Big Idea

The argument in one line.

Claude Fable 5 earns its 2x token cost only on long, high-stakes work — on easy tasks it is nearly indistinguishable from Opus, but on the jobs that used to take a full day of babysitting it becomes the first model where the bottleneck shifts from the AI to how clearly you can articulate what you want.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

An AI automation practitioner or developer actively using Claude Code who wants a real field report on Fable 5 before upgrading usage.
Someone deciding whether the 2x token cost is justified for their workflow and needing concrete build evidence, not benchmarks.
A technical founder who uses AI for complex, multi-hour builds and wants to know exactly where this model outperforms its predecessors.

SKIP IF…

You want a feature rundown or capabilities spec — this is a field report from a practitioner, not an official overview.
You are new to Claude Code and have no baseline to compare against; the delta this video covers requires prior experience with Opus.

TL;DR

The full version, fast.

After a full day building with Claude Fable 5 (Mythos), the pattern is unmistakable: the model earns its cost only on hard, long-running work. A 41-minute unsupervised agentic build produced a complete business hub app that tested its own server, fired real notifications, wiped its test data, and wrote its own limitation documentation. A second test produced a functional 3D CAD editor with a built-in Copilot in one prompt. The three honest problems are real: session limits burn fast (119K tokens in a single turn), safety guardrails silently downgrade the model to Opus mid-build without notification, and hard requests take minutes to complete. The practical verdict: specialist tool reserved for your two or three hardest jobs — the shift from babysitting tasks to handing over goals is the real story.

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 – 00:36

01 · 24 hours in: the honest review

Setup and promise: not hype, real usage, own money. The shift changes how you use everything.

00:36 – 01:22

02 · You stop babysitting it

With previous models you gave a task and checked results. With Fable you hand over a responsibility and walk away.

01:22 – 01:51

03 · Karpathy calls it a step change

Andrej Karpathy called it a major version bump. Working software now comes out on a tab.

01:51 – 03:04

04 · How to get the most out of it: 3 tips

Three rules from Anthropic prompting guide and day-one testing: point at hardest problem, interview first, tell it why.

03:04 – 03:33

05 · Test 1: building a full business hub app

A local-first business hub web app for a solo service company, built in one shot with effort extra-high.

03:33 – 04:42

06 · Letting it interview me first

Prompt: you are my technical cofounder. Interview me before writing any code. Three rounds of questions, then a full build spec in 3:36.

04:42 – 05:36

07 · The build spec and running the build

Said go and watched files populate: node modules, agents.md, calendar grid, events editor, client pages.

05:36 – 07:49

08 · 41 minutes later: what it built

Full walkthrough: dashboard, drag-drop pipeline, clients, projects, money/invoicing, calendar, settings. It tested its own server, fired real Mac notifications, wiped test data, wrote its own limitations doc.

07:49 – 09:40

09 · Test 2: a tool building a tool

One prompt in Claude desktop: build a 3D CAD editor and design a printable phone stand inside it. Result: Forge — full editor with objects panel, inspector, STL export, built-in Copilot.

09:40 – 10:18

10 · It keeps flipping back to Opus 4.8

Safety system silently downgrades to Opus mid-build. User had to switch back manually every time.

10:18 – 10:49

11 · The final 3D-printable part

Watertight, support-free phone stand (90x78x98mm). Added lip and cable slot unprompted. Then warned about a flaw in its own export button.

11:43 – 12:00

12 · Problem 1: burns through usage

119K tokens in one turn. A full day of Opus usage gone in one hour.

12:00 – 12:25

13 · Problem 2: guardrails too jumpy

Filters trip on innocent work, silently swap to Opus. Anthropic says fixing it is a top priority.

12:25 – 12:51

14 · Problem 3: it is slow

A single hard request can take several minutes. By design — the tradeoff is the output quality.

12:51 – 14:27

15 · Where I land after a full day

Specialist, not daily driver. Point at hardest, longest, most expensive-to-get-wrong work. The job changed from babysitting tasks to handing over goals.

Atomic Insights

Lines worth screenshotting.

The first model where the bottleneck is not the AI but how clearly you can describe what you want.
On easy tasks, Fable 5 is barely distinguishable from Opus — the gap only opens on problems that used to take a full day.
Letting the model interview you before it writes any code is the single biggest quality lever available.
A 41-minute unsupervised build is not a bug — it is the entire value proposition of the model.
Fable 5 wrote its own limitation documentation, tested its own server, and warned about a flaw in its own export button.
Telling the model why — who the output is for and what it feeds into — consistently produces output closer to what you pictured.
Safety guardrails silently downgrade to an older model mid-task without notification; watch the model picker on long builds.
119K tokens consumed in a single turn means your daily session limit is gone in one hard job — treat it as a timer, not a buffet.
The model shifted from task-mode to responsibility-mode: instead of investigating one crash report, it watches every crash and keeps the app healthy.
Fable 5 is a specialist, not a daily driver — expensive and slow by design, worth it only on the work where getting it wrong is costly.
The job changed from giving tasks to handing over goals — that mental shift is what unlocks the model.
A 3D CAD editor with a built-in Copilot was produced in one prompt inside the regular Claude desktop app.
The model added features to the phone stand without being asked — then warned that its own export button was unreliable for printing.

Takeaway

Give it a goal, not a task.

WHAT TO LEARN

The model that stops needing supervision demands something harder from you: knowing exactly what to ask for and why.

Point Fable 5 at your hardest problem — on easy tasks the difference from previous models is negligible, and the gap only opens on the work that used to take a full day.
Let the model interview you before it builds anything; interview-first runs produced results that were night-and-day better than prompt-and-go in every real test.
Tell the model why — who the output is for, what it feeds into, and what success looks like — and you get back something much closer to what you pictured with far less micromanaging.
Treat your session limit as a countdown timer, not a free buffet: 119K tokens in one turn means a full day of budget is gone after a single hard build.
Watch the model picker on long runs — safety filters silently downgrade to an older model mid-build without notification; if a response feels off, check which model is actually responding.
The shift from babysitting tasks to handing over goals is the real unlock: once the model stops being the bottleneck, the hard part becomes how clearly you can describe what you want.
Self-testing, self-documentation, and self-correcting behavior — the model warned about a flaw in its own export button — change what you personally need to verify after a build.

Glossary

Terms worth knowing.

Fable 5 (Mythos): The internal codename for the Claude model tier released in mid-2026, positioned above Claude Opus. Designed for long-horizon agentic tasks and consumes approximately 2x the session tokens of Opus per request.
/effort extra-high: A Claude Code slash command that sets the effort level to maximum before a build run, signaling the model to allocate more reasoning and planning budget.
Agentic build: A build session where the model runs unsupervised for an extended period — planning, coding, testing, and revising — without human check-ins between steps.
Model picker: The dropdown in the Claude desktop interface that selects which model handles a conversation. Fable 5 can silently flip this back to Opus when its safety filters trigger.
STL export: Standard file format for 3D printable parts. The model built a CAD editor, designed a phone stand, and produced a verified STL file from software that did not exist one hour earlier.
Build spec: The structured plan a model produces after an interview phase — listing architecture, features, data models, and implementation steps before any code is written.

Resources

Things they pointed at.

01:22channelAndrej Karpathy

01:51linkAnthropic Fable prompting guide

13:47productReprises AI ↗

13:47linkSystems to Scale School community ↗

Quotables

Lines you could clip.

11:08

“I was sitting there just trying to come up with something hard enough to actually trip it up and I kept losing.”

Punchy reversal — model outpacing the user is the thesis in one line→ TikTok hook↗ Tweet quote

12:36

“The model basically stops being the hard part and now the hard part is just knowing what to point it at.”

Clean thesis statement, no setup needed, quotable as text→ newsletter pull-quote↗ Tweet quote

13:42

“We went from effectively babysitting tasks to handing over a goal and walking away.”

The paradigm shift in one sentence — instantly shareable→ IG reel cold open↗ 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.

00:00So it's been about twenty four hours since Claude Fable came out, and I've barely done anything else but utilize it. I've had it been building real applications. I've watched it run on its own for literally half an hour straight without me touching anything.

00:11I've watched it just quietly rip through my plan like it was nothing. And I'll be straight with you, this is not the launch day hype video. I already did the what it is type of breakdown yesterday, so that'll be linked down below if you are interested in seeing what this is all about.

00:24This is going to be the honest one where I've had a full day in with actually using this, throwing my own money on the line, showing what actually holds up, and the one shift that changes how you use this thing, and the stuff that is just genuinely going to be annoying you. Let's get into it. Now, first thing that I had to unlearn is how I actually talk to this model.

00:39And this is gonna be a reoccurring theme throughout the video, so just stay with me. Now, with every model leading up to now, including Opus, you were in a loop of just constantly checking. So you give it a task, it does the task for you, and then you just go make sure it didn't do something dumb.

00:52So you were babysitting. So I was constantly asking myself if it was doing the right work. But with Fable, that question I've been noticing, it really just goes away for after almost an entire day on it, I mostly trust that the code it writes actually works.

01:04Example directly from the Anthropic team, one of their lead engineers said that he used to tell it, go look at this one crash report, that is a task. Now he doesn't. So it just sits there watching every crash that comes in, and its whole job is just to keep the app from actually crashing.

01:19So he gave it a responsibility instead of a a task. And this isn't just them hyping up their own thing as you would probably expect them to do anyways.

01:26Andre Karpathy, honestly, one of the most respected voices in this whole space, he called it a major version bump deserving step change. Now, just a fair warning, he actually just joined Anthropic a few weeks back, so take that with a grain of salt. But the thing that he said that stuck with me is that working software now comes out, in his words, on a tab.

01:44So you just ask for the thing like a dashboard, a tool, a one off application, and it shows up. And after today, that's exactly how it feels.

01:51Before I show you what I built, here's how you actually get the most out of it. And this is going to be coming from two places, Anthropic's own prompting guide for Fable, and me testing this thing the whole day, and just seeing what actually moved the needle. Number one, and this is the really big one, pointing it at your hardest problem.

02:05Anthropic literally says, if you are only going to test it on easy stuff, you're just gonna be underselling it. And I've definitely felt it all day using it, where on small tasks, I've barely felt any difference from using Opus. And the bigger and messier the job that I gave it, the more that it actually pulled ahead.

02:20So do not waste it on the little things. Use it for the really, really hard stuff. Number two, let it ask you questions before it builds.

02:26So their advice is to pick something harder than you would normally hand a model, and then let it scope the whole thing out and interview you first. Now I've been trying builds both ways today. In the interview first runs that I have been doing, it came out completely night and day better.

02:40So it's the single biggest upgrades to all of my results, and you will see exactly what it looks like in just a second. Number three, tell it why. So you need to be giving it the reason behind the task, who it's going to be for, and what it is feeding into, and it just gets it more often.

02:53So every time I gave it the, like, actual why, giving it the reasons, I got back something closer to what I had pictured. So less micromanaging, much more context.

03:03That is the whole game now. Okay. First test, let's get into the thing that this model is supposed to be best at, which is just long messy work.

03:09So I'm going to have it build a full app all in one shot, and it's gonna be from my actual world. So I run service businesses and I live in client meetings. So Claude code, Fable five, it is loaded and first move is gonna be the slash effort extra high.

03:22So that's the level built for runs just like this. So let's just navigate to the model, select Fable, and we'll do the effort extra high.

03:31Not letting it build anything where first, I'll just make it interview me. So I'm going to list out this prompt right here saying, you're my technical cofounder. I want a local first business hub web app for a solo service company.

03:40Before writing any code, interview me. Just ask every question that you need about the features, the data, the design, one batch at a time until you fully understand the build. Do not write the code until I say go.

03:49By the way, I'll have all resources and all the assets inside of this video available for free to download inside of our school community. So you can get access to all of that and so much more inside of that community. Alright.

03:59So we have our list of questions. So let me just go ahead and enter one by one. I've just submitted all of that.

04:04Let me run through it and give it back to Claude. Just a little bit later, it's asking some more questions, just getting more specific and getting really into all the nuance. So I've just entered in some of those answers for that.

04:14And you'll notice that within the status line for all of these different answers that I'm just providing it, It's gonna be chewing on this for quite a while. So it might come out to be a couple minutes. We'll see.

04:23But either way, this thing is going to be thinking longer than anything I have used. That's been pretty much like the reoccurring theme that I've seen from using Fable thus far, but it's as you would expect given you're providing with a lot of data and making sure that it's just going slow.

04:37Now we have our third list of questions. Let me get through this. And yet again, we're getting more lists of questions back to us.

04:42And just like that, about three minutes thirty six seconds later, after interview and interview, it has then provided me with the build spec as you can see here. So assuming that everything was good, I'm gonna say go and use slash goal. And I'll run this off.

04:57One of the coolest things that you'll notice when I actually do slash goal is it recognizes that that is not a skill. So in this case, I had to use it a little bit differently. So it's just stating back to me that it is not a skill.

05:07But you'll notice that with what Fable had actually built out for me, so this entire build spec, the plan that it just wrote, that is the goal. So just a couple of minutes into this, you can see all the different files actually populating inside of my folder.

05:20So we have the node modules. I mean, everything wow. It's actually built out quite a bit pretty fast.

05:25Then we have the agents.md folder. We have our calendar grid, events editor, client pages, and the last item it should give itself to just verify everything end to end.

05:36Alright. Now, after about forty one minutes later, here's what it came up with. So obviously, about forty one minutes is kind of unheard of when we are looking at different models like Opus.

05:46It wouldn't have taken nearly as long, but it wouldn't have given us the output that we have right here. Before I show you all of the different features, let me first just walk through what it actually did. Now, the biggest thing is it tested its own server end to end.

05:59So it confirmed a second Tuesday of the month meeting, typically lands on the right four dates, it listed them out, it ran its own quick ad example, it confirmed the events to the right client, even fired a real Mac notification and inspected every screen in a hidden browser. It wiped its test data so I would just get a clean first run, and it even flagged its own weak spots and wrote its own limitation sections, and that used to be my job or another developer's job.

06:22Anyways, let's go back into the application. You can see it's extremely clean. So we have the dashboard right here.

06:28Wednesday, June 10. Here's everything that's outstanding. Overdue, the revenue in June, the retainer, the MRR, follow ups, everything that's going on today.

06:37So pretty clean overview, would say. Here's our pipeline. We should be able to click on everything, we can click on the new concept name.

06:43Let's try to input something random here. Let's do that. One, three.

06:46Deal value, let's say, $10,000 and there we go.

06:50So we should be able to drag this back and forth. Everything looks good there. Our clients, you can see high is now populated inside of here exactly how we were hoping.

07:00Our projects, we could try to throw in a random project here.

07:05Let's do that. Tasks. Looks like tasks is actually broken, but let's go into money.

07:12Everything looks good here. We can create an invoice. We can create a quote.

07:15Create a container, a retainer as well. Here's our calendar, and here is the settings. So this one right here, I mean, you can call this AI slop all you want, but I mean, this is extremely thorough.

07:26I mean, it better be for taking about forty one minutes, but this is exactly what I had requested for. That is test number one. It is literally just one prompt for me.

07:34Took about forty minutes and I was completely unattending this. I did not intervene once inside of this chat right here. I mean, was just going back and forth with itself and it is quite a long list.

07:44I can't even go all the way up to the top. But as you can see, it is pretty insane. Now moving into the second test.

07:50This is the one that I keep coming back to and the one I keep thinking about because it's a tool building a tool. So this time, I'm not just gonna be in Claude code inside of my IDE. I'm just gonna use the regular Claude desktop application.

08:03Fable five is now picked out, and we're just gonna do one prompt. So I'm gonna say build me a three d CAD editor, the software engineers use to design printed parts, and then design a real part inside it. Here's the actual prompt that I am throwing in.

08:15Again, you guys can get access to all this stuff inside of my free school community. Alright. So sometime later, here's the result.

08:21Now, you'll notice that it didn't just throw up a viewer. So it actually builds a real editor, It named it Forge.

08:27On the right side, we have an objects panel just listing every part of the stand, the base, the lip, the back rest, and it's all clickable and interactive. So we then have an inspector for the exact positions.

08:40We have an exports to the STL button up top, and then down here, we have a Copilot built into its own application. The command's already loaded. You can make the backrest 10 millimeters taller, turn the whole stand matte black if we wanted to, so I could just click one and the model changes right in front of me.

08:57And you'll notice that it didn't hand me just one design and call it a day. So the colors, it kept cycling while it tuned the shape, and at one point, it even asked me like, do you want me to adjust the angle, widen the cradle for a thicker case, or add an AirPod slot? So it's the same interviewing behavior as the big build just in miniature.

09:15Now, to get a bit honest here, I will say that it got pretty messy in the middle where I asked for a tweak and it threw a pile of random shapes into the scene. So some cones, some spheres, like a cube. Like it was just thinking out loud on my canvas.

09:27So I did have to tell it and kind of just like put it back on its path, clear the scene, get back to the stand. So it is incredibly powerful, but it's not just magic just yet. So you're still steering just a little bit.

09:40The one thing that kinda got under my skin that nobody else I've seen is flagging this. If you look at the model picker, so I set this chat to Fable five. So mid build, it kept quietly flipping itself back to Opus 4.8, and that's just the safety system.

09:53So Fable hands certain things to the older Opus, and the filters, they're so cautious right now. They trip on just totally normal work like this where I kept switching it back by hand.

10:02So if you are running this, doing something pretty big and complex, just watch that picker. It does not stay where you put it and I mean, we know why the reason is because they don't want this to find any like security vulnerabilities and do anything that's gonna be, you know, doom or gloom. Right?

10:18But anyways, if we look at the final output, just look at this render. So read the line underneath. It's 90 by 78 by 98 millimeters.

10:26It's watertight, prints support free. It even added a lip for the phone and a cable slot, so I never even asked for either of those. And it told me how to print it, the material, 20% infill, and then it made me laugh.

10:39It warned me about its own application. In the editor's export button, it just welds shapes as is. So for any actual printing, you wanna be using the verified file that it had attached instead.

10:49So it then quality checked itself over its own tool. We then have a printable file from software that did not exist an hour earlier. So after a full day of testing, the pattern that I keep running into.

11:00The longer and harder the job, the bigger this is going to be winning. So on just a quick one liner, you're barely going to notice the difference over Opus, but on the stuff that used to take a full day of just back and forth, that's where it really pulls away, and it's actually pretty insane when you put it to the test.

11:13And by the end of the day, actual problems had flipped. So I was sitting there just trying to come up with something hard enough to actually trip it up and I kept losing. So it's the first model where the bottleneck actually just stopped being the model and it started being just me and how clearly I can actually describe what I actually want.

11:30So if you're gonna ask me like where does this shine, it's gonna be long, messy, high stakes work. So any big migrations, big research projects that runs for hours or whole application from an idea like that's the lane, that's what you need to reach for it with this.

11:43Now, part that actually matters if you're about to go and use this because I will say it's not all sunshine. Problem number one, and it's a pretty big one, this thing burns through your usage like nothing you have ever seen. So application, it literally warns you right under the box, like Fable takes about two times the usage of Opus, and that is not a joke.

12:01Like, I burned through pretty much my whole session in about in just hour of real building. And those are limits that normally carry me through an entire day just completely gone. After watching that build eat a 119,000 tokens in one turn, I get why.

12:14So that free window that you get on your subscription right now, treat it like a timer, not just a completely free buffet. But what I would say to you is just find the two or three jobs where it's clearly worth it and save it for those. Alright.

12:25Now problem number two, it's the guardrails. So right now, the filters, they're just way too jumpy and they can catch totally innocent work and mind tripped on just a simple phone stand. And in Thropic, they did say they're fixing those false alarms and that's a top priority.

12:38But until then, if an answer feels like a notch dumber than you expected, you probably got Opus without actually knowing it. You just have to switch back to fable. And problem number three, it's not fast, not at all.

12:48So it thinks a lot and a single hard request, it can run for several minutes before you get anything bad. The trade off, of course, it's the output and it's usually worth the wait. But if you want quick and snappy like this is the wrong tool and that's by design.

13:01So you need to have patience. So where do I actually land on this after a full day? Well, I'll come back to you with two things.

13:07One, this is the real deal, and it is a specialist, not at all a daily driver. So the builds that I ran, they're just things that the last generation just couldn't one shot. But it's expensive, it's slow, so you're not gonna want us to be pointing it to your easy, high volume stuff if you couldn't tell already as I've been just hammering it away.

13:24You're going to be pointing it at your hardest, longest, most expensive to get wrong work, and on that, it pays for itself in a single good run. Number two, the real story is not the benchmark chart. It's that the job changed.

13:36So we went from effectively babysitting tasks to handing over a goal and walking away. And after today, like that's the shift that I cannot unsee. The model, it basically stops being the hard part and now the hard part is just knowing what to point it at and honestly, that's a really fun problem to have.

13:52So that's my honest take about twenty four hours in of really using this thing. So I'm really curious to hear what you guys have been using this for, any feedback that you guys have. And let me know, of course, if you wanna see any other videos on this stuff.

14:03I'm gonna be hammering this thing away and continuing to use it no matter how expensive it is and pointing it at all my heart problems. But with that being said, make sure to check out our free school community link will be down below in the description. And if you're a business owner looking to implement AI systems to ultimately scale your business in 2026, you can book in a call with our team.

14:19Link will be down below in the description. We'll show you how we can drive leverage and increase your bottom line. But with that being said, thank you guys for watching, and I'll see you in the next video.

The Hook

The bait, then the rug-pull.

Twenty-four hours, two full application builds, real money on the line. This is not a launch-day hype video — it is the honest one, from someone who barely stopped using the thing long enough to record a video about it.

Frameworks

Named ideas worth stealing.

01:51list

The 3-Rule Fable Prompting Protocol

Point it at your hardest problem — easy tasks will not show a meaningful difference from Opus
Let it interview you first — give it the task, let it scope and ask questions before it writes a single line
Tell it why — who the output is for and what it feeds into, not just what to build

Three rules distilled from Anthropic official prompting guidance and one full day of real builds. The interview-first rule produced the biggest single quality jump.

Steal forAny complex agentic build session; pre-prompt template

12:51concept

Specialist vs Daily Driver split

Fable 5 is not a replacement for Opus — it is a complement. Reserve it for the 2-3 hardest jobs per day where getting it wrong is expensive. Use Opus for everything else.

Steal forAI tool selection framework for teams managing session budget

CTA Breakdown