Big Idea

The argument in one line.

Claude Fable 5 costs twice as much as Opus 4.8 and overthinks by default, so the only sustainable strategy is to fence it to the 20% of work where raw intelligence matters — planning and final review — and run the execution bulk on cheaper models.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You hit your Claude Pro or Max limit within hours of Fable launching and want to keep using it without upgrading.
You code in Claude Code and want to stretch sessions without losing output quality.
You run a small business or agency and need to treat AI token spend like any other operating cost.
You want the Anthropic-official framing on agentic loop workflows before building your own.

SKIP IF…

You use free-tier Claude for casual chatting — this is aimed at heavy power users paying per token.
You want model benchmarks; this is pure workflow and cost-management advice.

TL;DR

The full version, fast.

Claude Fable 5 is priced at $50 per million output tokens — 2x Opus 4.8, 4x Sonnet — and compounds cost by running longer reasoning loops than any prior model. The answer is the 10/80/10 rule: use Fable for the first 10% of a task (architecture/planning), Opus 4.8 for the 80% execution middle, then Fable for the final 10% review. Layered on top is loop engineering, a pattern from an Anthropic engineer where Fable sets a goal, cheaper sub-agents (Haiku/Opus) do verification passes against a rubric, and the task stops automatically when the rubric passes. Practical guardrails: watch the Claude Code model selector (silently defaults to Fable), use low/medium effort (still beats Opus on max), delete overly rigid rules files, and set a hard spending cap before June 22.

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 – 01:14

01 · Hook — the best model, the worst bill

Fable nearly burned his limit in hours. Artificial Analysis shows $50/M output tokens — 2x Opus, 4x Sonnet, 1.67x GPT-5.5.

01:14 – 03:16

02 · June 22 — Fable leaves standard plans

Anthropic removing Fable from Pro/Max/Team on June 22, switching to pay-per-token above limits. Stakes for power users.

03:16 – 04:13

03 · What to actually use Fable for

Fable for high-stakes strategy, architecture, important decisions. Opus 4.8 stays the daily driver for creative work, tweets, copywriting, grunt work.

04:13 – 07:13

04 · The 10/80/10 System

Core framework: Fable for first 10% (plan), Opus 4.8 for 80% (execute), Fable for final 10% (review). Architect analogy. PDF available via newsletter.

07:13 – 13:47

05 · Loop Engineering — the Anthropic model

Lance Martin (Anthropic) article on loops. /goal sets goal + rubric, Fable spins sub-agents on cheaper models, self-corrects until rubric passes. /loop runs it on a recurring interval. Only in Claude Code.

13:47 – 17:35

06 · General token-saving advice

Don't chat on Fable. Use effort settings (low/medium beats Opus max). Delete overspecified rules files. Give the why not the what.

17:35 – 20:35

07 · Settings draining your credits

Model selector trap in Claude Code. /usage command to check limits. Hard spend cap before June 22. Cookie Monster GitHub widget for Mac menu-bar monitoring.

Takeaway

Six rules for using Fable without burning through your limit.

WHAT TO LEARN

The most powerful AI model available also overthinks by default — sustainable use comes down to model allocation, effort settings, and one habit of checking which model you actually have selected.

Fable 5 costs twice as much as Opus 4.8 per output token and tends to run longer internal reasoning loops — prompting it the same way you did older models compounds cost fast.
The 10/80/10 rule: use the most capable model for planning (10%) and final review (10%), then switch to a cheaper model for everything in the middle where iteration volume is highest.
Fable at low or medium effort still outperforms Opus 4.8 at max effort — start low and move up only if the task genuinely requires it.
Over-specifying instructions makes Fable perform worse, not better — it reasons well from a short goal statement, so trimming bloated rules files saves tokens and improves output.
Loop engineering (the /goal and /loop commands in Claude Code) offloads verification to cheaper sub-agents, so the expensive model only runs twice per task cycle — at the start and end.
The model selector in Claude Code can silently default to Fable — checking which model you are on before a long session is the single highest-value habit for controlling costs.

Resources

Things they pointed at.

01:34toolArtificial Analysis ↗

07:28linkLance Martin — Designing loops with Fable 5

19:19toolCookie Monster by David Pflugpeil ↗

Quotables

Lines you could clip.

00:26

“How can you actually use Fable without breaking the bank?”

Clean one-line hook, no context needed→ TikTok hook↗ Tweet quote

04:47

“Think of Fable as the architect. Even if you have the best builder in the world, if the plan is dodgy, the builder's just gonna execute on a dodgy plan.”

Vivid analogy that makes the 10/80/10 framework instantly intuitive→ IG reel cold open↗ Tweet quote

15:18

“Fable five on low to medium beats Opus models on extra high.”

Counterintuitive claim with immediate practical payoff→ TikTok hook↗ Tweet quote

17:35

“Hard problem, spawn agents to solve, have other agents pick the best answer, repeat.”

Clearest one-liner summary of loop engineering produced anywhere→ 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.

analogystory

00:00I've been using the new Claude Fable five for the past two days, and I've gotta say it's the best AI model that I've ever used. The things it can do is absolutely mind blowing. But there's one big problem.

00:11It's expensive, like, really expensive. In the first few hours of using it, I almost burned through my entire Claude usage limit because Fable is exactly two times the cost of Opus 4.8.

00:23And because it's so smart, it actually overthinks running loops which end up burning more tokens than any other clawed model. So how can you actually use Fable without breaking the bank?

00:34Well, today, I'm gonna show you the exact ten eighty ten system that I use that is based on the official advice from developers at Anthropic. Plus, I'm gonna show you the default setting that's burning people's limits.

00:46I've seen reports of people wasting thousands of dollars because they accidentally had this setting turned on. That one single change is gonna make Fable so much cheaper for you. As a business owner that runs a company of 30 people, it was imperative that I went down this rabbit hole to work out how to actually use this model the most efficiently or I'd probably be wasting tens of thousands of dollars in tokens.

01:06So no matter if you're using this for personal use or whether you're running a small business, I think the advice in today's video is going to apply to a plethora of users. And some of this advice doesn't just apply to Claude Fable, but will likely apply to many of the major models.

01:21Because a lot of these models are just becoming more and more expensive due to the sheer cost of compute. So firstly, let's establish the problem with Fable. If you go onto artificial analysis, you could see it's by far the most expensive model coming in at $50 per output token.

01:36If you look at GBT 5.5, it's $30 and it's twice as expensive as Claude Opus 4.8 max.

01:43It's over three times as expensive as Claude Sonnet, and it's four times more expensive than Gemini. So although it's the best model I've ever used, although it's extremely smart, it comes at a massive, massive cost.

01:55And this cost is actually going to be passed on to users, not just through their subscriptions, but also through the additional add ons because they can't afford to have their normal subscribers using Fable five because of the compute limits.

02:10So, essentially, what they're gonna do on June 22 is remove it for everyone's plans. Just give them a little bit of a taste of the good stuff now, and you're gonna actually have to pay per token above the limit in a couple of weeks. So I think this video will become even more important as time goes on because the real power users of AI will be forced to pay extra.

02:28So you really wanna make sure if you're being billed above your subscription that you're using this in the most efficient manner and actually in accordance with what Anthropic developers are actually telling you. Because I'll tell you right now, if you use this model like the past Opus models, you are going to absolutely burn tokens and you're not even gonna get effective results.

02:47This is a completely different model from other models and the way that you prompt and even the way you think about prompting has to be completely different. And today, I'm gonna break down why that thinking needs to be completely different. So before we get into that, how am I actually using Fable?

03:01That is a really important thing that you need to decide even before you implement more advanced strategies, you know, what you're actually gonna use the model for. So based on my initial tests and based on the literature that's out on Claude Fable and all of the benchmark results, we know for a fact it's very, very good at high level strategy, knowledge work, coding, data analysis.

03:22So anything where you need the best model possible, like a really important business decision, a really important personal decision, or if you're coding an in-depth application, Fable is going to be the best. But Opus 4.8 is still great for creative work.

03:34When I was writing tweets this morning, I didn't necessarily find Fable to be better. In fact, in many cases, I found it to be more verbose and the preexisting skills that I had already created were working just as well on Opus 4.8.

03:47So for any tweet writing, creative work, copywriting, and a lot of grunt work actually, I'm still gonna be using Opus 4.8. For most of my day to day chatting, I'm still gonna be using Opus 4.8.

03:57For the bulk execution task, which I'll get into the exact framework behind that in a second, I'm still gonna be using Opus 4.8. Where I am going to use Fable is on any truly difficult task and any task where I demand perfection. If there is something important for you that demands perfection, you may want to consider using Fable.

04:15And to save tokens, you may want to consider running a system like this, which is the ten eighty ten system. This is the very simple coding framework that I came up with based on the official literature from anthropic developers to help you save money and maximize your subscription while still getting the good parts out of Claude Fable.

04:33And it's quite simple. It's basically the first 10% of any task. So whether this is a workflow, whether this is an application you're coding, or whether this is a loop, which we'll get into later, which is one of the new ways that you need to be prompting Claude Fable.

04:47I'm basically running a split approach where the first 10% is planned by Fable. So if I'm working on a build, Fable does the planning to make sure that the architecture is perfect. It's the smartest model.

04:57So I want the smartest model doing the initial planning. Think about, you know, when if you would build a house, You know, you need a great plan. You need the architect to have great drawings or you won't get a great outcome.

05:08Even if you have the best builder in the world, if the plan is dodgy, the builder's just gonna execute on a dodgy plan. So think of Fable as the architect that can architect a proper structure for for whatever task you're doing or whatever workflow you are doing. So that's where I use Fable.

05:22For the actual execution, this is where I'll use a model that is slightly cheaper or more cost effective, Opus 4.8 or potentially Haiku if I'm doing real grunt work. And this is the execution layer that is actually the thing that burns tokens.

05:36If you actually look at where most people burn tokens, it's not in the planning. It's not even in the the critical review phase at the end. It's in the back and forth and the execution and, you know, all of the the minor stuff that crops up in the middle chunk of completing a task.

05:51So by using a cheaper model during the execution phase, you get the best parts of Fable, which is its strategic brain, and you don't have to deal with the token bloat. So this approach and this mindset going into any sort of build, I think, going to be imperative when it comes to using Fable. And if you actually use loops to execute workflows on your behalf, this is going to be a very effective approach that it can actually do automatically for you.

06:16And then, of course, the final 10% are reviews. So at the very end of a build, at the very end of a workflow, you can use the smartest model, which is Fable to review the build or review the workflow to make sure that the output is solid and in line with the initial architecture. So this is the ten eighty ten approach, which by the way, if you access the link in the description below, I wrote a full document based on the exact approach that you can actually take and just drop straight into Fable.

06:43So whenever you are doing something, Fable is gonna be a 100% sure that you wanna follow this approach. It outlines the exact way that it should think in order to give you the right outcome whilst not burning a bunch of tokens. So if you just join the newsletter below and then join the Instagram community in the pinned drive, you'll find the link to a PDF which you can just drop straight into Fable, which has this ten eighty ten approach built into it.

07:06So that's the basic framework of how I'm approaching work on Fable. But what do the official Anthropic developers have to say about using Fable?

07:15Well, that's where something brand new comes in and that is loops. Loops are the new way to use Fable. I'm gonna be doing a full video on this, but I'll give you a broad overview today and explain how it works.

07:27Lance Martin, who currently works at Anthropic, did a full article on how to design loops.

07:34And he's actually said, so this is kinda crazy, people inside Anthropic are saying that it's changed the way that they've worked at Anthropic. Mythos class models like Claude Fable five have changed the way we work. I wanna share two tips for getting the most out of the models.

07:47If you're gonna listen to anyone, listen to the actual coders who built this at Anthropic. Now the advice he gives is firstly to use self correction loops, and secondly, to optimize your memory to serve these loops.

08:00But I've basically taken this whole article and I've broken it down so you can really understand what loops are. So the way you need to think about loops is this basically being a change in the way that you used to have to prompt these AI models versus how you need to engineer workflows now. So the old way was basically you would prompt, Claude would answer, then you would check the answer, then you would re prompt and you'd respond again, and then you just keep prompting again and again and again, manually taking the data into account to get Claude to give you the outcome that you eventually wanted.

08:31So, essentially, you would be the loop. With loop engineering, you essentially give it a goal upfront.

08:38Fable will run a system and spin up sub agents based on this goal. It will use verifier grade, so it'll use a separate model similar to the approach that we discussed before. So, know, this execution piece, Fable's not doing all of that.

08:52It's actually spinning up sub agents on Haiku and Opus cheaper models to actually verify and make sure that it self corrects. Once it's confident that it has achieved the outcome, then it will finish the task.

09:06And if it's a repetitive task, it'll feed back to the agent, and it will go through this process again and again and again. And as Lance pointed out from Anthropic in his document, it actually remembers over time. So Fable really excels when it comes to memory.

09:21We could think about this as the outer loop that spans across sessions. Claude writes to memory during a session, and those memories can be retrieved in future sessions. So so no matter what you do, it's going to remember, especially if you have a proper memory configuration like I've discussed in other videos.

09:37It's gonna remember what you've done. It's gonna use that to get better. So these actually become recursive self improving loops, which although they were possible on previous models, the reason why so many people at Anthropic speak about this now is because the new Mythos class models are so smart and intelligent at working in this way.

09:55They're much better at taking information into account and self improving, and they're much better at identifying when a job is properly done. And that extra step up in intelligence is making these loops much more viable, whereas in the past, you know, there'd be a lot of broken loops.

10:12And the fact that Anthropic is literally telling you that you need to do this and there's been so much talk from top AI developers that this is the new way to prompt, that is now forcing me to shift a lot of my workflows, even, you know, my video production workflows, even my coding workflows, even my business development workflows that we use to find prospective clients and reach out to prospective clients.

10:31We're now building loops for the agents to self learn and self correct and self verify throughout the process. So if you're a bit confused, and I'm gonna do a full video on this soon, so do make sure you subscribe to the channel because I've got some great mythos slash fable content coming very soon. But if you're a bit confused, if you knew goal, basically, this was the command that you could enter into Claude code to achieve a final outcome.

10:54So instead of just telling it, hey, you know, write me a newsletter, you could go forward slash goal, build a newsletter that loads perfectly, every link works, sign up form sends a real test and stops after, you know, x amount of turns.

11:08So this was, you know, a singular goal. So you tell it to do something, it would go out. It's it'll if you use Fable, spin up agents, then it'll achieve that goal.

11:15If you use loop, this will actually rerun your prompt at a specific interval that you predefined, and then an agent will do that same process every time. But every time it does it, it'll get smarter.

11:27So the thing with goal is if you do it once, it doesn't have a chance to learn, and then you have to manually go forward slash goal to get it to do that specific task. If you go forward slash loop, it actually does it on its own at a predetermined interval, and then it will get better, learn from the results of something, and then do it again.

11:45So, you know, imagine the use cases here for video editing, which has been a big advancement of Fable. Imagine the advances here for scraping. Imagine the advancements here, something that I'm very interested in for AI trading because these loops will essentially enable the AI to back test strategies faster, look at the efficacy of a strategy, and then compare that to prior strategies, and you can run all of these loops automatically without, you know, having a static finish line like with goal.

12:12So how all of this actually helps you in terms of saving your limits and not spending too much money is that this will essentially spin up cheaper models to grade every single turn. As I showed you here, this self correction is not happening by the Fable model.

12:29It's happening by a cheaper model like Haiku or Opus. You can even specify and tell it what model you want to use. And then it stops the second the rubric passes, so you don't have any re prompts.

12:38The problem with the old way of prompting is that you'd have lots of re prompts. And every time you re prompted, you're using Fable. This new ten eighty ten approach is factored into these loops, and you can be very specific that you wanted to follow this approach.

12:52That's what I'm helping you do with the PDF down below as well if you access that and actually put it into Fable to ensure that you are running your prompts in the most efficient way possible. So you can actually use forward slash goal so you can state what task you wanna be done, give it the context, the success criteria, the constraints, and the checkpoints.

13:10And then at the end, you can use forward slash loop. You can say forward slash loop every fifteen minutes, one hour, one day. So you could say, you know, scan the market or scan my x feed or, you know, create this edit of this specific clip every single day or, you know, scan my watch list of these these trading view stocks every single day at 9AM, and then it could have a recursive loop to do a specific task over and over again.

13:32So basically creating energetic workflow inside Fable. Fable is very, very, very, very, very good at this stuff. Much better than the previous models.

13:42So those are loops. I'm gonna show you some live examples. I'm gonna do a full breakdown in a future video, so make sure you're subscribed.

13:47Now let's move into some general advice on how to save money on tokens because although this will save you a lot of money and make Fable more efficient, there are some really small tweaks that you need to do to make your model even cheaper. And, also, there are some common pitfalls that you're likely going to fall into if you're not aware of it that can cost you hundreds or even thousands of dollars in the future, especially if you end up paying as you go once these models are no longer included in your subscription on the twenty second.

14:17So the first thing to establish is that you can't prompt Fable the same way you prompt other models as I've made very clear. Use Fable for your hardest tasks. So, obviously, the loops are great for using that 10 eighty ten rule.

14:30But generally speaking, the way you wanna think about Fable is you wanna use it for your hardest task that requires the most processing and that requires the smartest model possible. Anthropic themselves literally say that easy tasks are a waste of tokens. So don't chat on Fable.

14:46Now a trap that you may fall into is when you open the Clawd app, it might default have Fable up. Just be very cognizant of what model you actually have selected because I think it could be very, very easy to accidentally use Fable without even realizing that you're accidentally using Fable. Because they're trying to incentivize people to test the model, but you may be chatting to it without even noticing.

15:04I even had that happen to me this morning. So you have to be very careful about that because you can just blow past your limits doing that. The other thing that's important, especially if you are coding on Claude code, is controlling the depth with the effort setting, not bigger prompts.

15:18So Anthropic has actually said that Fable five on low to medium beats Opus models on extra high. So, you know, with Opus, you know, we're also competing against Codex, so we wanted to make sure that we're on extra high to get the best possible output. You don't need Fable on extra high.

15:32Medium and low are still gonna do a better job than Opus. So if you're coding, you still want a better coding model, but you don't wanna exert a bunch of tokens, maybe start on Medium, see how it goes, see if you're getting errors, see if it's adept to solve the particular problem that you're trying to solve, which it likely is if it's basic stuff like my personal OS guides, lot of my training bots, that will all be fine on low to medium on Fable.

15:53And only if you want something more advanced and, you know, if you're noticing it's running into issues, then jack up the model to high and to max. But, you know, move up instead of just defaulting to max at once. The other thing you need to do is delete old rules and skills.

16:07So instruction following is now so strong on Fable that if you over specify, it can make it worse. So, you know, the folder configuration that we've discussed in videos with memory and instructions, make sure your instructions aren't too rigid because Fable is actually very good at following very short prompts.

16:25You can even, you know, be quite vague. You don't need to be as specific as you were before, and it will it's so smart. It can reason and work out what you're trying to say.

16:32So if you have too many constructions and and instructions and too much context, it can actually make the model more bloated and it has to process more data to give you an outcome. Thus, you're going to burn more tokens.

16:44This is also very important. Give it the why and not the what. It does better work when it knows the goal and not just the task.

16:50That's very, very true, and that's also where the forward slash goal and the forward slash loop, if you have interval strategy comes in. And by the way, just so you guys know, and I'm gonna do a future video on this, this only works on Claude code. So if you're doing forward slash goal, you basically enter a goal.

17:08You know, you enter the goal, you enter the context, you enter, you know, what you wanted to verify, and then to enter the interval, you go forward slash loop, and then you enter the interval that you want after that. I'm gonna do a full video on it, but it only works on Claude code.

17:20So if you're trying to use these commands on the main chat and you're confused, that is likely why you are running into that issue. And I really like the simplification of of loop and goal as well.

17:29Hard problem, spawn agents to solve, have other agents pick the best answer, repeat. That is essentially you know, if if you're feeling overwhelmed about looping, that is essentially all it's doing.

17:40And if you're able to do this, as Enzo points out here, you'll be able to see the full power of Fable. Now this is a video about saving money. So let me tell you the two things to be aware of that could drain your credits.

17:51One, we've already touched on. Make sure you're very cognizant of what model you are actually selecting in code.

17:59So it's down here. Make sure that you select Opus 4.8. You want 4.8, only use Fable for the really, really hard stuff.

18:06Something you can do is actually check your usage in Claude code. So you can go usage, click enter, and you could see your exact limits.

18:14This is very important to keep checking so you know when you're running into your limits. The other thing that I recommend doing is going into settings on your application and clicking usage or you can go onto the website and click usage. This will show you about your current session limits and your weekly limits, especially once you start plugging into your credit card.

18:32If you don't set a hard limit, which I recommend doing if you do end up adding your credit card. So that will only happen once you turn usage credits on. I strongly recommend keeping an eye on it because someone actually ran into I saw an x, like, a thousand dollar bill that happened because he hadn't set a a cap on his usage limit.

18:50So in the future on the twenty second, if you do wanna expand your Fable usage because you find it really impactful for your business or for yourself, just make sure you're keeping an eye on the usage limits and even better setting a hard cap on spending. There's even a setting where where you can just top up pre top up instead of having unfettered access to your card.

19:07It it depends on your jurisdiction and the setting that you pick. But those are some of the traps that you need to avoid if you don't wanna absolutely burn money on tokens. And then I've even seen some cool workarounds for token monitoring, like David, for example, and it's in his GitHub here, github.com.

19:22You can pause the video and read this out. David, p f l u g p e I l, forward slash cookie hyphen monster. If you go into that GitHub, you can actually download a little widget in the command bar on your Mac if you're on Mac, which will show you your exact session limits across the different models as well, which I think is quite handy because if you don't wanna keep manually doing forward slash usage or you don't wanna keep clicking, you know, settings and then going into usage, if you just wanna be able to see it at all times, it can actually show up on your command bar, which is also helpful as well for extra monitoring.

19:56So a lot of this is also about monitoring and awareness because that is going to make you obviously more aware of how many tokens you are burning and thus you can tweak your behavior if necessary. But I think some of the tips that I gave you today are gonna help a lot. I'm gonna do a full video on loop with some real examples.

20:13I'm gonna show you the real power of that using Fable in future videos, but I want to get this out to you today to make sure that you're not burning money and you're actually getting the most out of your subscription. So make sure you subscribe. Link below to the full PDF if you wanna put it into Claude to make your coding approach or your workflow approach even more efficient.

20:31And I'll see you in the next video. Have a lovely rest of your day. Peace out.

The Hook

The bait, then the rug-pull.

Two days into Claude Fable 5 and he had almost burned through his entire usage limit. The model is extraordinary — and at $50 per million output tokens, it will quietly empty your wallet if you prompt it the same way you did every model before it.

Frameworks