Big Idea

The argument in one line.

Fable 5.2 only earns its double price on judgment-heavy review work, so run everyday work on Opus at medium effort with cheap helpers and call Fable in at the end.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code daily on a subscription and keep hitting your weekly limit faster than you expect since Fable 5.2 arrived.
You run sub-agents or agent teams and have never checked which model and effort level each helper actually runs on.
You want a concrete rule for when delegating work to cheaper models saves money and when it quietly costs more.

SKIP IF…

You only chat with Claude in the app and never touch Claude Code, sub-agents, or skills.
You are on pay-per-token API billing with a fixed model choice and no sub-agent setup to tune.

TL;DR

The full version, fast.

Fable 5.2 is the strongest model but costs about twice Opus, so using it for everything drains limits fast. The presenter tests Anthropic's guide and lands on a system: default to Opus 5.5 at medium effort, put file search back on Haiku because the built-in search agent now inherits your session model, keep sub-agents on cheap models with detailed briefs, and delegate only when a job is big enough that cheap middle work outweighs Fable's planning cost. A test showed Fable orchestrating cost five times more than Fable alone on a small job. Fable is best spent on the final 10 percent review, and on auditing your skills whenever a new model ships.

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 – 00:47

01 · Intro

Fable is the best model but twice the price of Opus. The video tests Anthropic's guide and promises a one-prompt setup.

00:47 – 02:55

02 · What you're paying for

Per-million-token pricing across Fable, Opus, Sonnet and Haiku, and how that shows up in subscription limits and resets.

02:55 – 05:45

03 · The hidden subagent cost

The built-in search agent now inherits your session model up to Opus. Replacing it with a Haiku Explore agent cuts search cost about 50 times.

05:45 – 08:30

04 · Choosing your effort level

Medium is the default for good reason. High costs nearly triple for eight points, and sub-agents inherit your effort setting.

08:30 – 09:57

05 · Is Fable worth it?

Same onboarding audit on both models. Opus found all seven planted issues for 20 cents, Fable also found eight extras for 35 cents.

09:57 – 12:40

06 · The 10-80-10 rule

Fable plans and reviews while cheap models do the middle. A channel-audit test showed orchestration costing about five times more than Fable alone.

12:40 – 15:22

07 · When should you delegate?

Delegation only pays off when the middle of the job is large. Small jobs should run on one model in one sitting.

15:22 – 17:00

08 · How to brief cheaper models

Cheap models guess when prompts are vague. Have Fable write a full brief with goal, rules and fallback before handing off.

17:00 – 18:59

09 · The new model audit

On every new model release, have it audit all your skills, prompts and workflows, fixing what it is sure of and listing the rest for review.

18:59 – 19:18

10 · Outro

Recap and sign-off.

Atomic Insights

Lines worth screenshotting.

Fable 5.2 costs $8 per million input tokens and $40 per million output, exactly double Opus, so a job that costs $0.80 on Opus runs about $1.60 on Fable.
The built-in search agent now runs on your session model up to Opus, so even in a Fable session every file lookup is billed at Opus prices instead of Haiku.
Pointing a custom Explore agent at Haiku made the same search about 50 times cheaper, under one cent instead of roughly 15 cents.
Custom sub-agents have no model cap, so if you do not set one they copy your main model and a file-reading helper gets billed at full Fable prices.
Sub-agents inherit your effort level by default, so one session left on high makes every helper it spawns think on high too.
On Anthropic's coding benchmark, Fable on high scored 45.3% at $6.72 per task versus 37.6% at $2.48 on medium: eight points more for nearly triple the cost.
Going one effort level past high can cost as much as moving to the bigger model, so try Fable on medium before Opus on max.
Opus and Fable both found all seven planted problems in a client onboarding audit, but Fable surfaced eight extra issues for about 15 cents more.
Pay Fable for judgment: do the work on a cheaper model and spend Fable on the final 5 to 10 percent review.
In a channel-audit test, Fable alone cost 58 cents, Opus alone 51 cents, and Fable orchestrating cheaper helpers cost about $2.50, nearly five times more.
Delegation is like hiring a project manager to make a sandwich: the briefing, waiting and checking cost more than doing it yourself on a small job.
Fable's planning and checking cost is fixed, so delegation only pays off when the middle of the job is large, like reading 50 transcripts where Haiku reads at about a eighteenth of Fable's price.
When Anthropic told Fable what a sub-agent costs and when one is worth it, it spawned roughly half as many sub-agents.
Prompt engineering is mostly dead for smart models, but cheaper models take your words literally and guess the rest, so they need a full brief.
A good brief for a cheap helper states the goal, what done looks like, the rules, what not to do, and what to do when stuck.
Skills written badly by a weaker model drag down every model that uses them, so having Fable rewrite your skills lifts Opus and Sonnet too.

Takeaway

Spend Fable only where judgment matters

WHAT TO LEARN

Default to Opus at medium effort with cheap helpers, delegate only big jobs, and pay for Fable on the final review.

02What you're paying for

Fable costs twice Opus per token, so a subscription limit drains twice as fast on equal work. Reserve it for work where extra judgment changes the outcome.

03The hidden subagent cost

The built-in search agent now follows your session model up to Opus. Replace it with a Haiku Explore agent to cut search cost about 50 times.
Custom helpers copy your main model unless you set one. Give search, read and summarize helpers a cheap model, since mistakes there are easy to spot.

04Choosing your effort level

Check your effort level before each session. Medium is the default, and helpers inherit whatever you left set, so a stale high setting multiplies across every helper.
Eight extra benchmark points on high cost nearly three times as much. Try a bigger model on medium before pushing effort past high.

06The 10-80-10 rule

Plan and review on Fable, let cheaper models do the middle, but only when the middle is big enough. Small jobs cost more once planning and briefing are added.

08How to brief cheaper models

Write full briefs for cheaper models with goal, finished state, rules, and a stuck fallback. They guess when told too little, and guesses cost accuracy.
Keep Fable's final review pass even on 90/10 splits. It is the most valuable part of the workflow.

09The new model audit

When a new model ships, have it audit your skills, prompts and workflows. Better skills lift every model that uses them.

Glossary

Terms worth knowing.

Sub-agent: A helper copy of Claude that Claude Code starts in the background to do a side task, such as searching files, so the main session stays focused.
Effort level: A setting from low to max that controls how long a model thinks before answering. Higher levels use more tokens and cost more.
Explore agent: Claude Code's built-in search helper that finds files and lines in your project. You can replace it with your own version pinned to a cheaper model.
10-80-10 rule: A delegation pattern where the strongest model plans the first 10 percent, cheaper models do the middle 80 percent, and the strongest model reviews the last 10 percent.
Skill: A saved set of instructions Claude follows for a particular task. A poorly written skill makes every model that uses it perform worse.

Resources

Things they pointed at.

00:00linkAnthropic model selection guide

01:47channelBoris Cherny, creator of Claude Code

02:30channelAndrej Karpathy

08:02tool/usage command in Claude Code

17:00toolAlex Albert's new-model audit prompt

18:05productFable For Cheap skill (Skool free asset library) ↗

Quotables

Lines you could clip.

09:30

“Pay Fable for judgment.”

short, memorable rule of thumb→ TikTok hook↗ Tweet quote

11:46

“Think about it like hiring a project manager to make you a sandwich.”

vivid analogy for delegation overhead→ IG reel cold open↗ Tweet quote

15:35

“The less you tell them, the more they guess, and the more they guess, the more they're wrong.”

tight explanation of why cheap models need briefs→ 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.

Fable 5 .2 is the best AI model in the world right now, topping every single benchmark. But it's almost twice the price of Opus. So if you use it for everything, you're going to chew through your limits super fast.

But the good news is Anthropic have just published their own guide on which model to use and how to keep costs down. So in this video, I'm going to be testing their advice as well as compiling other advice from the cloud creator, Boris Cherny, and other AI figureheads like Andrej Kapathy to finalize the ultimate system for using Fable 5 .2.

for cheap. And at the end of today's video, I'm going to be giving you the one prompt setup guide that you can apply to your cloud code right now to utilize this system. And it's the same system that I'm using internally in my AI tech and media company with over 30 employees.

We're streamlining all of our workflows with Fable 5 .2. So to understand how to save money on Fable, you first need to know what you're actually paying for. So like all LLMs, you essentially pay per token in and token out.

And Fable costs $8 per million tokens in and $40 per million tokens out. Opus is exactly half that price at $4 .20, Sonnet at $2 .10, and Haiku is just $0 .10. So there's a massive spread between the cheapest Claude model and the most expensive Claude model.

To put that into real terms, a typical job of about 75 ,000 words in and 15 ,000 words back costs around $1 .60 on Fable $0 .80 on Opus and just $0 .02 on Haiku. Now, if you're on a subscription like most people are, you won't see those dollars, but you will fill them in your weekly limit.

Because on Max and Ultra, Fable is included for up to half your normal weekly limit, and it burns through that limit a lot faster than other models. Whereas on Pro, it's only included for launch week until the 16th of November, unless it's extended. And then after that, it's going to run on paid usage credits where you'll be paying at the rate I just described.

Boris Choney, the creator of Claude Code, said that there are going to be Fable resets depending on how long you've been subscribed to Claude. So the longer you've been subscribed, you will get more resets. But a reset only buys you more time once.

So the way to actually get around the cost is to use Fable for the only part of the job where it's genuinely better. and then let the cheaper models handle everything else. That way you can get Fable -level results without paying Fable prices for an entire workflow, which is the system I'm going to be breaking down today.

And that's exactly how Manthropic has set it up in their own guide. They say to start most of your work on Opus at medium effort, and then only bring in Fable for the hardest projects when Opus on high still isn't good enough, or... for long jobs that you are running continuously on your own, where Fable shines.

And the people who use Claude the most, for example, Andre Kapathy, who is now at Anthropic, are doing the same thing. He says his own default in Claude code is Opus 5 .5 on medium effort. And for most tasks, I also use Opus 5 .5 on medium effort, apart from the areas where Fable really shines, which I'll get into very shortly.

Okay, so step one, which is simple, is to stop paying fabled prices for everyday work. But even if you've done that, there's actually a hidden cost not many people know about on Claude, because part of what you're paying for isn't even the model you picked at all. That's because Claude code doesn't do all of the work in your main model.

So whenever it needs to look something up, it hands off that job to a helper agent called a sub -agent, which is basically another copy of Claude running in the background, so you don't really see what's happening in real time. So the one you probably use all the time is the built -in search agent. So every single time Claude needs to find a file or a line in your project, it sends the search agent off to look, and in one session that could happen dozens of times depending on what you're doing.

Now obviously searching through files is one of the simplest job that there is. You don't need the smartest model in the world to find a file, which is why the search agent always used to run on Haiku, the cheapest model Anthropic makes. But in Anthropic's new guide, it says that's changed.

The search agent now runs on whatever model your session is on up to Opus. So even if you're on Fable, every single search is being executed by Opus. So you can see what it costs here.

I'm in a Fable session right now, and I've asked it to find something in my files. And if you look at the task list, the search agent is running on Opus. That one search cost me about 15 cents, and that's just to find a particular piece of information.

And that's even before Claude has done any of the actual work. But what you can do, and it's one of the fixes that can save you a lot of money over time, especially if you're doing research -heavy work, is actually... putting search back onto Haiku, where it used to be, and Anthropix's official guide tells us exactly how.

So you can make your own agent called Explore, which replaces the inbuilt one, and in its file you add a line that selects the model, Haiku. And in my Freeschool community down below, you'll get the one prompt setup guide which can set all this up for you, including the rest of the setup which we're about to get into. Then I run the exact same search again.

This time it's on Haiku, it finds the same thing, and it costs less than a cent. So it's about 50 times cheaper for the same answer. And with the helpers that you make yourself, it's actually worse.

Anthropic capped the built -in search agent at Opus, but your own helpers don't have that cap. So if you don't tell them which model to use, they just copy your main one. For example, if you're in a Fable session, that means a helper that only reads files is now being charged at full Fable prices.

So the rule that I follow is that if a helper is only searching, reading, or summarizing, it uses a cheap model. Because a mistake there is easy to spot. And all of these fixes are built into the setup I'll give you in the school community for free down below.

If you look at this tweet, which Boris Cherny, the creator of Claude, reposted, which is from ClaudeDevs, you can see that Anthropic are clearly thinking about things the same way. Because they actually just moved the status updates for background sub -agents onto Haiku, and now each update costs about 99 % less. So you can bring this same principle into your work.

So what we just discussed is which model is doing the work. Clearly, that's a component that needs to be optimized to save money on Claude. But the second thing that's important in terms of optimizing spend is selecting the effort level.

So this is essentially how hard each model thinks before it gives you an output. So there are five levels, low, medium, high, extra high, and max. The higher you go, the longer the model thinks and the more tokens you pay for.

So Anthropic's official model guide actually suggests that before you move models, you try adjusting your effort level. Medium is already the default in the Claude app and in Claude code, so I leave most of my work on Medium. However, obviously make sure you're checking this before you go into a session, especially a long task because there's been times where I've accidentally set it on high from a previous session.

So just make sure you double check that because you can burn a lot of tokens not being aware of this. Now, high, of course, scores better than medium on most benchmarks. For example, on Anthropix Frontier Coding Benchmark, Fable on medium scored a 37 .6 % at $2 .48 per task, but on high, it scored 45 .3%, but it's at $6 .72 per task, so it's almost triple.

only for a small percentage increase. So you get eight percentage points more, but you pay nearly three times as much. And if you start going above the high effort setting, it just gets expensive really quickly.

Anthropix says that one more level past high can cost as much as using the bigger model. So there's an example where a user ran the same scene prompt on high and then on max, and it cost them $24 versus $88 for max. So sometimes you're better instead of going from Opus High to jump onto Fable Medium, depending on the type of task that you're doing.

Now, Boris himself has actually commented on heavy sessions. Boris noticed this trend when he was looking into why people were hitting their limits faster than they expected. He said that most cases came down to one or two heavy sessions, like on extra high effort, running agent teams or long workflows.

So this ties back into the helper point I made before, because he says that sub -agents inherit your effort level by default. So if your main session is set on high, every helper it starts is also on high. And if you're paying for that extra thinking on each single search, you're going to rack up a huge token spend.

So his advice is actually to keep an eye on the effort level you're on. And if you want to see where your tokens actually went, you can use forward slash usage, which breaks down your usage for each session. So you can actually run audit and then use this to develop a skill.

So before you go into a big task, you can tap into your skill, which is already optimized. Obviously, I'm giving you that initial optimization today in the asset down below. Now you know about effort levels, let's discuss when Fable is actually worth it.

So assuming Opus is your default, using cheap helpers and medium effort, you've already cut out the majority of the waste. But that leaves the real question, which is, when is Fable actually worth paying twice as much for?

Well, Anthropic's suggestion is basic, but it's just to test it. Run the same task on both models and compare what you get and what it costs. So that's what I'm doing.

For example, right now, I'm giving Opus and Fable the same prompt, which is to audit my client onboarding process and tell me what's broken. And I've actually hidden beforehand several real problems in the process so I can actually score how well each one does. You can see now Opus found all seven, costing me about 20 cents.

Fable also found all seven, but it cost around 35 cents. However, where Fable does better is everything around the seven fixes. So it actually found eight more issues that Opus didn't, like the fact there's no step to remove a client's access when they leave.

Although to be fair, Opus also catches two things that Fable missed. Still, Fable is up net six if you want to use that as a benchmark. So for most of your work, Opus gets you the same core answer for less, which is obviously why Anthropic is suggesting you start there.

But on something... like this where you're reviewing something important where that extra analysis or that extra judgment is going to move the needle then in my opinion in this example it's worth the 15 cents so what i recommend is that you actually pay fable for judgment so you can complete your work on opus and then you can put fable in for that final five to ten percent review to flag any issues or concerns so for example i'll do this with scripting i'll script on opus and then at the end i get fable to review the script for errors and make improvements i'll also do this for I will use it as an auditing agent and then I'll let the cheaper models do everything else There's actually an old designation rule which I've spoken about on the channel called 1080 10 Which is essentially where the leader in this case fable sets the direction or the planning at the start the team So the cheaper models do all the work in the middle and then the leader comes back at the end

to check the work that the cheaper models did. And Anthropics Guide says to set it up the same way. It puts a leading team of subagents on Fable, and it puts the subagents that search files, read logs, and summarize on Haiku or Sonnet, and then Fable can come back in for the review.

You don't have to build anything to get Claude to do this, you can just ask Claude Coe to plan the job and hand the pieces to cheaper helpers and spin them up by themselves. So if you look at my agents running right now, you've got Fable at the top orchestrating, then you've got Opus, Sonnet and Haiku doing the work underneath and Fable is going to review the work that these models did.

But here's actually the caveat, something you might not expect when I was running this in testing. Because on paper you might think, okay... This should be a lot cheaper and in previous models it was because Fable's only now doing a fifth of the work.

However, I ran a test which got me some really interesting results. So I ran a test to audit my channel. So look at my best 10 videos compared to my median and plan my next videos based on why those videos succeeded.

Now when I ran everything on Fable, it cost 58 cents and took about a minute. But when I ran everything on Opus, it cost 51 cents and took about a minute and a half. Makes sense, okay?

In this case, it's worth it to use Fable. But when Fable was orchestrating, so managing the other models like 108010, and then handing off to other models, and then reviewing the work, it actually cost 250, which was five times the amount almost that the pure Fable run took. Which at first when I saw that, I thought something had gone wrong.

So I tested it again. But then what I worked out is it's all down to efficiency. So think about it like hiring a project manager to make you a sandwich.

They have to work out what you want, write the brief, hand it to someone, wait for them, check the sandwich, relay the feedback back, check it again, which actually just costs more than making it yourself. It's the same principle for a lot of delegation. And that's exactly the same thing that's happening in this case.

So to delegate, Fable has to plan the job and write a brief for every single helper, read every report that comes back. And all of that management is being done by the most expensive model. And on top of that, every helper starts from scratch and rereads the files every time there's a review.

So you're just paying for the same setup over and over and over. And if you think about how models work, tokens in and tokens out, that's why doing this can actually cost you more. And Anthropic actually, in their official docs, I dug into this even further, says this in their official guide.

So for work that fits in one session, they say that the planning and merging cost more than they save. Now, I know what you're probably thinking at this point. Miles, does 108010 actually work or not?

Well, the answer is that it does, but only when the job is big enough to be worth managing. So the reason for this is that Fable's planning and checking is a cost that you're going to have to pay either way, whether the job is big or small. What changes is how much work is in the middle.

That's the part that the cheap models are doing for a fraction of the price. So for example, on a big job, like going through 50 video transcripts, Almost all of the cost is reading, and Haiku reads for about an 18th of what Fable charges.

So the money that you save on all of that reading is far bigger than what Fable spends on managing it, so you end up coming out well ahead with $10 .80. But on a small job, like writing an email, fixing a bug, planning a video, even reviewing like 5 -10 videos, there's not enough to hand off, so you're paying for Fable to plan, brief, check, And there's not enough in the middle for you to actually save money on.

So that's exactly what happened in my test. It was just 10 videos worth of numbers, which is a minute of work for one model. So Fable spent more managing a job that only costs 58 cents to do on its own.

Anthropic, according to their official guide, actually tested this themselves. On a job searching for bugs across a huge code base, they had Opus, at the time, before Fable was out, leading a team of cheaper Sonnet workers. at a cost about a third less than Opus doing it on its own for the same result.

And the same logic applies to Fable. Whereas that was a big code base review. If it was one or two or three lines of code, it could actually just be more efficient to get Fable to do all of the work.

So here's how I actually use this in practice. If it's something one model can finish in a single sitting, Like my test or, you know, a little bit of work, I don't split it at all.

I'll just run it on the model that I want. So Opus 5 .5 on Medium or Fable if I think it really needs the judgment. But if it's a big job with lots of separate pieces, for example, a build where I'm building like a trading terminal or I'm building like a big landing page where, you know, I want the planning and taste, but there's enough work to do in the middle, that's where a structure like 108010 will pay off because Fable plans it, the cheap models do the grunt work, and then Fable checks the work at the end.

And to save you guys the hassle of needing to make that judgment call every single time you do a piece of work, I've actually written the rule into the setup prompt, which I'll give to you, which can help you define what work is worth handing off. Because Anthropic actually found that when they told Fable what a sub -agent costs and when it's actually worth using one, it started roughly half as many sub -agents.

So that judgment is also really important, making sure that you only split up the work when there's a job big enough to save you money. Okay, so now you know which model to use and now you know the system behind it Here's something that's extremely important that's underrated on the smartest newest models like fable 5 .2 Claude is really good at working out what you want And you know if you watch videos of mine, i've said that in many cases prompt engineering is dead That is 100 true for the smart models now, but where most people go wrong is on the cheaper models because cheaper models aren't as good at filling in the gaps as the smart models like Fable.

So Fable can often work out what you mean, even with little context, but Sonnet or Haiku would literally take your words and just guess the rest. So the less you tell them, the more that they guess, and the more they guess, the more they're wrong, essentially. So on the left of screen is how most people delegate, which is usually something like, research competitive videos and summarize.

That leaves the model to have to guess which competitors, how many videos, what counts as good, what's bad. And on the right is a brief written by Fable. It has the goal, what the end result looks like, the rules, what not to do, what to do if it gets stuck.

So I recommend before doing a big task, you can actually plan and orchestrate your prompt with Fable and then hand it off to the other agents to get work that's more tailored and specified towards your actual needs. That's also how the setup I'm giving you tells Fable to brief its helpers. So you actually use Fable's inference to tell the helpers how to act, meaning every job it hands out.

comes with a proper brief like this. But even if you give less smart models the perfect brief, they can still get things wrong, which is why that last 10%, the check at the end that Fable does, matters so much. And that is actually the most important part.

Even if you get the other models to do 90 % of the work and don't do 10, 80, 10, and you just do 90, 10, that last 10 % from Fable is still the most important part. Now you're ready for some real alpha. This one's an absolute banger.

Alex Albert at Anthropic has a habit I love. Whenever a new model comes out, He runs one prompt with the brand new model on each of his projects to find what's broken and fix what it's sure about and listing the rest for him to manually review.

So I've actually taken that further and I've had Fable, and this is true whenever a new model comes out, audit my whole setup, which is all of my skills. all of my prompts, all of my workflows, all of my code. Skills are basically just saved instructions that Claude follows for a particular task.

So when they're written badly by a less intelligent model, then every model that uses them does a worse job, no matter how good the model is. So if you get Fable to rewrite your skills, every other model will actually do a better job. This is also a hack if you want to continue using Opus, but you just get Fable to do one review of your code base, that's like extremely high leverage in my opinion.

Alright, so that's the whole system but now let me give you the Claude skill which I promised at the start because I've packaged all of this into a skill that sets this up for you. If you want to claim it, just go to the free school and then click on courses then free asset library. It's for free.

Once you've added the skill to your Claude, you just type it in, so forward slash fable for cheap, and it makes three changes to Claude code for you. The first is a search agent, which is the fix from earlier using Haiku, so your file search stops running on Opus. The second is the rule for when Claude should split up a job, so it only uses helpers when the job is big enough to save money, and each helper gets a cheaper model.

And then the third is a fable reviewer, which you can call in by asking for a fable review. So it only runs when you ask, So you're only paying for Fable's judgment on the really important jobs.

So if I run it right now, you can see that when Claude searches my files, you can see it's running on Haiku. And then when I give it a small job, it just does it itself without spinning up any helpers. And when a new model comes out, I can run my audit prompt and it goes through my whole setup, tells me what to fix without changing anything until I tell it to.

And as you can see in front of you, these are all of the commands that you can use using the skill. So hopefully today you, A, learned something about how AI actually works under the hood. Subscribe for more content like this.

And B, you can actually use this for yourself to get better outputs from Claude whilst saving money at the same time. I will see you in the next video. Hopefully you have a lovely rest of your day.

Peace out.

The Hook

The bait, then the rug-pull.

Fable 5.2 tops every benchmark and costs nearly double Opus. The presenter tests Anthropic's own cost guide and builds a system that keeps Fable for the moments it is actually worth it.

Frameworks