this is something we can extend to a lower level, right?
00:05:48We could have Opus build something and have Terra take a look or have Luna take a look, so on and so forth.
00:05:53The idea is we never want the model that builds to be the model that evaluates
00:05:58because frankly, it's just not going to do a great job.
00:06:00And these models in general tend to grade their own work poorly.
00:06:05I mean, in a very favorable light.
00:06:07Fable's always going to think Fable's work is great.
00:06:09Astra's always going to think Astra's work is great, so on and so forth.
00:06:13So that's why we're doing it.
00:06:14And now we'll jump into the skill.
00:06:15But first, a quick word from today's sponsor, me.
00:06:19So inside of Chase AI+, I have just released both a Claude code and a Codex masterclass.
00:06:24So if you're someone who's trying to figure out how to use either of these tools
00:06:28and you don't come from a technical background, this is the perfect place for you.
00:06:32I update this every single week.
00:06:33We focus on real examples.
00:06:36So if this sounds like something you really want to master, definitely check us out there.
00:06:40It will be a link in the pinned comment.
00:06:42Hope to see you there.
00:06:43So all the skills you need are found in this GitHub repo, the Claudex loop GitHub repo.
00:06:47I will put a link to it in the pinned comment.
00:06:49Now, the two big ones we're going to talk about today is Claudex loop,
00:06:52which I've done videos on the past on, and the new Claudex route.
00:06:56So let's talk about Claudex route first because this is the new one.
00:06:58Big picture, if you run Claudex route, and you can run this inside of Claude code or inside
00:07:03of Codex, and you give it a brief description of what you're trying to do, like, hey, I'm
00:07:07trying to do this particular task.
00:07:09Can you use Claudex route to figure out what model I should use?
00:07:11Well, the skill isn't going to tell you, here's the model that makes the most sense for you.
00:07:15Maybe it's a very simple task, so it's going to use Luna.
00:07:18Or maybe it's something that, hey, requires a lot of ambiguous work.
00:07:21It's pretty difficult.
00:07:22Let's use Astra.
00:07:23So it takes away any sort of thought you need to do in terms of what model to use, because
00:07:27sometimes it's kind of like a gray area.
00:07:29And this is basing it off the actual model information that has been posted for all the
00:07:34models by the Frontier Labs, including some of the usage guidelines.
00:07:37And it takes into account, like, the actual cost.
00:07:40And what that looks like in reality is, let's say I'm inside of Claude code, and I do slash
00:07:44Claudex route, and I say, hey, I'm starting this new project.
00:07:48I think it's going to be pretty complicated.
00:07:49What model should I use?
00:07:50And it says, go use Astra.
00:07:52Well, it's just going to send all the relevant information to a headless instance of Codex.
00:07:58So think of it pulling up Codex in the CLI, but it's invisible.
00:08:01Codex is going to do its thing, do all of its work, and then it will send all the information
00:08:05back to Claude code saying, here's what I did.
00:08:07Here's where that sort of code is.
00:08:09Now, this also works in the reverse.
00:08:11Like I said, you can start this from Claude code or Codex, doesn't really matter.
00:08:15Now, the next skill you need to pay attention to is Claudex loop.
00:08:18I've done full videos on Claudex loop, taking you from the very beginning to the end.
00:08:22So I will link that above if you want to watch that.
00:08:24For now, I'll kind of just go over some of the changes I've made to it and what it sort
00:08:28of does at a high level.
00:08:30So if Claudex route is for like one-off features, like, hey, what model should I use?
00:08:33Go send them to do it.
00:08:35Claudex loop is sort of for really big implementations.
00:08:38That walks you through, that kind of holds your hand from planning stage all the way to execution.
00:08:43And it's all about that idea I talked about earlier of, hey, if model A executes, then
00:08:49model B checks its work.
00:08:51So first things first, if I'm inside of Claude code or inside of Codex, it's going to use
00:08:57the opposite model to sort of be the inspector.
00:09:00So, hey, if I'm inside of Codex, right?
00:09:03Who's going to be doing the plan?
00:09:05Well, it's going to be Astra.
00:09:05Who's going to take a look at their plan?
00:09:07It's going to be Claude code and vice versa, depending on what platform you're on.
00:09:11From there, it goes through four stages.
00:09:13So first things first, it's going to do some reconnaissance, which just means it's going
00:09:17to do deep research.
00:09:18And let's say you're saying, hey, I want to clone something like Whisperflow.
00:09:22Well, it's going to spawn a bunch of subagents, figure out what Whisperflow actually is, what
00:09:26we need to think about.
00:09:27Has anyone already cloned this?
00:09:29And like really sort of validate your assumptions.
00:09:31From there, it's going to ask you a few questions.
00:09:34Be like, okay, what do you want to change about it?
00:09:36What do you care about?
00:09:37What is your vision for this project?
00:09:39Before finally, it executes the plan.
00:09:41Well, it doesn't execute the plan.
00:09:42It builds out the plan.
00:09:43So Astra builds out the plan, sends it to Claude code.
00:09:46They have their back and forth for several rounds until they finally reach an approved verdict.
00:09:51From there, we kind of repeat that process.
00:09:53And it's going to ask you if you want Astra or Fable to actually execute it.
00:09:58One of them executes the plan.
00:09:59And then the opposite one takes a look at the executed plan and says, hey, here's what you
00:10:04missed.
00:10:04Here's what looks good.
00:10:06So at the end of the day, the Claudex loop is really just meant to have these opposite
00:10:12models sort of check for blind spots that you're not going to be able to figure out.
00:10:15Because let's be honest, for many of us, what we're building inside of Claude code and Codex
00:10:20goes well beyond the scope of our coding abilities.
00:10:24So are you going to double check what Codex comes up with and Astra comes up with?
00:10:29Are you going to double check what Fable comes up with?
00:10:31Probably not.
00:10:32But it will give you something of like a warm and fuzzy feeling.
00:10:35Knowing that Astra looked at Fable's work or Fable looked at Astra's work.
00:10:40And that is sort of the whole point of the Claudex loop.
00:10:42I think it's the easiest way to kind of get the best out of both of these models without
00:10:46some like super convoluted setup.
00:10:49So if you want to get this working for yourself, just go ahead, copy the URL here, point Claude
00:10:55code at it, point Codex at it, and it will install for you.
00:10:58So that's where I'm going to leave you guys.
00:10:59Remember, if you want to see the Claudex loop actually executed from beginning to end,
00:11:03check out that video I linked earlier.
00:11:04I go into depth showing here's what GPT found.
00:11:07Here's all the errors.
00:11:08Here's what it looks like when they have their back and forth.
00:11:10And here's sort of what the final setup sort of looks like.
00:11:12But besides that, I think we're in such an awesome place where we have these dueling models.
00:11:17Like Fable 5.1 is great.
00:11:19GPT-6 Astra is also great.
00:11:21So I don't really think there should be an either-and thing.
00:11:24And I think the sort of pricing of all these can kind of scare you off.
00:11:28Like, should I be on a 20X here and a 20X there?
00:11:30Should I be paying 400 bucks a month?
00:11:32I don't think so.
00:11:33I think for a lot of people who have kind of been, you know, Claude-pilled for the last
00:11:376, 9, 12 months, you've been probably on a $200 plan.
00:11:41I would suggest going half and half.
00:11:42Do the 5X with GPT, do the 5X with Anthropic, and really see what you like.
00:11:48You know, maybe you like using them in tandem.
00:11:50Maybe you love GPT-6.
00:11:51I think in general, though, when it comes to all these AI tools and how fast everything
00:11:55is moving, we need to be tool agnostic.
00:11:58And it's really hard to be tool agnostic if you don't play around with all of them.
00:12:01So that's my suggestion.
00:12:03Definitely check out the skill.
00:12:04Let me know what you think.
00:12:05Let me know if you have any improvements for it.
00:12:07Always trying to keep it updated.
00:12:08And besides that, I'll see you around.

Key Takeaway

Combining GPT-6 Astra and Claude Fable 5.1 through the Claudex loop and route skills cuts token costs while leveraging adversarial cross-evaluation between frontier models.

Highlights

  • Frontier models GPT-6 Astra and Claude Fable 5.1 both cost $10 per input and $50 per output.

  • OpenAI models Terra and Luna offer lower costs and higher token efficiency compared to Claude Sonnet 5.

  • Luna costs $0.20 per input and $1.20 per output while matching or exceeding Sonnet 5 performance.

  • The Claudex loop skill runs an adversarial process where one model builds and the opposite model evaluates the work.

  • The Claudex route skill determines the optimal model for a specific task based on frontier lab benchmarks and pricing guidelines.

Timeline

Model Comparison and Cost Analysis

  • GPT-6 Astra and Claude Fable 5.1 serve as the top frontier models for large planning tasks but carry high API costs.
  • Claude Sonnet 5 underperforms on benchmarks relative to its token consumption.
  • OpenAI models Terra and Luna provide efficient alternatives to Anthropic options for budget-conscious workflows.

Fable 5.1 and Astra require $10 per input and $50 per output, making them expensive for routine tasks. Sonnet 5 struggles with token efficiency and performance compared to Terra and Luna. Luna delivers high effectiveness at $0.20 per input and $1.20 per output, matching Sonnet 5 performance on DeepSuite benchmarks.

Adversarial Model Evaluation

  • A separate model inspects and evaluates the work produced by the primary building model.
  • Models rate their own outputs too favorably without external adversarial review.
  • Cross-model evaluation prevents blind spots and reduces long-term iteration costs.

Fable-generated plans face evaluation by Astra without shared context or Anthropic idiosyncrasies. This back-and-forth loop continues iteratively until both models reach an approved verdict, safeguarding project quality beyond a single model's capability.

Implementation via Claudex Skills

  • The Claudex route skill automatically selects the appropriate model for a task based on complexity and cost.
  • Claudex route operates via a headless instance of Codex within Claude Code or CLI environments.
  • Tool selection relies on official usage guidelines and pricing data published by frontier labs.

Running Claudex route with a task description eliminates guesswork regarding model selection. The skill analyzes whether a task requires a lightweight model like Luna or a heavy frontier model like Astra before executing the workflow.

Workflow Execution with Claudex Loop

  • Claudex loop guides large implementations from the reconnaissance stage to final execution.
  • Reconnaissance spawns subagents to research project requirements and validate assumptions.
  • Tool agnosticism and alternating model usage maximize efficiency across different AI ecosystems.

Claudex loop handles complex tasks by cycling through reconnaissance, planning, building, and evaluation stages between opposing models. Users benefit from splitting subscriptions across ecosystems rather than paying for a single high-tier plan.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video