11:39Jack Roberts · Demo
Fable 5.1 vs ChatGPT: One-Shot Website Design Put to the Test
Five identical design briefs, one shot each, no retries — Fable 5.1 against ChatGPT 5.6 Sol, plus the exact stack behind the wins.
September 2ndThe argument in one line.
Claude Code the harness and Claude the model are separable, so a proxy that swaps in free API keys from NVIDIA and other providers can keep the same coding workflow running well past the paid subscription's usage limit.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code daily and regularly hit its usage limit before the day is over.
- You're comfortable pasting a setup prompt into Claude and signing up for a couple of free developer accounts.
- You want to save your paid Claude usage for the hardest tasks and offload routine work to free models.
SKIP IF…
- You need consistent frontier-level output on every task; free-tier models trade quality for cost.
- You're not willing to manage API keys across several providers.
The full version, fast.
Claude Code's harness, the rules, skills, memory, and prompts around the model, can run on any model that answers in Claude's API shape, which is what lets a free GitHub repo swap in free-tier keys from NVIDIA, OpenRouter, and other providers once the paid Anthropic quota runs out. The setup is one prompt pasted into Claude, followed by a free NVIDIA API key that needs no card and allows about 40 requests per minute, and the tool automatically reroutes when one tier runs dry. The tradeoff is real: free models fall behind on design and architecture calls and can misreport which model actually answered, so the sustainable pattern is running roughly 80% of grunt work on free tiers and saving the remaining 20% for frontier models.
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 →Where the time goes.
00:00 – 00:29
01 · The hook: paid Claude Code stops cold
Claude is expensive, and once you hit your limit you're stopped dead. A new GitHub repo promises 1.3 billion free tokens a month inside the Claude Code harness.
00:29 – 01:33
02 · Model vs. harness
Performance comes from two separable layers: the model (Opus, Fable, etc.) and the harness (the rules, skills, prompts, and tooling built around it). Only the model needs to come from Anthropic.
01:33 – 02:02
03 · Meet the free keyring skill
One setup command flips a workflow from a paid monthly bill to $0 by routing through free-tier providers instead.
02:02 – 02:39
04 · Why NVIDIA over OpenRouter
OpenRouter needs a card on file and can rate-limit or serve leftover capacity. NVIDIA's free tier needs no card and allows about 40 requests per minute with frontier open models.
02:39 – 03:38
05 · Install the skill and sign up for NVIDIA
Paste the provided prompt into Claude to install the tool, then create a free account at build.nvidia.com.
03:38 – 04:30
06 · Generate the NVIDIA key and add more providers
Generate a named, scoped API key on NVIDIA's site, paste it into the tool's config, and apply it. Additional provider keys (OpenRouter, Gemini, Groq, GitHub) slot in the same way.
04:30 – 05:41
07 · Building the keyring, and a masterclass plug
Each added key is one entry on a keyring rather than a single point of failure. Jack recommends a shortlist of providers, then plugs his paid Claude Code masterclass.
05:42 – 06:36
08 · Running Free Claude Code and testing it
Launch the tool from the terminal and send a trivial test prompt. The response falsely claims to be Opus 5 and reports a fake $0.48 charge, when it's actually routed through NVIDIA's free tier.
06:36 – 07:19
09 · Automatic failover
When one free tier runs dry, the tool automatically reroutes to another, which solves the classic failure mode of free AI access: a short session before the workflow just stops.
07:19 – 08:29
10 · Connecting Zapier as an MCP server
Jack wires his existing Zapier MCP server into the free setup as an authentication layer, so cheaper models can reach his other tools with scoped-down access.
08:29 – 09:32
11 · The 80/20 rule: free for grunt work, frontier for taste
Frontier models are still worth paying for on the highest-stakes calls. The recommended split is roughly 80% of tasks on free or cheap tiers, 20% reserved for frontier models.
09:32 – 11:13
12 · Three caveats: hallucination, shrinking tiers, model cosplay
Free models hallucinate more and need review by a stronger model. Free tiers shrink or disappear without warning. Free models miss most on taste, and a model can misreport which underlying model actually answered.
Lines worth screenshotting.
- Claude Code's harness, the rules, skills, memory, and prompts around the model, can run on a different model entirely once the paid quota is gone.
- A free NVIDIA API key from build.nvidia.com needs no card on file and allows roughly 40 requests per minute.
- The GitHub repo behind this setup works like a keyring: add API keys from NVIDIA, OpenRouter, Gemini, Groq, and GitHub, and it automatically reroutes to whichever tier still has quota.
- When you ask a rerouted model what it is, it can falsely claim to be Claude Opus, because every provider it proxies claims to be Claude inside the harness.
- Free tiers get pulled or shrunk without notice, so any setup built entirely around one free tier degrades in reliability over time.
- Free models fall furthest behind on taste, meaning design and architecture decisions, not raw code output.
- A workable split is roughly 80% of tasks on free or cheap models and 20% reserved for frontier models on the highest-stakes work.
- The only way to confirm which model actually answered a request is to check the provider's billing or usage log, not to ask the model what it is.
Free AI models can replace paid usage, within limits
COST STRATEGY
A model and the harness built around it are separate things, which means a proxy that swaps in free API keys can keep a paid AI coding workflow running, as long as you know exactly what you're trading away.
02Model vs. harness
- A model and the harness wrapped around it, the rules, skills, memory, and prompts, are two separate layers, and only the model needs to come from Anthropic for the harness to work.
- Understanding this separation is what makes it possible to keep using a familiar AI coding setup after a paid subscription's usage limit runs out.
03Meet the free keyring skill
- A single setup prompt pasted into an AI coding tool can install a proxy that reroutes requests to free-tier API providers when the paid quota is exhausted.
- Toggling a workflow from a paid model to free-tier providers can take usage cost from a fixed monthly bill down to $0, with real tradeoffs attached.
04Why NVIDIA over OpenRouter
- Free API access from a hardware vendor's developer platform can require no credit card at all, unlike some model marketplaces that ask for one upfront.
- A rate limit of roughly 40 requests per minute is generous enough for typical coding-assistant usage, since most sessions don't need a request every second.
05Install the skill and sign up for NVIDIA
- Free developer accounts on major AI infrastructure platforms are usually a five-minute signup with no payment method required.
06Generate the NVIDIA key and add more providers
- Naming and scoping an API key at creation time, rather than reusing one generic key everywhere, makes it easier to revoke or rotate access to a single provider without breaking the others.
07Building the keyring, and a masterclass plug
- Treating multiple provider API keys like a keyring, rather than committing to one provider, means a workflow doesn't fully break when any single free tier changes its terms.
- A curated shortlist of providers, rather than adding every option available, keeps a keyring manageable, since some pairings are more reliable than others for coding work.
08Running Free Claude Code and testing it
- Testing a new AI setup with a trivial prompt first, before running real work through it, is the fastest way to confirm the routing actually works.
- A tool reporting its own identity, or its own cost, is not reliable confirmation of what actually happened, the model can misreport both.
09Automatic failover
- Automatic failover between free tiers solves the biggest weakness of any single free AI plan: a short session before you hit a wall and the workflow just stops.
10Connecting Zapier as an MCP server
- Routing a workflow's tool access through a dedicated authentication layer, rather than granting it directly, makes it possible to scope down exactly what a cheaper or less-trusted model is allowed to touch.
11The 80/20 rule: free for grunt work, frontier for taste
- Frontier models are still worth paying for on the calls that matter most, architecture and design decisions, even when everything else runs free.
- Treating a free tier as a total replacement for a paid frontier model, rather than a supplement to it, is a false economy: the quality gap shows up on exactly the work you can least afford to get wrong.
- A rough 80/20 split, cheap or free models for routine work and frontier models for the hardest 20%, is a workable default for balancing cost against output quality.
12Three caveats: hallucination, shrinking tiers, model cosplay
- Free and low-cost models hallucinate more often, so any output from one should be reviewed by a more capable model or a human before it ships.
- Free API tiers get cut or shrunk by providers without warning, so a setup built entirely around one free tier is fragile in a way a paid subscription isn't.
- Free models perform worst on taste, meaning design and architectural judgment calls, and best on repetitive, well-specified execution work.
- A model can misreport which underlying model actually answered a request, so the only reliable way to confirm what's running is to check the provider's own usage log, not to ask the model directly.
Terms worth knowing.
- Harness
- The layer of rules, skills, prompts, and tooling wrapped around an AI model, in this case everything Claude Code adds on top of the underlying Claude model.
- Keyring repo
- A GitHub tool that stores API keys from multiple AI providers and automatically switches Claude Code's traffic to whichever key still has free quota left.
- Automatic failover
- The proxy's behavior of rerouting a request to a different provider's free tier the moment the current one runs out of quota.
- Model cosplay
- A rerouted request answering as if it were Claude, even though a different provider's model actually generated the response.
Things they pointed at.
Lines you could clip.
00:00
“The truth is that Claude is expensive and once you hit your limit, you are stopped dead in your tracks.”
02:28
“With NVIDIA, it truly is a $0 start, so it's very easy to get signed up.”
09:32
“Sometimes models, guys, they will lie to us. They will hallucinate.”
11:00
“I actually tripled my usage inside Claude. It is one of the best systems I've seen.”
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.
metaphoranalogy
Imagine if you could use ClaudeCode 100 % for free. The truth is that Claude is expensive and once you hit your limit, you are stopped dead in your tracks. However, a brand new repo just came out that gives you 1 .3 billion tokens to use inside the ClaudeCode harness every month for $0.
And in this video, I'll show you exactly how to set it up in one prompt. So you can use unlimited ClaudeCode harness, build more save time and get light years ahead of everybody else.
So if you haven't already, grab that beautiful coffee and let's dive straight in. So this is an actual new system. I'm going to cover the repo and why this is a better way to do it.
So one thing to understand with this is that effectively we can use... clawed code harness for free using nvidia and this incredible github repo but first of all you have to understand this if you're really going to get the benefits out of it even my dog agrees he's barking in the background so when it comes to performance it is a combination of two things there is the model so that could be clawed opus 4 .8 Fable 5, insert the blank.
And then we have the harness. The harness is the sets of rules. It's where you build.
It's all the prompts around it. Classically speaking, it's this. It's using inside the code harness with all of your skills and all of the other beautiful things that Anthropic has put behind it.
But you only get so much usage of these models. So what happens when you run out? Or what if you want to spend over a billion tokens?
Well, you can do that. for free with this system. And the way that I built this skill that I'm going to give you down below for free, it'll be the second link in the description.
You just give it one command, you toggle it off and you go from paying all your subscriptions down to $0. And when you make it to the end of the video, you are going to have such a ease of life hack and it is going to save you so many dollars. Now we're going to be using NVIDIA, which is incredible.
And you're going to get a 100 % free API key. You do not need to put down any debit card information. It is free.
I will explain exactly the business model so you can understand how that works. but it's completely free and you get access to 1 .3 billion tokens from a combination of 50 different providers. Now, you might be wondering why we're doing this versus OpenRouter.
Basically, OpenRouter, obviously you need your card, which I think is fine. Obviously, they should be making money. You get free models, but you can hit a daily wall and it'll cap you out and you can get rate limited.
And some have accused of just getting the leftover sometimes, but I do love OpenRouter. With NVIDIA, it truly is a $0 start, so it's very easy to get signed up. There is...
basically never a wall with you're going to hit with nvidia and you get frontier open models and it is 100 100 for free and they do this because Hopefully one day you go and actually rent from NVIDIA because you love the service so much. So with that in mind, let's get started and get this downloaded.
So now we've done that, let's go ahead and get this whole thing running. So I want you to click the link down below. Once you've done that, you're going to see this here.
This will explain exactly what it is, what the problem is, how it works, and how we set up. And all you're going to do is come down, copy this, and then head over to Claude. Just drop that in, hit enter, and this will do the rest.
So we're going to come down here and click yes, run it, recommended. Now this is using this GitHub. The one that's got over 50 ,000 stars and is blowing up right now.
Then once that is done, it will open up this window on your own computer. And all we're going to do now is give it a connected account. And we're going to be doing it with NVIDIA, which is going to be incredible.
And then click on configure. it brings you down, then head over to this website. I'll put a link down below.
It's build .nvidia .com. You're going to go ahead and sign in and create an account. Now, you're limited up to 40 requests per minute, which means you can hit the API essentially once every 1 .5 seconds.
That's a lot of requests, guys. I ask a lot of questions. I don't ask that many questions.
In other words, it's very, very generous. So come over and click on API keys. And by the way, if you are new here, I'm Jack.
I built and sold my last tech startup with a gazillion customers. Now I'm building my own AI startup. And on the channel, I share here the stuff that actually works.
So come over and click on Generate API Key. Give it a name. So I'm going to call this one Free Cloud Code.
Expiration for 12 months. And then click on Generate Key. And then you're going to have the API key.
You're just going to copy this like so. Come back over here and just drop. in right there and once you've done that click apply in the bottom right hand corner and now that is fully saved now basically you can just go ahead and enter in all the api keys from all of the different providers so we're going to have open route to grow climb plus and obviously token limits change day by day and month by month sometimes so once you add in all the api keys this does currently give you access to 1 .3 billion tokens per month so this is the idea we've got all the tokens to all these different models you can add them all in and then it will pull through and pull a different model based on various different kind of like standards so let's say i want to add an open route to api key awesome let's go ahead and grab that so i come over here i click on a new key for example give it a name like claude free and basically guys we'll just create these free counts from all these different locations which is cool and effectively once you've done that you can just provide it to the model itself and use that and then once you've got it you can come down and throw it in here if you want to
configure throw it in and repeat the process now in reality you wouldn't do all four to two think of this as like the key ring and then each of these api keys are the keys so i would add in personally nvidia open router is a classic gemini grok github is a good one as well by the way If this all sounds like I'm and you want to get even more out of Cloud Code AI systems, I'm going to put a link down below for the full Cloud Code masterclass that will take you from foundation setups, building websites, skills, and content I have never shared on YouTube, including full access to the agentic operating systems, all of these crazy memory systems, and of course, the design operating system, a gargantillion incredible things.
I'll put a link below. If you want to level up, don't tell your competitors that. It will give you a very unfair advantage.
So now what have we done? We effectively gone ahead and we have now given it all the keys we need. So now we can head back over to Claude and actually use this really cool feature.
So to go ahead and run this, all we're going to do is command spacebar and we're going to open up the terminal like so. And then all we're going to do in the terminal, guys, is type in this command. And again, this will be available in the Notion doc, just FCC Claude.
And this right here is free code code, which is awesome. Yes, I trust this folder. We can go ahead.
And now we are literally using our new free models. So to test it, we can say something like, hey. Say Pong.
Now, when you ask this model, what is it? It will tell you Opus 5. That's because it's rooting the ticker.
It will also show you, you spent 48 cents. That's not true. It's basically actually connected to Nemetron via NVIDIA, and it's got the full cycle of different free models it can access.
And just like in QR code, you can do forward slash. mcp and access all of your mcp servers now before i show you one of the really cool use cases with this you need to understand why this is actually so cool so what's really cool is if a tier runs dry it will automatically reroute it which is great because one of the biggest limitations with free models so to speak is that they can be on for a very short period of time you'd be coding to your heart's content and then it's like sorry you can't use it the cool thing about this system we plug all the different free stuff in as and when it appears and then just automatically tracks you meaning you get kind of hassle -free coding on the freer models and i've also included this really cool section that lets you connect any of your mcps that you're already using inside cloud over to this new free code.
So if, say, for example, I wanted to go ahead and connect Zapier, I could basically copy this, come over here, type this in like so into the chat window, bring it over here, and all I'll do is enter in my URL. Now, of all the ones that I would use, because these models, guys, are not as powerful as the Frontier, obviously, you're trading on performance, cost, and speed.
Right. The thing about Zapier is I can actually, it's my authentication layer. I can use it across my agentic operating system.
I can use it across Cloud Code. I can use it everywhere to give all the specific access I need. And it has stuff that I just can't access in other places, like the school API, for example.
So what I like to do is come down and create a new MCP server specifically for my new free Cloud Code sections. I'll put a link down below so you can grab this. I come over and I literally, guys.
generate the token, copy it, and then I can use that inside of Free Cloud Code to do things. So it has the ability to access different things. But one of the reasons I really like this for Zapier is because I can literally in the app say exactly what I do and do not want it to have access to.
For example, I can specifically remove certain requirements. Yes, you can also do this in Chord, but it's very handy for me to have multiple different levels of authentication that I like. And I can use the same thing in multiple different apps.
And connections is one thing, but if you're not using it properly, you're never going to get the best results. So think about it like this. You do want the frontier models for work.
Like you just genuinely do. And it is a... what we'd call a false economy to think that you want to do free for everything.
A tiny bit of money can get a very long way. So you never want to go all free. You want to have frontier for your best, best tasks.
And if you're on the $200 plan, dude, rock on, you're kicking butt. It's actually unlikely you're going to need this if you watch this video on screen, which I highly, highly recommend that you do. It's going to blow your mind in terms of how it can actually increase your token usage.
But typically speaking, if you really want to crush the free model gauntlet, you want to do 80 % on the cheaper models. So that could be Haiku or this free tier, and then really save the 20 % for the frontier. Now, what's really important to bear in mind is the caveat.
You are trading... For cost, since we're bringing the cost down, your performance is going to come down. But also sometimes models, guys, they will lie to us.
They will hallucinate. So just bear that in mind with your free models. And you need to be extra careful to double check, spin up sub -agents and cross -reference for any work you're doing on free, anywhere possible.
Have that reviewed by a more intelligent frontier level model. Now, also, I want to equip you so you walk into this with both eyes open. Free tiers.
they do shrink. So for example, do you remember OxAlpha? Fantastic, right?
That's great. Some of the best labs open things up for free, but they will take them off the market at time, which is one of the reasons why this kind of keychain repo is awesome because we have so many, you're less interrupted and things run a bit smoother after like 10 minutes of setup. Also, the amount of free they give you can change from time to time, which is awesome.
Free models perform, I'd say, the worst on taste. So if you're looking to design your next beautiful piece of art the next mona lisa Maybe we don't bring in the free models for that, but it can be incredible for grunt work running on your local AI labs 24 -7, just have a load of chucking on infinitely to do that, which is awesome.
And bear in mind that it will sometimes say it is Opus 5 when we pull the Scooby -Doo mask off, we can tell it is in fact something completely different. So it's just reading from the surface level data it's got. It's not the model that's actually being used.
And you can always check that by just asking Claude. Now using the Claude Code harness is one thing, but I did one... And I actually tripled my usage inside Claude.
It is one of the best systems I've seen. And so the next thing we need to do is get you running on that system in one prompt, which we can do in this video right here.
The Hook
The bait, then the rug-pull.
Jack Roberts opens with the pain every heavy Claude Code user knows: the workflow stops cold the moment the usage meter hits zero. His fix is a GitHub repo that reroutes Claude Code through free API keys from NVIDIA and other providers, so the same harness keeps answering after the subscription taps out.
Named ideas worth stealing.
01:07concept
Model vs. Harness
AI coding performance comes from two separable layers: the underlying model (Opus, Fable, etc.) and the harness wrapped around it (rules, skills, prompts, memory, tooling). Because only the model call needs to hit Anthropic's paid API, the harness can keep running on a different, free model once the paid quota runs out.
Steal forexplaining vendor lock-in and cost structure for any AI tool built as a harness over a model, not just Claude Code
09:59list
Three Things Nobody Tells You About Free AI Models
- Free tiers shrink
- Free models miss on taste
- Models lie about who they are
The three hidden costs of building a workflow on free-tier AI models: providers cut or remove free allowances without warning, free models fall furthest behind on design and architecture judgment, and a rerouted model can falsely claim to be a different, more capable model than the one that actually answered.
Steal forsetting expectations before recommending any free-tier AI tool
08:40model
80/20 Free vs. Frontier Split
- 80% grunt work on free or cheap models
- 20% taste and architecture calls on frontier models
A cost-control heuristic: route routine, well-specified work to free or cheap models, and reserve frontier models for the smaller share of tasks where design judgment and architecture decisions matter most.
Steal forbudgeting AI usage across any project with mixed task difficulty
How they asked for the click.
VERBAL ASK
05:17product
“I'm going to put a link down below for the full Cloud Code masterclass that will take you from foundation setups, building websites, skills, and content I have never shared on YouTube.”
soft mid-video pitch for his own paid course, positioned as the next step after already delivering the free tool, not a hard sell
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
