Big Idea

The argument in one line.

Claude Code's harness and its underlying model are separable: a local proxy plus a planner/executor skill let Claude Fable orchestrate GPT-5.6 Sol for the token-heavy work, cutting real build costs by roughly a quarter to a half without losing any custom skills.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code daily and hit weekly usage limits on Fable or Opus more often than you'd like.
You already pay for a ChatGPT subscription and want those tokens working for you instead of sitting unused.
You've built up custom skills, slash-commands, or a 'second brain' inside Claude Code and don't want to rebuild it in another harness to save money.
You're comfortable installing an open-source local proxy and editing model aliases from a terminal.

SKIP IF…

You're not hitting usage limits and see no reason to manage a second subscription plus a proxy just to save money.
You want a zero-setup fix — this requires installing CLIProxyAPI, logging into a second account, and configuring model aliases.

TL;DR

The full version, fast.

Anthropic's Claude models are currently the most expensive per-task of any frontier lineup, while OpenAI's GPT-5.6 Sol has closed most of the intelligence gap at roughly half the cost and with more generous usage limits. Rather than abandon Claude Code's skills and commands to switch harnesses, an open-source local proxy (CLIProxyAPI) lets Claude Code route requests to Sol under an OpenAI subscription while keeping the same interface. A companion skill, FabSol, applies an advisor/executor pattern: Claude Fable plans and reviews, GPT-5.6 Sol does the actual build work, and each model draws from its own separate usage pool. Across repeated test builds, the FabSol combo cost 27-43% less than running Fable alone, with comparable output quality. The larger lesson: don't lock the harness and the model into the same vendor decision — staying model-agnostic is itself a cost-saving strategy as providers keep undercutting each other.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:57

01 · Claude's problem

Anthropic's Claude models are expensive and burn through usage limits fast, just as OpenAI's GPT-5.6 lands within striking distance on benchmarks at roughly half the cost.

00:57 – 03:13

02 · GPT-5.6 Sol vs Fable: benchmarks and cost per task

Independent benchmarks from Artificial Analysis show GPT-5.6 Sol closing in on Claude Fable 5's intelligence score while costing less than half as much per task — and Anthropic's models are the priciest on the chart.

03:13 – 04:04

03 · The dilemma

OpenAI's usage limits stretch further than Anthropic's at a similar price and come with extra manual resets, but switching harnesses away from Claude Code means abandoning a second brain of skills and commands.

04:04 – 06:37

04 · Sol inside Claude Code

A local proxy lets Claude Code send requests to GPT-5.6 Sol under the OpenAI subscription while keeping the same skills and slash-commands — retrieving a skill took about 12 seconds in Claude Code versus 48 seconds when the same workspace was pointed at Codex directly.

06:37 – 08:22

05 · Under the hood

The wiring runs through CLIProxyAPI, an open-source local proxy with 40,000+ GitHub stars that logs into the OpenAI account and exposes it as an endpoint Claude Code can target with a model flag.

08:22 – 09:34

06 · The FabSol skill: Fable plans, Sol executes

Modeled on Anthropic's own advisor/executor pattern, the FabSol skill has Claude Fable plan and review while GPT-5.6 Sol does the actual build work, splitting cost across two separate usage pools.

09:34 – 11:13

07 · Side-by-side test

Across seven repeated builds — an ROI calculator, client dashboard, order tracker, and more — the FabSol combo consistently cost less than Fable working alone, with savings of roughly 27-43% on a comparable landing-page build.

11:13 – 12:33

08 · Why being model agnostic is the move

The closing argument: don't lock into one provider — Claude Code can orchestrate GPT-5.6 Sol, Sonnet, open-source models, or whatever comes next, and provider competition is what keeps usage limits generous.

Atomic Insights

Lines worth screenshotting.

Anthropic's own models are the most expensive per-task of any frontier lineup on independent benchmarks, which is why heavy Claude Code users hit usage limits fastest.
GPT-5.6 Sol scores close to Claude Fable 5 on independent intelligence benchmarks while costing less than half as much per task.
A coding harness and the model running underneath it are separable decisions — you can keep Claude Code's skills and commands while routing execution to a different provider's model.
Retrieving the same custom skill took about 12 seconds in a familiar harness versus 48 seconds when the identical workspace was pointed at an unfamiliar one — the harness itself, not just the model, drives speed on long-horizon tasks.
CLIProxyAPI is an open-source local proxy with 40,000+ GitHub stars that lets one CLI coding tool route requests to multiple providers' model subscriptions.
An advisor/executor split — an expensive model plans and reviews, a cheaper model executes — is a pattern Anthropic's own developer team has published gets roughly 92% of the expensive model's solo score at about 63% of the price.
Splitting execution across two providers' subscriptions doubles your effective usage ceiling, since each model draws against its own separate limit pool instead of sharing one.
Across seven repeated test builds (an ROI calculator, a client dashboard, an order tracker, and more), a planner/executor combo cost 27% to 43% less than the same expensive model doing every step alone.
On a real landing-page build test, the two-model combo produced output at least as polished as the single-model build, at a lower cost.
OpenAI has been offering manual usage-limit resets on top of its weekly reset cadence, which can meaningfully stretch how far a subscription goes if you're not already using it.
Locking your coding harness and your model provider into the same vendor means you inherit that vendor's pricing and limits with no leverage to route around price increases.

Takeaway

Claude Code doesn't require Claude — the harness and the model are separable.

MODEL ROUTING

A local proxy plus a two-step planner/executor skill let Claude Code delegate execution to a cheaper model while keeping every skill and command you've already built, cutting real build costs by roughly a quarter to a half.

01Claude's problem

Claude Code defaults to routing every request through Anthropic's own models, which are currently the most expensive per-task of any frontier lineup, so heavy users burn through weekly limits fastest on the harness they're most invested in.
A rival model closing the intelligence gap while costing half as much is a structural pressure point worth tracking, not just a one-off promo.

02GPT-5.6 Sol vs Fable: benchmarks and cost per task

Independent benchmark sites, not vendor marketing pages, are the more trustworthy place to compare frontier models on intelligence, speed, and cost per task.
Cost per task, not sticker price per token, is the number that predicts how fast you'll hit a usage limit — look for a 'cost per task' or 'weighted average cost' metric specifically.

03The dilemma

Subscription usage limits and reset cadence differ meaningfully between providers even at similar price points — check reset frequency and whether manual resets are offered before assuming two subscriptions are equivalent.
Switching your daily-driver harness has a real cost beyond price: every custom skill, command, and 'second brain' integration has to be rebuilt or ported, which is often reason enough to keep the harness and swap only the model underneath it.

04Sol inside Claude Code

Keeping the same harness while swapping the underlying model preserves tooling-specific speedups — the same skill retrieval took about 12 seconds in a familiar harness versus 48 seconds in an unfamiliar one pointed at the same workspace.
A model swap is worth doing even for modest workflows if it removes friction on long-horizon tasks, since the harness's native command handling compounds over many requests.

05Under the hood

Open-source local proxies exist specifically to let one coding harness route requests to multiple providers' subscriptions — check GitHub star count and recent activity as a rough trust signal before wiring your credentials through one.
A model alias or flag at launch is often all that's needed once the proxy and login are set up — the added complexity is front-loaded into a one-time setup, not ongoing friction.

06The FabSol skill: Fable plans, Sol executes

The advisor/executor split — one strong, expensive model plans and reviews while a cheaper model does the token-heavy execution — is a documented pattern worth adopting regardless of which two models you pair.
Splitting execution across two separate subscriptions doubles your effective usage ceiling, since each model draws from its own limit pool instead of sharing one.

07Side-by-side test

Run the same build twice — once with your default expensive model doing everything, once with the planner/executor split — and compare actual token cost, not estimates, before trusting a cost-saving claim.
In repeated tests across several small app builds, the split setup saved roughly a quarter to just over 40% in cost while producing comparable output quality on a real landing-page build.

08Why being model agnostic is the move

Treat your coding harness and your model choice as two separate decisions — locking into one vendor for both means you inherit that vendor's pricing and limits with no leverage.
Provider competition is what keeps usage limits generous and prices falling — staying flexible enough to route around whichever provider is currently expensive is itself a cost-saving strategy, not just a one-time hack.

Glossary

Terms worth knowing.

CLIProxyAPI: An open-source local proxy application that logs into a model provider's account and exposes it as an endpoint a coding CLI can route requests through, letting one harness bill against multiple providers' subscriptions.
FabSol skill: A custom Claude Code skill that splits a coding task between two models: Claude Fable plans and reviews the work, while a cheaper model (here, GPT-5.6 Sol) does the actual execution.
Cost per task: A benchmarking metric that measures the weighted average dollar cost of completing a standardized set of tasks on a given model, used to compare real running costs rather than just per-token pricing.
Advisor/executor pattern: A multi-model workflow where a stronger, more expensive model gives guidance and review while a cheaper model performs the bulk of the token-heavy execution work, lowering total cost with a small quality tradeoff.
Usage limit reset: The point at which a subscription's consumption cap refills, either on a fixed weekly schedule or, with some providers, as an additional manual reset a user can trigger a limited number of times.

Resources

Things they pointed at.

06:37toolCLIProxyAPI

01:21toolArtificial Analysis ↗

05:33productFabSol skill download (RoboNuggets) ↗

Quotables

Lines you could clip.

00:00

“Claude has a problem, and it's that their models are too expensive. And they eat up your usage limits like there's no tomorrow.”

states the whole thesis in one breath, no context needed→ TikTok hook↗ Tweet quote

08:21

“If you use Fable five as an adviser slash orchestrator and use a lower, cheaper model as an executor, then at the end of it you actually get something like 92% of Fable five's score at only 63% of the price.”

a specific, citable stat that carries the whole video's argument→ IG reel cold open↗ Tweet quote

10:59

“Fable is really good and is probably the most frontier amongst all of the frontier models right now, the fact of the matter is that it is also really expensive.”

concedes the counterargument before pivoting to the fix→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphorstory

00:00Cloud has a problem, and it's that their models are too expensive. And they eat up your usage limits like there's no tomorrow. Meanwhile, OpenAI now has GPT five dot six, almost at par with Fable, but at half the cost and with usage limits that go a much longer way.

00:15Well, what if I told you that you can keep using Cloud Code, keep the skills and the commands you built up over time, but link it with GPT Soul in your OpenAI subscription? Well, that's what I did. So today, I'll teach you how to set it up, how to use Fable and Soul together for cheaper but at the same power, and why being model agnostic is one of the best moves that you can do today.

00:34Let's dive into it. So Claude Code is an amazing harness. There's a reason why it's the most popular one in the market right now.

00:40But over the past week, what has become evidently clear is that it has one glaring problem, And that is the fact that with Cloud Code by default, you are obviously locked into using the Anthropic models, which have become more expensive when compared to other models that are almost at par with its competence and power. And I say it's become more evident over the past week precisely because of OpenAI's GPT five dot six model release, which if you look at their website here, you can see how much this model has been challenging Fable in terms of the benchmarks that they are publishing, where you can actually see five dot six sol, OpenAI's strongest model now, exceeding Fable in a lot of these benchmarks.

01:18And in fact, you go to artificialanalysis.ai's benchmark page, which is this independent website that does a lot of these tests and actually measures these models in a more independent fashion so that we're not looking at OpenAI's website, which is obviously a bit more biased towards their own models. You can see here that gpd5.6 is almost as good as CloudFable five.

01:39And this obviously depends on how you're using these models, but at least when I was using it, I can attest that it is pretty close. And another model that was launched just in the last week is Grok four dot five from SpaceX AI, and it is also coming in pretty close, I would say, versus g p d five dot six and CloudFable five.

01:56But the real headline that you should pay attention here is apart from just raw intelligence, if we go to the cost per task, basically, what they did here, this independent study, is they run these models on a variety of tasks, and then they just got the weighted average cost of it. And you can see how much Fable five is more expensive versus GPT five dot six here.

02:13It's more than two times the cost. And in fact, if we look at the breakdown here on the price and cost, you can see that the most expensive models right now are dontropic models. You have Fable, have Opus, you have Sonnet, where all of those models are more expensive on a per task basis even versus GPT five dot six sole at max effort.

02:32And so if you're running into your usage limits quite a lot with Anthropic, this is the clear reason why. The other thing that I like about what OpenAI is doing apart from just putting pressure to Entropic in terms of the pricing is that if you've used OpenAI's models before, which I have been doing and just comparing it versus Entropic's models, you can definitely tell how much further this weekly usage limit goes versus the same Antropic subscription at the same price.

02:55And also recently, OpenAI has been really generous when it comes to resets. So I think this weekly usage, they actually just reset this again. And apart from that, they introduced this concept of resets that you can trigger on your own.

03:06And you can see here just over the past weeks when they were giving this out, I was able to accumulate five resets that are ready for me to use. And so this introduces a bit of a dilemma. Right?

03:16Because what I want, because I'm already subscribed to ChatGPT and I have all these resets and this raw intelligence available to me is I want to make full use of this subscription, and I want to make use of all these resets. However, the main operating system that I am using, the main harness I am using is Cloud Code, and I have all my skills there.

03:34I have my second brain already set up and linked properly to Cloud Code. And it doesn't really make sense for me yet to port everything over to Codex and use Codex as my daily driver, especially if a lot of my work and a lot of really the surface I'm used to is already that Cloud Code instance. And so the question that I ventured out to answer is how can I still use Cloud Code, the same surface that I'm used to, but also making use of OpenAI's powerful models that are challenging Fable, which right now is available to us at a much cheaper price?

04:04Well, I did come up with a solution for my personal workspace, and you can see here in Cloud Code that you can actually see that in action. So apart from Fable, Opus, and Haiku here, you can see this g p D 5 dot 6 sole being logged within my Cloud Code model usage. So this line is actually pulling from my OpenAI sub end usage limits, which is exactly what I want to happen.

04:24And just to show you that in action, you can see I have Claude code in my terminal here, and the model that it is using under the hood is GPT5Dot6Sol, but the harness that I am using is still Claude code. And because I'm still using Claude code as the harness, even though I have an OpenAI model underneath working under the hood, the clear benefit to this is that I can actually now use all of the skills and all of the commands that I have built up over time.

04:46So to give a quick example, I'm using this skill called slash align, and that's basically a simple skill that will prompt Cloud Code to ask me multiple questions so that we are aligned on the final intention of things. And you can see that it was able to pick up that skill and give me a response back. And this is more efficient versus if you were to use Codex.

05:04So to give the same example, you can see I'm in Codex right now. And what I just did here is I just pointed Codex to my same workspace, to my same second brain. But since Codex is an entirely different harness that doesn't natively have this slash align command, you can see it worked for a much longer time, forty eight seconds, in order to retrieve that skill and actually follow it.

05:22So this is a significantly longer time, especially for long horizon tasks versus in Cloud Code, where you can see that it only took like twelve seconds in order to give us a response back. So 75% faster, basically.

05:34And so that's where the benefit lies because I am now using my OpenAI subscription over at Cloud Code, but I can enjoy all of the skills and all of the commands that I have built up over time. The second brain system that I've also linked via Cloud Code and optimized via Cloud Code, that is now accessible to me. And it's actually quite easy to set up, and I'll be talking through what actually happens under the hood.

05:54Just But to help you out and to make it even easier for you, I've exported the guidelines and everything that I've set up, which you can just grab as a zip file below, and you can just point your cloud code to it so that you can try out this wiring yourself. And by the way, if you want to learn how to build and sell AI systems that businesses actually pay for, then that's pretty much all we do over at the Robo Nuggets community, where not only do you get access to the Cloud Living Masterclass, which we update every week and takes you from zero to mastery with the latest on AI, but you also get access to our agents as a service course, which walks you through how to actually get paid for all these AI skills that you are learning.

06:26You also get to be part of a genuinely great community of AI builders. In fact, you can see just some of the recent wins our members are getting from the program right here. So if you want to start earning from AI, then check that just in the pinned comment below.

06:37Now back to the video. But essentially, what happens under the hood is that Cloud Code, that is still the harness, but in order to use GPT five dot six sol or any other model with a subscription like Grok four dot five, for example, we're actually using this tool called CLI Proxy API. And CLI Proxy API is this open source project which is already quite popular on GitHub.

06:58You can see it has 40,000 stars already in here, and what it does is quite simple. It basically builds you a local program, like a small application in your own device that when you set that up for the first time, that will actually open your browser for you. That will let you log in to OpenAI in order to link your account into that application that you now have in your device.

07:17And so whenever Cloud Code needs to access gpd5.6sol, normally, in a default Cloud Code setup, it only has the capability to send requests straight to Entropic. But with this setup, you're now giving it an option to route a request through this program, which sends it to the sol model through your OpenAI subscription.

07:34And in case you're curious if this is even allowed by OpenAI, then you can probably hear it straight from their engineering lead. So over at X, where I first heard about CLI proxy API, you can see Thibault, who is the engineering lead at OpenAI and the head of Codex, actually sharing that if you are not yet ready to migrate to Codex, then you can stay in the presence of your orange crap, which is Cloud Code, and just point it at gpd5.6sol.

07:56And basically, he just outlined here the CLI proxy API method that I was mentioning earlier. And so now if I go to my terminal and just add a tag here for the model that I want to use, which is five dot six Sol, then I'll be able to launch Cloud Code using that model and link to my OpenAI subscription. Now even though using Sol directly using Cloud Code is already quite useful, the way I'm getting most value from this setup is actually by combining Entropic and OpenAI's models together.

08:22And so what I did is to just make this skill where in its essence, I just have Claude Fable doing the planning, while Sol is the one that is doing the execution, the actual work to be done in that task. And this has been quite useful because if you've used Fable before or if you've read some documentation, especially from the Cloud Code team around it, then you would know that Fable is actually a great orchestrator.

08:43And just to show you that original thread, I pulled it up over here at x. This already has, like, 5,000,000 views in here, where the Cloud Devs account is saying that if you use Fable five as an adviser slash orchestrator and use a lower cheaper model as an executor, in this case, they used Sonnet five, then at the end of it, you actually get something like 92% of Fable five's score at only 63% of the price.

09:06And so what I ended up doing is pretty much the same thing, but instead of using Sonnet five, I thought to use g p d five dot six sole precisely because of the huge amount of usage that OpenAI is giving us. And so the way it works is really simple. We're in step one with the planning that is all being done by CloudFable, and then it hands it off to sole using that integration with the CLI proxy API that we set up before.

09:27And then as an optional step, Fable can do the checking in order to make sure that it is actually up to spec with the build that we assigned to it. And so when I ran it across a couple of tests, you can see here that Fable solo does end up costing much higher versus this FAB Soul skill, which is that Fable plus GPT Soul skill that I created where Fable is handing off to Soul to do the work.

09:48And across these several tests, all of them ended up costing much less versus if I were to just have Fable do it alone, which kind of makes sense because Fable is, as we showed earlier, roughly around two times the cost of soul per task. And just to show you an actual build, what I have here are two sessions. For the first one, I was testing out using the FAB Soul skill.

10:08So this one is gonna be an orchestrator where I've asked it to do this task where Fable is the one doing the advising while Soul is doing the working. And then this other one, this other session, is one where it is the builder. So I ask it not to use FABSAL and to just use the model that it is currently on, which is right now Fable at extra high.

10:26And for both of these sessions, basically what I did is to ask them to create a landing page for Skylight, which is a SaaS app that turns marketing agencies scattered campaign data into live, client ready report dashboards. And if you were to just look at what both of these sessions did, it is pretty similar. Right?

10:43So this is the one that is using Fable five, and it is pretty good. Like, you can see there's a couple of animations here that it was able to introduce to us, and this is the one where Fable is the orchestrator while g p t five dot six is the one that is working.

10:56And so consistent to what I was observing before, g p d five dot six is actually quite good already and is almost reaching Fable's levels. And at least for this build, I actually think that the one from the fab sole skill build where you had two models working together turned out a bit better versus the one from just Fable five.

11:13And so the point being, even though Fable is really good and is probably the most frontier amongst all of the frontier models right now, the fact of the matter is that it is also really expensive. So try this FAB Soul combo out, see if it works for your own setup and for the work that you do because, obviously, each of the stuff that we do is quite different.

11:30So you may experience different results, but I'll provide the link to this skill down below so that you can test it out and see if it's going to be useful to you. So there, I hope that was useful. I hope that sort of expanded your mind a bit in terms of where we are going.

11:43Because with a lot of these tools, even though we are using Cloud Code and you prefer to use Cloud Code as a harness, I think one of the best ways for you to get ahead and for you to really understand how the space is evolving is to not be just shackled into just one provider. Especially since the space moves around a lot, OpenAI is right now, in my view, doing a really good job with offering such a good model with five dot six sol at such cheap prices given how much usage they're giving us.

12:08And overall, I think that is good so that Anthropic is feeling that bit of pressure so that they can offer these same models to us at the same usage rates as well. But that's it for this one. As usual, thanks for watching until the end.

12:21I'll see you all next time. Thank you.

The Hook

The bait, then the rug-pull.

Claude Code is the harness he's most invested in, but Anthropic's models are the most expensive per task on the market — so instead of switching tools, he wired Claude Code to run GPT-5.6 Sol through his existing ChatGPT subscription, and built a skill that splits every task between the two.

Frameworks

Named ideas worth stealing.

06:37model

The Wiring

Claude Code (interface, skills)
CLIProxyAPI (local proxy, port 8317)
Fable 5 (Claude subscription)
GPT-5.6 Sol (ChatGPT subscription)

Claude Code sends every request to a local proxy, which routes it to either the Claude subscription or the OpenAI subscription depending on which model is aliased in the request.

Steal forany workflow where you want one harness to bill execution against two separate subscriptions

08:22list

The Flow (FabSol skill)

1. Split the work with Fable
2. Sol builds — subagents execute on the cheap model
3. Fable checks — reviews the diff, runs the tests, ships

A three-step orchestrator/executor pattern where the expensive model only plans and reviews, and the cheap model does all the token-heavy execution.

Steal forany multi-model coding setup, not just Claude plus OpenAI

CTA Breakdown

How they asked for the click.

VERBAL ASK

05:33link

“I've exported the guidelines and everything that I've set up, which you can just grab as a zip file below, and you can just point your Claude Code to it so that you can try out this wiring yourself.”

Soft in-video CTA dropped mid-tutorial right when the viewer would want the actual files, followed a few seconds later by a separate sponsor pitch for the RoboNuggets community.

MENTIONED ON CAMERA

01:21toolArtificial Analysis ↗

05:33productFabSol skill download (RoboNuggets) ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.