Big Idea

The argument in one line.

A frontier model's advantage is mostly its written operating discipline, not its raw intelligence, so extracting that discipline into a skill file lets cheaper models match its output quality for a fraction of the cost.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code, Claude Projects, or similar agentic coding tools and want to cut API/subscription costs without losing output quality.
You already orchestrate multi-agent or sub-agent workflows and want a cheaper way to route tasks across models.
You're deciding whether to keep paying for the most expensive model tier by default instead of matching model cost to task difficulty.
You want a concrete, copyable method (not just a philosophy) for making a weaker model behave like a stronger one.

SKIP IF…

You don't use any AI coding assistant or agent framework, so model-routing and skill files have no application for you.
You're looking for a Fable 5 feature walkthrough rather than a cost-optimization technique.

TL;DR

The full version, fast.

The creator spent thousands of dollars testing Claude Fable 5 against Opus and Sonnet and found that when a smart model orchestrates cheaper sub-agents, results are nearly identical to an all-Fable run at a fraction of the cost — proving intelligence isn't the moat, process is. After Fable 5's system prompt leaked, he distilled its operating habits (verify before trusting memory, answer before asking, calibrate effort to task size, work through five gates: scope, evidence, attack, verify, report) into a portable 'skill file' he calls Fable Mode. Loading that skill into Opus 4.8 makes it plan, self-check, and report the way Fable does, without Fable's price tag. He also keeps a scored table of models (cost, intelligence, taste) so an orchestrator model can route sub-tasks to the cheapest model that can still do the job — in one real test, swapping Sonnet/Opus workers for Haiku workers cut cost roughly 3x with identical results.

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 – 00:27

01 · The model isn't the moat

States the core thesis: Fable 5 is powerful, but it isn't the actual advantage.

00:27 – 01:03

02 · Karpathy on Sonnet beats a beginner on Fable

Comparison: an expert instructing a weaker model beats a beginner instructing a far stronger one — instruction quality, not model IQ, wins.

01:03 – 01:34

03 · Fable orchestrating Fable vs. Fable orchestrating Opus/Sonnet

His own dynamic-workflow tests: results were about the same whether sub-agents were Fable, Opus, or Sonnet, even though all-Fable cost exponentially more.

01:34 – 02:13

04 · Teacher, not workhorse

Reframes the frontier model as a senior engineer packaging its judgment for junior models to inherit, instead of a workhorse you push on every task.

02:13 – 02:42

05 · Model routing: matching intelligence to task cost

Introduces model routing — tasks need different amounts of intelligence, and overpaying for max intelligence on easy tasks wastes money.

02:42 – 03:12

06 · The plan: extract how Fable thinks, run it on smaller models

States the actual plan of the video: extract Fable's thinking process and let cheaper models execute it the same way.

03:12 – 03:52

07 · What the leaked system prompt reveals

Walks through real excerpts from Claude Fable 5's leaked system prompt: don't trust memory over verification, a mentioned file may not exist, answer before asking (one question max), own mistakes without an apology spiral.

03:52 – 04:20

08 · Effort calibration: how much work per task

The prompt's effort budget rule — roughly one tool call for a signal fact, three to five for medium tasks, five to ten for deep research.

04:20 – 05:25

09 · Accuracy vs. cost across effort levels

Uses the FrontierCode benchmark chart comparing Fable 5, Opus 4.8, and GPT-5.5 across effort levels — Fable on low is close to Opus on high, and maxing effort can make a model overthink and get worse.

05:25 – 06:26

10 · Extracting the Fable Method into a skill

The actionable instruction: take a session whose output you loved but couldn't explain, have the model analyze its own process, and turn that into a reusable skill file — 'Fable Mode' for Opus.

06:26 – 07:14

11 · Fable Mode: the five gates

Names the extracted skill file's five-gate discipline — Scope, Evidence, Attack, Verify, Report — as the working discipline any model can run.

07:14 – 07:36

12 · Adversarial planning is the differentiator

Distinguishes ordinary step-by-step planning from adversarially planning for every possible failure mode, and how that lets Sonnet execution loop back to a Fable planner with equivalent quality.

07:36 – 08:01

13 · The one prompt to build it yourself

Gives the exact reusable prompt for generating your own skill file that transfers judgment, planning, verification, and reasoning habits, and where to get his free 'Fable Mode' file.

08:01 – 08:31

14 · Scoring your model toolkit (cost / intelligence / taste)

Shows a scored table for Fable 5, Opus 4.8, Sonnet 5, and Haiku 4.5 so an orchestrator model can route each sub-task to the cheapest model that clears the bar.

08:31 – 09:26

15 · Real test: Haiku scouts cut cost 3x

Describes an actual orchestration test comparing Sonnet, Opus, and Haiku sub-agent workers under an Opus orchestrator — the Haiku version was about 3x cheaper with the same result.

09:26 – 09:59

16 · Own your process, not the model

Closes on the broader point: access to any specific model can be revoked (as happened briefly with export controls on Fable 5/Mythos 5), so the durable asset is your process, methodology, and eventually owned hardware/local models.

Atomic Insights

Lines worth screenshotting.

A beginner using Fable 5 will still lose to an expert like Andrej Karpathy using the older, weaker Sonnet 3.7 — instruction quality beats model intelligence.
Running the same dynamic workflow with a Fable orchestrator directing Opus sub-agents versus Fable directing Fable sub-agents produced roughly equal results, despite the all-Fable run costing exponentially more.
You can't keep a frontier model's raw intelligence, but you can keep its process — the reusable habits are the actual asset.
Claude Fable 5's leaked system prompt includes the instruction: partial recognition from training does not mean current knowledge, so the model should verify rather than trust memory.
Fable 5's prompt tells it to address even an ambiguous query before asking for clarification, capping follow-up questions at one.
Fable 5's effort-calibration rule: roughly one tool call for a signal fact, three to five for a medium task, five to ten for deeper research or comparison.
On the Claude Fable 5 / Mythos 5 release benchmark, Fable 5 on its lowest effort setting scores similarly to Opus 4.8 on its highest effort setting, at a somewhat higher cost.
Maxing out effort ("xhigh"/"max") on either Fable or Opus can backfire — the model overthinks, second-guesses itself, and produces worse output than a moderate effort setting.
The creator's extracted skill, 'Fable Mode,' runs any hard problem through five gates in order: scoping, evidence, attacking (adversarial self-challenge), verifying, and reporting.
Genuine planning (list the steps, execute) is different from adversarial planning (play devil's advocate against your own plan, surface every way it could fail) — Fable's edge comes from doing the latter.
A model orchestrating cheap sub-agent workers, given the Fable Mode prompt, can plan out every failure mode in advance and delegate pure execution downstream, closing most of the quality gap with an all-frontier-model run.
The prompt template for building this yourself: 'Write a complete installable skill file that makes Opus 4.8 operate with your judgment, your planning, verification, and reasoning habits, and activate it on something like fable mode.'
Keeping a scored table of every model in your toolkit — cost, intelligence, taste (creativity/UX judgment) — lets an orchestrator route each sub-task to the cheapest model that can still hit the bar.
In one orchestration test, an Opus orchestrator delegating to Haiku sub-agents ('scouts') instead of Sonnet or Opus workers cut cost roughly 3x while producing the same result.
Since companies and individuals will increasingly operate under fixed AI budgets, matching task difficulty to the cheapest sufficient model will separate teams getting far more output for far less spend.
Because access to frontier models can be revoked or gated (as happened when Fable 5 and Mythos 5 were briefly restricted under export controls), the durable asset to own is your process, methodology, and eventually your own hardware/local models — not the model itself.

Takeaway

The model's edge is its written process, not its raw intelligence.

MODEL ROUTING

You can extract a frontier model's operating discipline into a portable skill file and hand that discipline to a much cheaper model, closing most of the quality gap for a fraction of the cost.

01The model isn't the moat

Fable 5 is a genuinely strong model, but it isn't the actual advantage over other models.

02Karpathy on Sonnet beats a beginner on Fable

An expert instructing a weaker model beats a beginner instructing a far stronger one — instruction quality outweighs raw model intelligence.

03Fable orchestrating Fable vs. Fable orchestrating Opus/Sonnet

Fable-orchestrating-Fable and Fable-orchestrating-cheaper-models produced about the same results, even though the all-Fable run cost exponentially more.

04Teacher, not workhorse

Treat a frontier model like a teacher packaging up judgment for others to inherit, not a workhorse you push to its limits on every task.

05Model routing: matching intelligence to task cost

Tasks need different amounts of intelligence, and paying for maximum intelligence on an easy task wastes money for no quality gain.

06The plan: extract how Fable thinks, run it on smaller models

The actionable move is to extract how the frontier model thinks and let cheaper models execute that same thinking.

07What the leaked system prompt reveals

Verify before trusting memory, answer before asking for clarification (one question max), and own mistakes without an apology spiral — these are learnable habits, not proprietary intelligence.

08Effort calibration: how much work per task

Effort budget scales with task type: roughly one tool call for a simple fact, three to five for a medium task, five to ten for deep research or comparison.

09Accuracy vs. cost across effort levels

A cheap model on its lowest effort setting can score close to an expensive model on its highest setting, but maxing out effort on any model can backfire, causing it to overthink and produce a worse result.

10Extracting the Fable Method into a skill

Take a session whose output you loved but couldn't fully explain, have the model analyze its own process, and turn that process into a reusable skill file.

11Fable Mode: the five gates

A five-gate discipline — Scope, Evidence, Attack, Verify, Report — is the transferable working method behind a frontier model's judgment.

12Adversarial planning is the differentiator

Adversarial planning (hunting for every way a plan could fail) is what separates a frontier model's plan from an ordinary step-by-step plan, and it's the part worth copying into cheaper models.

13The one prompt to build it yourself

The reusable prompt for building this yourself: ask the model to write a complete installable skill file that transfers its judgment, planning, verification, and reasoning habits, activated by a phrase like 'fable mode.'

14Scoring your model toolkit (cost / intelligence / taste)

Keep a scored table of every model in your toolkit across cost, intelligence, and taste so an orchestrator can route each sub-task to the cheapest model that clears the bar.

15Real test: Haiku scouts cut cost 3x

In a real test, swapping Sonnet and Opus sub-agent workers for Haiku workers cut cost roughly 3x while producing the identical result.

16Own your process, not the model

Access to any specific AI model can be revoked or restricted without notice, so the durable asset to build is your own process, methodology, and eventually owned hardware or local models.

Glossary

Terms worth knowing.

Skill file: A written document (like Claude's 'skill' files) that encodes a working method or discipline that any model can load and follow, transferring behavior rather than intelligence.
Model routing: Matching each task to the cheapest model capable of handling it, instead of defaulting every request to the most expensive/most capable model available.
Effort level: A setting (e.g. low/medium/high/max) that controls how much reasoning, tool-calling, and iteration a model applies to a given task, trading cost and time against output quality.
Dynamic workflow: An agentic setup where a model designs its own multi-step process and spins up sub-agents on the fly to execute parts of it, rather than following a fixed script.
Orchestrator / sub-agent: A pattern where one model (the orchestrator) plans and delegates, while separate model instances (sub-agents or workers) execute individual pieces of the task and report back.
System prompt: The hidden instructions a model provider gives the model before any user message, defining its default behavior, caution rules, and working habits.

Resources

Things they pointed at.

05:44linkClaude Fable 5 / Mythos 5 release blog (FrontierCode accuracy-vs-cost benchmark)

05:48tool"Fable Mode" skill file (creator's own, shared free)

07:36linkFree Skool community — all YouTube resources, linked in description

Quotables

Lines you could clip.

01:03

“You can't keep the model's intelligence, but you can keep its process.”

the single-sentence thesis of the whole video→ TikTok hook↗ Tweet quote

03:38

“Partial recognition from training does not mean current knowledge.”

a verbatim leaked system-prompt line with strong standalone meaning→ IG reel cold open↗ Tweet quote

08:15

“A skill file won't transfer raw intelligence. It transfers how Fable plans, checks its mistakes, and reports. That's the part worth keeping.”

honest caveat that also reinforces the core claim→ 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.

metaphoranalogy

Since we all got access to Fable five, I've spent a few thousand dollars in usage credits just playing around with it, understanding how it works, and understanding, more importantly, how we as people can get the most out of a model that's powerful. And I think one of the most important takeaways I've had is that, yes, Fable five is an incredible model.

Don't get me wrong. But the model isn't really the moat. So think about it like this.

You've got a beginner with AI, and you've got someone like Andre Karpathy, for example. If you give the beginner Fable five and you give Karpathy Sonnet 3.7, Carpathi will build something better than the beginner even though the beginner's model is exponentially better.

And that's because of the fact that it's way more important the way you instruct it and the systems you build and the loops you build around the model. Here's another example I want you to think I've obviously tested Fable five a ton. I've also tested Opus a ton and Sonnet a ton.

One of my favorite things to do lately is to use the dynamic workflows that Cloud Code lets us, you know, spin up. And I've done a ton of dynamic workflows where I said, hey, Fable, design a workflow and then all of your little sub agents, have those be Fable as well. Then And I would do the exact same tests with Fable orchestrating a bunch of Opus sub agents and Fable orchestrating a bunch of Sonnet agents.

And what I found is that when I run those dynamic workflows, the results are about the same. Even though the Fable runs costed me exponentially more. So anyways, the point I'm trying to make, you can't keep the model's intelligence, but you can keep its process.

So today what I wanna talk about is basically, really quickly, how can you turn a model like Opus into something that feels more like Fable? And the first thing is that you have to think of this model more like a teacher rather than a workhorse. When we all got access to Fable, I think we were basically just pushing it to its limits.

And then Thropic even put out some stuff about, like, seeing how it's really good at, like, long tasks. You know?

Working towards a goal, planning it out, executing it, and then verifying it. And it is really good at that, but the whole idea of model routing is basically finding that balance.

You know? This task takes this much intelligence. So why would you need a model that has this much intelligence and cost you this much for something that you could, you know, grab a much smaller, cheaper model, and and you could get the task done at the same level of quality.

And that is the whole name of the game, and that's going to be a very important skill to master as we head into, you know, the next years of AI. So rather than having Fable, you know, plan everything out and do everything, we're trying to extract the way that Fable thinks and then let other smaller models think it like that and execute like that.

I've had Fable go through my setups and make improvements and go through my skills and improve them, and I realized that I was just treating Fable like kind of like a cofounder or more like an officer at my company rather than just an employee. Kind of like a senior engineer that is about to retire, and it's trying to package up everything that it knows to hand over to the new cohort of junior engineers that's gonna come over or come in and take its place.

So one thing that happened recently was the system prompts got leaked from Cloud Fable five. So I read through this whole thing. I had Fable five read through this whole thing, and we picked out some important things that we noticed from this prompt itself.

Partial recognition from training does not mean current knowledge. Meaning, just because something's in your memory, you should probably verify it. Very similarly over here, a prompt implying a file is present doesn't mean one is, so it's told to check that things actually exist.

So basically, making sure at every step of the way that what it's doing and what it's done is accurate. Address even an ambiguous query before asking for clarification. So answer first, then ask.

One question max. Acknowledge what went wrong. Stay on the problem.

Maintain self respect. One for signal facts. Three to five for medium tasks.

Five to 10 for deeper research comparison. So basically talking about effort. So, yes, we have the discussion around what model is right for the task, but we also have the discussion around what effort level is right for the task.

And something that I like to look at is this example, which was on the release blog for Cloud Fable five and Mythos five. So this compares Fable five and Opus 4.8 as well as GPT 5.5 when it comes to the score on the y axis, the cost on the x axis, and then we see each of these models and we see different effort levels.

So if you're one of those people that has just turned on Fable five and you're just using it on the default effort level or same with OPUS and you never play with those, then definitely start playing around because it gets a little interesting. Like here, you'll notice that Fable five on low is pretty similar to Opus 4.8 on high.

Now Fable five is a little bit more expensive and a little bit higher quality as you can see here, but they're kind of similar. But that doesn't always mean that higher effort is actually better. I've had a lot of times where with Fable, I'm just basically using it on high because when I use x higher max or even on Opus when I use x higher max, it starts to go way longer, get way more expensive, and then it overthinks and it, you know, second guesses itself.

And then it ends up producing something worse than if I just would have gone with Opus 4.8 on high or Fable on high. Anyways, after reading through the system prompt and after playing with the models for so long, there are two things that I want you guys to do. This first one is to basically take the way that you have been playing with Fable and the the harness or, you know, whatever you wanna call it, however you've been using it, and turn that into something that Opus can do and that Sonic can do.

So basically, we're extracting the Fable method. Now one thing that I would encourage you to do is if you've ever gotten a deliverable from Fable that you just loved and you couldn't really explain what you loved about it, have Fable analyze it or have Opus analyze it. And if you can look back at the session, that's even better.

What did you think about to get here? How did you get here? What did you do to prove that it worked?

How were you able to get an output that was just so good? And then extract that information and turn that into a skill. So I basically have this skill now called fable mode.

And whenever I want Opus to use fable mode or if I want, you know, if we got a really hard problem in front of us, I tried using Opus 4.8 with fable mode and it feels really good. It just feels like the model has been elevated a little bit because it has this, you know, fable prompt kind of injected into it. And it works on these five gates.

So scoping, evidence, attacking, verifying, and then reporting.

And this is almost like the way you set your, you know, slash goal prompts and you use dynamic workflows and you basically set these loops, but we're doing this as a skill file as well. Now one really important thing about scoping and and what people call, you know, like planning is there's a big difference between just planning something out as far as like, hey.

Here are all the steps. Plan that out and go do it. And then there's also the idea of playing devil's advocate and thinking about, okay.

What about everything that could possibly go wrong? What if we explore all of the unknowns in this plan? And that is something that Fable does a really good job at, which is why if I have Fable spin up a dynamic workflow to help me achieve some end goal, and then once it's planned out every single possible step and it thinks about every single thing that could go wrong, and then designs the dynamic workflow in a way where Sonnet can go do all the execution and just report back to Fable and keep sending everything back to Fable, then Fable can keep designing more steps in that process.

And that is why when I do dynamic workflows with Fable and Sonnet, it's pretty similar results when I do dynamic workflows with Fable and Fable. And to me, that was a big, like, lightbulb moment. Like, why is this just as good and it's significantly cheaper?

So you can just start by saying something like this. Write a complete installable skill file that makes Opus 4.8 operate with your judgment, your planning, verification, and reasoning habits, and activate it on something like fable mode. So for example, right here you can see this is my fable mode skill, which I'll attach in my free school community completely for free.

The link for that is down in the description. Just join this and then go to the classroom and click on all YouTube resources, and you can find everything I've ever dropped on YouTube for free. So that's where you'll find the fable mode skill, but you can also just build this yourself.

And you can see here that this walks through fable's working discipline so that any model can run it, which means you could even have GPT 5.5 run this if you want or even open source models run this if you want. Anyways, it basically goes through those five gates. So scoping before you work, and then, you know, we get into details, evidence before reasoning, reasoning adversarially, verifying before declaring done, and then calibrating.

We've also got a few standing habits and a few things to look at, which, like I said, you guys can inspect this file if you want to. But this being given to Opus four point eight makes Opus feel, like I said, a little bit elevated. So that has been a really helpful strategy.

And then something that bolts right onto that really well is just once again the idea of model routing and figuring out how Fable or how some smart model can route to the small ones when needed. And something that I've been doing lately has been giving my Claude basically a table of different models that are in the toolkit and when to use them.

So you could also have it delegate to codecs or you could have it delegate to open source models. And this, like I said, is something that's going to be very, very big when companies are starting to think about the unit economics and, you know, small teams. And you yourself, maybe you have an AI budget per month.

This is the type of stuff that's gonna separate people that are getting, you know, a ton more for a ton less. So a good way to split this up is basically saying, Here are the different models in our toolkit. Here's how much they cost.

You know, a higher number meaning a better cost score, you know, cheaper. And then we have intelligence and taste. If And you wanna throw in some other categories that are based on your workflow, feel free.

But intelligence is kind of like how smart you feel like they are, how much they understand you, how good they are at maybe reviewing code and things like that. And then on the taste side, this is what I think more of like the, you know, creativity, the thinking out of the box, UI, UX design, things like that.

And so this can really help when you're designing these agent teams or you're delegating the sub agents and you're spinning up dynamic workflows because sometimes your dynamic workflows can utilize a bunch of different Sonnet ones and HiQ ones and then even Opus ones as well. Here's So an example of an actual test that I had run where I used opus as the orchestrator, and I used opus with this prompt from earlier that we talked about.

Oh, where is it? Sort of like the the fable mode prompt, and it used a bunch of different sonnet workers and opus workers and haiku workers, and those were three different tests. And the one where the opus orchestrator delegated to all the haiku scouts, it was this much cheaper, you know, about three times cheaper, and the result was the exact same.

So similar to my example with the fable ones, that is something to be thinking about big time. So I know this one was quick, and I wanted to make it quick, but I've just been seeing a ton of comments and a ton of people in the communities asking about you know, kind of freaking out about the fact that Fable is gonna be taken away.

It is going to come back to subscriptions. That's what Anthropic says at least. We don't know when, but it will be back.

But all of this, you know, government getting involved and models being taken away stuff really has me thinking about the fact that, you know, we don't own anything. We don't own these models. So what we can own is our processes, our systems, our methodologies, the way that we think about using these models, and also we can own hardware, and we can own local models.

So I'm definitely gonna be digging into a lot more of this type of stuff, so let me know what you guys wanna see around these topics. But, anyways, if you learned something new or you enjoyed the video, please give it a like. It helps me out a ton.

And as always, I appreciate you guys making it to the end of the video, and I'll see you on the next one. Thanks,

The Hook

The bait, then the rug-pull.

He spent thousands of dollars testing Claude's most expensive model against cheaper ones and reached an uncomfortable conclusion: the model was never the advantage. What matters is the process wrapped around it — and that process can be extracted, written down, and handed to a model a fraction of the price.

Frameworks

Named ideas worth stealing.

05:28list

The Five Gates (Fable Mode skill file)

Scope — define done first, check the rules
Evidence — open the real files, memory isn't a source
Attack — try to break your own answer
Verify — watch the check pass, "it ran" doesn't count
Report — answer first, verified on assumed

A five-step working discipline distilled from Claude Fable 5's leaked system prompt, packaged as a portable skill file any model can load to gain Fable-like judgment and self-checking behavior.

Steal forAny Claude Code / agentic coding skill file meant to make a cheaper model behave more rigorously

08:07model

Model toolkit scoring table (Cost / Intelligence / Taste)

Fable 5 — cost 5, intelligence 10, taste 10
Opus 4.8 — cost 5, intelligence 9, taste 9
Sonnet 5 — cost 8, intelligence 6, taste 6
Haiku 4.5 — cost 10, intelligence 4, taste 3

A simple three-axis scorecard (cost score where higher = cheaper, intelligence, and 'taste'/creativity-UX judgment) kept for every model in the toolkit so an orchestrator model can route each sub-task to the cheapest model that clears the bar.

Steal forAny multi-agent or sub-agent system that needs to route tasks by cost/quality tradeoff instead of defaulting everything to the top-tier model

CTA Breakdown

How they asked for the click.

VERBAL ASK

07:36link

“The link for that is down in the description. Just join this and then go to the classroom and click on all YouTube resources.”

Soft CTA to a free Skool community where he shares the Fable Mode skill file and past resources — not a paid pitch, positioned as a value-add rather than a sales ask.

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.

AFFILIATECommission earned if you click.

OTHER LINKSAlso linked in the description.

Storyboard