Modern Creator
AI LABS · YouTube

Fable 5.2: Nine Rules for Using It Without Burning Your Limit

AI LABS distills Anthropic's own talks and guides into nine rules for running Fable 5.2 in Claude Code: start cheaper, define done, and let Fable lead while smaller models build.

Posted
2 days ago
Duration
Format
Tutorial
educational
Views
49.3K
796 likes
Part of the collectionThe Claude Fable 5.2 PlaybookEvery Fable 5, 5.1 and 5.2 breakdown, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

Fable 5.2 is the strongest and most expensive Claude model, so you get the most from it by saving it for hard jobs, defining done up front, and setting Claude Code up so it can run unattended.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You use Claude Code daily and keep hitting your usage limit before a long Fable 5.2 job finishes.
  • You run multi-agent setups and want to know which model should lead, build and review.
  • You maintain a CLAUDE.md full of older never and important rules and suspect it is hurting newer models.
SKIP IF…
  • You only use Claude through the chat app and never touch Claude Code, effort settings or CLAUDE.md.
  • You want independent benchmarks. Nearly every number here comes from Anthropic's own launch material.
TL;DR

The full version, fast.

Fable 5.2 launched November 9 at $8 in and $40 out per million tokens, 20% cheaper than 5.1, and scores 49.1% on FrontierCode against 45.5% for Opus 5.5. It still costs twice as much per token as Opus, so the first rule is to start on Opus 5.5 and escalate only when it fails. The remaining rules cover writing a verifiable finish line, using Fable for big, 3D and hardest jobs, making Fable the lead agent over Sonnet, Opus and Haiku workers, running the lead on high effort, pre-approving safe actions, demanding readable final reports, cleaning old CLAUDE.md rules, and compacting before you switch models.

Free for members

Chat with this breakdown — free.

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

Create a free account →
Chapters

Where the time goes.

00:00 – 00:40

01 · Cold open

The problem of running out of limit, and the three Anthropic sources behind the nine rules.

00:40 – 04:30

02 · What Fable 5.2 changes

Price cut, speed, fewer tokens, FrontierCode 49.1%, fewer false refusals, clearer writing and self-review via subagents.

04:30 – 06:40

03 · Rule 1: Don't start on Fable

Start on Opus 5.5 and escalate only when needed, with time and cost comparisons.

06:40 – 08:40

04 · Rule 2: Give it a finish line

Verifiable done criteria, a high bar, and review by a fresh Fable session.

08:40 – 09:40

05 · Rule 3: Three jobs for Fable

Big jobs, 3D and game work, and the hardest problems.

09:40 – 12:40

06 · Rule 4: Fable leads, cheaper models build

The lead agent pattern, model-by-task routing and sub-agent cost.

12:40 – 13:30

07 · Sponsor: Brevo

Transactional email API sponsor read.

13:30 – 16:10

08 · Rule 5: Run the lead on high effort

Why Fable stops early and the two fixes, effort command or a CLAUDE.md block.

16:10 – 18:20

09 · Rule 6: Set up so it keeps going

Auto mode scope, reversible versus irreversible steps, and answering questions precisely.

18:20 – 20:50

10 · Rules 7 to 9 and close

Readable final reports, cleaning CLAUDE.md, switching models, and the community plug.

Atomic Insights

Lines worth screenshotting.

  • Fable 5.2 costs $8 per million input tokens and $40 per million output, 20% below Fable 5.1, but still twice Opus 5.5 per token.
  • On FrontierCode at extra-high effort Fable 5.2 scored 49.1%, ahead of Fable 5.1 at 32.4% and Opus 5.5 at 45.5%.
  • On the same one-line 3D prompt, Opus 5.5 finished in 27 minutes while Fable 5.2 took 58 and used about seven times the tokens.
  • A finish line only helps if it is checkable. Make the page good gives Fable nothing to verify, a pass or fail test does.
  • Fable keeps working until it hits the bar you set, so a low bar gets you an average result and a high bar gets you more.
  • A review by Fable in a fresh session that never saw the build catches problems Opus tends to miss.
  • Fable holds its position under pushback instead of changing its review to please you.
  • Fable as lead with cheaper workers cost about 1.7 times Opus alone on a small feature, because the lead rereads everything to check it.
  • Telling Fable what a sub-agent costs cut the number it spawned by about half in Anthropic's test.
  • Low and medium effort make Fable more likely to stop early with a list of what's left. High effort is the fix, and going past it adds cost for little gain.
  • A step that can be undone is safe for Fable to take on its own. A step that can't be undone should wait for you.
  • Lines with the word never in an old CLAUDE.md are treated by Fable as hard rules, so write ask before deleting files instead.
  • Switching models mid-session throws away the one-hour prompt cache, so the first message on the new model costs full price.
Takeaway

Save Fable for the work Opus can't do

WHAT TO LEARN

Fable 5.2 pays off when it is aimed at big or hard jobs with a clear finish line and permission to run, and wastes your limit everywhere else.

03Rule 1: Don't start on Fable
  • Anthropic's launch post says to start with Opus 5.5 and only move up when Opus on a higher effort setting falls short.
  • On the same 3D prompt Opus took 27 minutes and Fable took 58 with about seven times the tokens, and on a feature test Fable cost about 58% more for equal work.
04Rule 2: Give it a finish line
  • Define done as something Fable can verify, split the job into tasks that each get a pass or fail test, and set the bar as high as you can.
  • Build with Opus on medium and review with a separate Fable session that never saw the build, since Fable resists pushback and rarely invents problems.
05Rule 3: Three jobs for Fable
  • Give it whole jobs far bigger than usual, like rewriting large parts of an app or planning how a full app fits together.
  • Use it for 3D scenes and hard game work, and for the hardest problems where it matched Opus at max effort for about half the cost per task.
06Rule 4: Fable leads, cheaper models build
  • The lead only decides and judges, so route simple tasks to Sonnet, harder ones to Opus 5.5 and file searches or log reading to Haiku.
  • Record model-per-task choices in CLAUDE.md and add a note on sub-agent cost, and skip this setup on small jobs where it costs about 1.7 times Opus alone.
08Rule 5: Run the lead on high effort
  • Fable stops early on long jobs more often at low and medium effort, so set the lead to high before a job you will leave running.
  • Anthropic's CLAUDE.md block halves unfinished sessions on medium but does not stop the list what's left then quit pattern, and going past high adds cost for little gain.
09Rule 6: Set up so it keeps going
  • Auto mode only trusts the current repo, so add every place Fable may work, including the test site, to CLAUDE.md.
  • Reversible steps proceed and irreversible ones wait, and when Fable asks, reply with exactly what you approve so it doesn't repeat the question.
10Rules 7 to 9: Reports, cleanup, switching
  • Have the final report written for someone who never saw the work, leading with key points and including screenshots and test results, optionally as a browser page.
  • Replace never rules with ask-first wording, run the prompt audit command, and compact before switching models because the prompt cache and Fable's reasoning do not carry over.
Glossary

Terms worth knowing.

Effort
A Claude Code setting for how long the model thinks before answering. Lower effort uses fewer tokens, higher effort thinks longer and costs more.
Lead agent
The main model that splits a job into tasks, hands them to worker sub-agents and judges what comes back, instead of doing the building itself.
Sub-agent
A separate Claude session spawned by the main one to handle one task, often on a cheaper model.
Prompt cache
A saved copy of the conversation Claude has already read, kept for an hour so repeat reads cost far less. It belongs to one model.
Auto mode
A Claude Code mode that approves actions on its own, but by default only trusts the current folder and repo.
FrontierCode
An Anthropic coding benchmark that checks whether maintainers of real public projects would accept the code a model writes.
Resources

Things they pointed at.

00:15channelLydia Hallie Claude Code best practices talk
00:20linkAddy Osmani guide on which Claude model to use
13:00productBrevo ↗
Quotables

Lines you could clip.

04:55
“The first rule with Fable 5.2 is that you don't start on it.”
short contrarian hook→ TikTok hook↗ Tweet quote
14:40
“Fable treats any line with the word never in it as a rule it can't break.”
specific, surprising, actionable→ 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.

metaphor
Ever since Fable 5 .2 came out, people have been using it for everything, but they keep running out of their limit before the work is done, because Fable has rules and most people don't know them, and the team at Anthropic has been sharing those rules. Lydia Halley from their Claude Code team gave a talk on Claude Code best practices.
Addy Osmani from Anthropic wrote a guide on which Claude modeled to use for each job, and Anthropic put out its own guide for Fable. We went through all of them, tested them on our own work, and narrowed them down to nine rules that unlock the max potential of this model two are how the people who build cloud code work themselves one is the opposite of what most people do with fable and most are about how you set fable up not how you prompt it if this is your first time we're a software company and this is our channel ai labs and in this video we're going to show you all nine rules why each one matters and how to use it in your own workflow this section is for those who don't know about fable 5 .2 so if you already know you can skip ahead to the first rule from the time timestamps in the description below.
So Fable 5 .2 is Anthropic's best model so far and it came out on November 9th. It's made for the hardest coding jobs and it beats Fable 5 .1 at almost everything. The first thing that changed is the price.
Fable 5 .1 costs $10 for every million tokens it reads and $50 for every million tokens it writes. Fable 5 .2 costs $8 and $40 so it's 20 % cheaper on both. Fable 5 .2 also writes its answers up to 30 % faster.
And on top of being cheaper, it needs fewer tokens to do the same work. You already know effort is how long the model thinks before it answers, and the lower the effort, the fewer tokens it uses. On medium, Fable 5 .2 finishes hard jobs with fewer tokens than 5 .1 did.
And even on low, Fable 5 .2 scored close to what 5 .1 got on a higher effort setting in Anthropix coding tests. And it doesn't just do the same work with fewer tokens. It also does better work than 5 .1.
Frontier code is a test that checks if the people who run real public coding projects would accept the code a model writes. When Anthropic tested both Fable models on it, 5 .1 scored 32 .4 % and 5 .2 scored 49 .1%, both on their highest effort setting. Anthropic says that 49 .1 % is the best score of any model that's out right now, and even Opus 5 .5 on its highest effort setting only scored 45 .5%.
And on almost every other test Anthropic ran, Fable 5 .2 beat the other top models too. Besides the higher scores, Fable 5 .2 also gets in your way less. Fable models have safety rules that stop them from helping with some kinds of jobs, mostly around hacking and biology.
But those rules sometimes block a normal request by mistake, because it looks like something that goes against their rules. Anthropic says Fable 5 .2 does that less often than 5 .1 did. The way 5 .1 wrote was a problem too.
our previous video, it wrote a lot more than it needed to. Its sentences were long with very few breaks, so its answers were hard to follow. Anthropic says 5 .2 writes more clearly than 5 .1.
Fable 5 .2 also starts subagents to go deeper into reviewing its own work on its highest effort settings, which the previous version never did. That habit is a big part of why Fable 5 .2 is so good at long jobs, and Lydia said the same thing in her talk. She said Fable is really good at checking what it builds and that both both Opus 5 .5 and Fable 5 .2 are great at long tasks, and that's not the only thing Fable 5 .2 is best at.
Anthropic also calls it their best model for 3D, and a lot of people are already using it to build full 3D games that look great. But even after the price drop, Fable 5 .2 still costs twice as much as Opus 5 .5 for every token. Anthropic also says it uses up your limits way faster than their other models.
So after numbers like those, you might want to switch to Fable for every... But the first rule with Fable 5 .2 is that you don't start on it. That comes straight from Anthropic's launch post which says to start with Opus 5 .5 for most of your work.
Anthropic also says you should only move to Fable when Opus on a higher effort setting still can't do what you need and the people who build Claude Code work this way too. Lydia calls Opus 5 .5 her favorite model for daily work and she starts every session on the normal settings which is Opus 5 .5 with auto mode. And you can see why when you give both models the same job a builder on x gave them the same one line prompt to build a 3d model opus 5 .5 finished it in 27 minutes fable 5 .2 took 58 minutes and used about seven times as many tokens and got the same results on everyday work opus gets you the same result for half the price per token and it doesn't use up your limit as fast when we tested both models on the same feature they both did work that was just as good but fable cost about 58 more but before Before going to the next rule, it would be great if you subscribe to the channel and hit the hype button.
This small gesture of support goes a long way for us. But to know if opus is enough, you need a way to tell if the work is really done. So the second rule is that you tell the model what finished looks like before it starts.
And that's something we've always said in our previous videos as well. Lydia explained why in her talk. Fable works for hours on its own and makes a lot of small choices along the way.
You don't find out about those choices until you dig into what it did. because its report won't tell you all of them. So a finish line gives Fable something to check those choices against.
Lydia's fix is to say what done looks like and ask Fable to split the job into smaller tasks. Each task gets its own test, which is anything that comes back as a pass or a fail. Fable needs things it can measure, because asking it to just make the page good gives it nothing to check.
So when you start a job, you need to write in your prompt what done means as something it can verify. And once you've written that finish line, you should set the bar as high as you can.
Fable keeps working until it reaches the bar you give it, so a low bar gets you a normal result. But when you set the bar really high, Fable does even better than you'd expect. Once the work is done, that's where the review comes in, and Fable is really good at checking work.
It does even better when it checks the work in a fresh session that hasn't seen how the work was built. On most of our builds, we got better results when Opus 5 .5 on medium effort did the building and a separate review agent on Fable 5 .2 checked. the work.
Fable's review was strict and it found a lot of problems in the build, even the kind Opus usually misses. The reason Fable is such a good reviewer is that it doesn't just agree with you. Most models go along with whatever you tell them, so if you push back on their review, they change their answer.
Fable sticks to what's actually true and what it can prove with a source. So even if you try to talk it into thinking something is wrong, it won't go along with it most of the time. And it doesn't make up problems or fixes that aren't there either.
Anthropic's own testing backs that part. because Fable was less likely than their other models to make things up or to say it had finished work it hadn't. The third rule is that you use Fable for three kinds of jobs.
The first kind is big jobs. Lydia's advice was that instead of giving Claude one small task at a time, you should give it a whole job that's way bigger than usual. Addy Osmani from Anthropic lists the same kind of work for Fable, like rewriting big parts of an app and planning how a whole app fits together.
The second kind is 3D and game work. Unity said Fable did best at building 3D scenes from scratch and at the harder game jobs, where it keeps making the result better instead of getting stuck.
So if you're building a game or a 3D scene, that's a job for Fable. The third kind is the hardest problems. On that coding test from earlier, Fable on high effort scored about the same as Opus on its highest effort setting for about half the cost per task.
So on the hardest jobs, Fable is the cheaper choice. When you hit one of those jobs, that's where you should use Fable. The fourth rule is that Fable leads and cheaper models build.
Anthropic calls Fable their strongest lead agent. They use it as a manager that splits the job up and hands out the smaller tasks to other models. Lydia said her team at Anthropic already works this way.
So instead of talking to eight sessions, they only talk to the lead Fable agent and it handles the rest. She also explained how the work gets split. The lead only has to decide what to do and judge what comes back.
So a model like Fable is a good fit for it. The sub -agents doing the work need a model that's good at the work. itself.
But the workers don't all need the same model, so the lead picks each one based on the task it gets. If a task Fable splits off is simple and clear, Sonnet can do it. If it's harder and needs more thinking, that task should go to Opus 5 .5.
And for small jobs like searching through files or reading logs, Haiku is enough. That's also how Addy splits up the models in his guide. So to set this up, you just need to open Cloud Code and pick Fable as your main model with the model command.
Then you tell it that it's the conductor. It does the thinking, splits the job into tasks, and gives each one to its own worker. Then it waits while the workers build their parts.
When each worker's result comes back, Fable checks it, and it only stops when it needs a decision from you. You can also ask Claude to write in your Claude .md which model should do which kind of task. That way, when Fable starts a sub -agent, it picks the right model for it.
But those workers aren't free, and Fable is quick to start them. Anthropic says that when your instructions ask Fable to hand work off, it does that quick. This means sub -agents can end up being a big part of what a session costs.
In Anthropic's own test, telling Fable what a sub -agent costs cut the number it started by about half. So we'd recommend adding a short note to your clod .md that says what a sub -agent costs. The note should tell Fable to only start one for work that can run at the same time, and this setup is only for big jobs.
On a small feature, this setup cost about 1 .7 times what Opus alone did because the lead reads everything again to check it. Addy says that for work that fits in one session, planning it out and putting the pieces back together costs more than it saves.
But before we talk more about this lead agent, let's have a word by our sponsor. Brevo. So you've just vibe -coded a working app and people can sign up, but nothing lands in their inbox.
No welcome email, no password reset, no receipt. Brevo is the email backend for your app, and it's honestly one of the best transactional email APIs for developers. You get a REST API, SMTP, and SDKs in seven languages, so your app sends real mail with a single call.
Lovable has a native connector, and with any other builder, it's a standard REST call. You get dynamic templates, real -time delivery events, and 99 % deliverability. And there's SMS and automation on the same platform.
So we pointed Cloud Code at our app and asked it to wire up Brevo. It added the call, built the template, and sent a real welcome email while we watched. So if your app still can't reach its users, Brevo is the tool you need.
Start free with 300 emails a day. The fifth rule is that you run the lead on high effort. Anthropic says that on long jobs, Fable sometimes stops before the work is done.
It might list what's left or offer to keep going. Or it might stop at something normal, like a command taking too long. And it does this more often on low and medium effort.
That's a problem for the lead, because it's exactly the kind of long job you don't want to do. watch. So when the lead stops early, the work just sits there unfinished until you come back.
The lead is also handing tasks to sub -agents and doing all the planning. So it should run on high effort to make fewer mistakes and to stop it from quitting early. There are two ways to fix this.
The first is to set the lead to high with the effort command before a long job you'll leave running. Anthropic says that high is for long jobs nobody is watching and it's where you get most of what the model can do. The second is to stay on medium and add a block of text.
from Anthropic to your Claude .md. That block tells Fable not to stop while it still owes you work and to keep going instead of offering to wait. In Anthropic's tests on Medium, the block cut unfinished sessions by about half.
But the block only fixes one of the two ways Fable stops early. Anthropic found that it mostly stopped Fable from asking if it should keep going. It didn't stop Fable from listing what's left to do and then stopping anyway.
So if your runs keep ending with a list of what's left, the block won't help. And that's when you need to set... the lead to high.
But whichever fix you pick, you shouldn't go past high. Anthropic says that extra high only raised scores a little and used a lot more tokens. And Addy says that past high, one more level can cost as much as a bigger model.
The max effort setting goes even further, and Anthropic says it's only for jobs where time and tokens don't matter at all. The sixth rule is that you set clod code up before a long job so it can keep going without asking. Lydia pointed out that by default, auto mode only trusts the folder and repo you're working on.
in. So if you ask Claude to put something on the test version of your site which lives somewhere else, Auto Mode blocks it. Then you're back to saying yes to things yourself.
To fix that, you can change Auto Mode settings. We showed the Auto Mode setup command in one of our earlier videos and it sets them up for you from your past sessions. But those settings only cover what Claude Code is allowed to run.
Fable also needs a list of the steps it can take without asking you. To sort them, Lydia's test is simple. If a step can be undone, Fable goes ahead.
If it can't, Fable waits for you. You need to add every place Fable is allowed to work into your cloud .md like your project folders, your repos and the test version of your site.
That way it can work in those places without stopping. You can also add a line to your prompt that tells Fable to ask you first about any decision it needs from you before working on it. So when Fable stops to ask about one of those decisions, don't just reply with go.
Anthropic says to tell it exactly what you're saying yes to so it doesn't ask you the same kind of question again.
So the seventh rule is that Fable's final report has to make sense to someone who never watched it work. Anthropic says the person reading Fable's final report usually hasn't seen any of the work behind it. So they give you a block of text for your clod .md that tells Fable to write for someone who doesn't know what happened.
It tells Fable to start with the most important points and use plain words instead of unnecessary jargon. And Lydia also asks Fable to show the actual results since you weren't there to see them.
That means screenshots of the app and the test results along with the summary. In one of Anthropic's tests, Fable's final summary often left out small changes. You can also ask it to make the report a page you can open in your browser, written in simple words that are easy to understand.
The eighth rule is that you update your setup for Fable. Your clod .md probably still has lines you wrote for older models, like ones with never or important in them. Lydia says these confuse the model and make the result And Fable treats any line with the word never in it as a rule it can't break.
So where you still want the final say, you can write ask before deleting files instead of never delete files. But you don't have to go through these lines yourself. Cloud Code has a command for this called prompt audit.
It checks your cloud .md and your skills against the model you're on. And it tells you what that model doesn't need anymore. We showed it in one of our earlier videos too.
The ninth rule is about when you switch to another model. Cloud keeps a saved copy of the conversation. it has already read for one hour.
When it needs the conversation again, it uses that copy instead of reading everything from the start, which saves a lot of money. But that copy belongs to the model that read it. So when you switch models, the first message on the new model has to read the whole conversation again at full price.
And when you move off Fable to any other model, the thinking Fable did before each answer doesn't come with it. So the new model keeps going without knowing why Fable made its choices. So when a task ends, you can run the compact command a fresh session and then switch models now all the skills and workflows that you see in our videos are all available in ai labs pro which is our community so if you found value in what we do and want to support the channel this is the best way to do it the links in the description that brings us to the end of this video if you'd like to support the channel and help us keep making videos like this you can do so by using the super thanks button below as always thank you for watching and i'll see you in the next one
The Hook

The bait, then the rug-pull.

People reached for Fable 5.2 on everything and ran out of limit before the work was done. AI LABS collects the rules Anthropic's own team uses and boils them down to nine.

Frameworks

Named ideas worth stealing.

05:00list

Nine rules for Fable 5.2

  1. Don't start on Fable
  2. Give it a finish line
  3. Use it for three kinds of jobs
  4. Fable leads, cheaper models build
  5. Run the lead on high effort
  6. Set up so it keeps going
  7. Final report for a cold reader
  8. Update your setup for Fable
  9. Compact before switching models

The video's full checklist, drawn from Anthropic's talks and guides.

Steal forA pre-flight checklist before any long Claude Code job
13:10model

Reversible step test

  1. Can it be undone: go ahead
  2. Cannot be undone: wait for the human

Lydia Hallie's simple rule for which steps Fable may take without asking.

Steal forWriting the permissions section of a CLAUDE.md
CTA Breakdown

How they asked for the click.

VERBAL ASK
15:35product
“All the skills and workflows in our videos are available in AI Labs Pro, our community.”

A soft plug near the end, after the nine rules are delivered.

MENTIONED ON CAMERA
13:00productBrevo ↗
Storyboard

Visual structure at a glance.

open
hookopen00:01
FrontierCode result
valueFrontierCode result02:18
rule tiles
valuerule tiles05:07
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.