Why Modern Creator?

Jesus Martinez · YouTube

Claude Fable-5 Is So Dangerous They Had To Lock It Down

A first-look reaction to Anthropic's Mythos-class model — what got locked out, what slipped through, and whether the pricing makes it worth it.

Posted

June 9th

3 months ago

Duration

08:31

Format

Reaction

educational

Views

2.9K

42 likes

Part of the collectionThe Fable 5 PlaybookAll 45 Fable 5 breakdowns, synthesized into one page.

Read the playbook

Big Idea

The argument in one line.

Fable-5 is a deliberately hobbled Mythos model — Anthropic redirects its most dangerous capabilities to Opus 4.8, which means you're paying frontier prices for a frontier model that silently downgrades itself on the exact tasks it was built for.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You're a developer or founder deciding whether to upgrade your Claude API tier to Fable-5.
You're already using Opus 4.8 for agentic coding and want to know whether Fable-5 is worth the 2x price jump.
You run a SaaS product and want to know whether Fable-5's bug-finding capability is relevant to your security posture.
You're curious about the business dynamics of Anthropic holding back the full Mythos model from general release.

SKIP IF…

You're not already using Claude APIs — this is a pricing and capability comparison, not an intro to LLMs.
You need a formal security audit — this video is commentary, not a cybersecurity methodology guide.

TL;DR

The full version, fast.

Anthropic released Claude Fable-5, a consumer-safe version of their Mythos-class model, after deciding the full Mythos Preview was too dangerous for general access due to its ability to discover thousands of cybersecurity vulnerabilities. The public version routes sensitive cybersecurity and biology queries through Opus 4.8 as a safety harness — meaning Fable-5 silently degrades on its headline use case. For most builders, the $10 input / $50 output per million token pricing is steep; Opus 4.8 remains the better value for ordinary tasks. The full Mythos 5 is being released only to vetted cyber defenders and critical infrastructure operators.

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 – 00:52

01 · The release and the hype

First-look introduction, confirms Fable-5 is live on claude.ai with effort levels from low to max, notes agentic coding not yet showing in the platform.

00:52 – 02:42

02 · Anthropic's official announcement

Reaction to the official launch video — why Mythos Preview was never released (cybersecurity vulnerability discovery), how it was used to fix critical software holes before public launch.

02:42 – 04:10

03 · The Opus 4.8 fallback harness

Explains the safety routing: cybersecurity, biology, and chemistry queries are silently rerouted to Opus 4.8. The safeguards intentionally over-trigger, catching harmless requests too.

04:10 – 05:05

04 · Mythos 5 vs Fable-5 — who gets what

Full Claude Mythos 5 (unconstrained) is only going to cyber defenders and critical infrastructure operators. The public gets the hobbled version.

05:05 – 07:00

05 · Benchmark review and first impressions

Walks through benchmark charts — agentic coding and cybersecurity are the standout gains; questions the real-world utility when those areas are harnessed on the consumer model.

07:00 – 08:31

06 · Pricing breakdown and final take

$10/$50 per million tokens — double Opus 4.8. The Stripe use case (50M line Ruby migration in 2 days) is the real target. For most people, Opus is still the right choice. Final verdict: important release, real impact mostly for power agentic coding users.

Atomic Insights

Lines worth screenshotting.

Anthropic never publicly released Claude Mythos Preview — the model found thousands of real cybersecurity vulnerabilities during testing, so they handed it to defenders to patch holes before release.
Fable-5 is the consumer-safe version: any query touching cybersecurity, biology, or chemistry is silently re-routed to Opus 4.8 instead of Fable-5.
The safety harness deliberately over-triggers — Anthropic admits it will catch harmless requests in those domains, not just malicious ones.
Claude Mythos 5 (the unconstrained version) is only being released to a small group of cyber defenders and critical infrastructure providers — the public never gets it.
Fable-5 pricing is $10 per million input tokens and $50 per million output tokens — less than half the price of Mythos Preview, but still double Opus 4.8.
Stripe tested Fable-5 on a 50-million-line Ruby codebase and compressed months of engineering into two days — the edge case is long-context, multi-week agentic projects.
For regular writing, conversations, or short coding tasks, Opus 4.8 is still the better value — Fable-5's advantage only shows up on complex, autonomous, multi-day work.
Opus 4.8 already exposed a $10 billion protocol vulnerability in a crypto project after four years of it going unnoticed — Fable-5 pushes that capability further.
Agentic coding on Claude.ai was not yet available at launch time — the model shipped without full platform integration.
Anthropic doubled Claude.ai usage limits for a few days after the Fable-5 launch to let people test it.

Takeaway

Fable-5 does less than its benchmarks suggest

WHAT TO LEARN

A model trained for cybersecurity that silently reroutes all cybersecurity queries is a different product than its benchmarks imply — and the price doubles to get there.

The safety harness on Fable-5 intentionally over-triggers, meaning you can pay frontier prices and get Opus 4.8 responses on the exact domain the model was marketed for.
Anthropic's decision to hold back Mythos Preview was the right call — a model that can autonomously find thousands of security vulnerabilities is a dual-use weapon, and they patched holes with it before releasing a hobbled version.
The $10/$50 per million token pricing only makes economic sense for long-context agentic projects — multi-day autonomous coding across massive codebases, not daily chat or short coding tasks.
The full Mythos 5 exists and is in the hands of vetted cyber defenders; the consumer version is structurally a different product despite carrying the same lineage.
Opus 4.8 is already finding nine-figure crypto vulnerabilities — the capability floor is already high enough for most security-adjacent use cases without upgrading to Fable-5.

Glossary

Terms worth knowing.

Mythos class: Anthropic's internal capability tier for their most powerful models, above the Opus tier. Mythos-class models have capabilities Anthropic considers too risky for unrestricted public release.
Fable-5: The publicly released, safety-harnessed version of Anthropic's Mythos model. High-risk queries are redirected to Opus 4.8 rather than handled by Fable-5 directly.
Mythos Preview: The pre-release version of the Mythos model that Anthropic never publicly launched after discovering it could autonomously find thousands of cybersecurity vulnerabilities.
Safety harness: A routing layer that intercepts queries in flagged domains (cybersecurity, biology, chemistry) and redirects them to a less capable but safer model — in Fable-5's case, Opus 4.8.
Agentic coding: AI-driven software development where the model autonomously plans, writes, tests, and refines code over extended sessions with minimal human intervention between steps.
Long context: The ability to process and reason over very large amounts of text or code in a single session — Fable-5 supports millions of tokens, enabling it to work across entire codebases.

Resources

Things they pointed at.

00:00productClaude Fable-5 ↗

00:52linkAnthropic Fable-5 launch announcement ↗

07:10productStripe Fable-5 case study (Ruby codebase migration)

Quotables

Lines you could clip.

01:03

“When we finished training and testing it, we saw that the model, Claude Mythos Preview, was finding thousands of cybersecurity vulnerabilities.”

The specific claim that explains why Anthropic held back the full model — this is the news.→ TikTok hook↗ Tweet quote

03:03

“Fable-5 uses this fallback mechanism because they don't trust us, essentially.”

Blunt editorial summary of the safety harness — quotable take.→ IG reel cold open↗ Tweet quote

07:10

“Stripe reported that Fable-5 compressed months of engineering in two days in a 50,000,000 line Ruby code base.”

Concrete proof point with a memorable scale — 50M lines, 2 days.→ 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.

story

We just got the biggest AI release of the year so far, a mythos class model. Keep in mind, this is a model that was so powerful that they had to make a dung down version to give to us is now out. This is Claude at Fable five.

I'm gonna be going over it, what we know so far, etcetera, etcetera, etcetera. Now you can actually use this live right now. You can, you know, click over here on new chat, click on Fable five.

There are different effort levels all the way from low to max, so test it out yourself. I couldn't get it working on code so far. I don't know why it's not there.

I'm assuming it's because they're still updating the app and stuff. But on chat, at least so far, it is it is indeed there. So let's see from Anthropic themselves what this big old model has to provide for us and what actually is BS and what actually is not BS.

Today, we're launching Claude Fable five,

the most capable model we've ever released to the public. Fable five is a Mythos class model with safeguards that make it ready for general use. We didn't broadly release our previous model with this level of capability because when we finished training and testing it, we saw that the model, Claude Mythos Preview, was finding thousands of cyber security vulnerabilities.

A model that can find flaws like that can also be used to exploit them. So instead of releasing it, we handed it to the people who protect the world's critical software and put it to work fixing the holes before someone could break through them. It was the right call for the moment, but it was never the goal.

We believe powerful AI should be safe and accessible. That's why we went to work on Claude Fable five.

Every Claude model has safeguards to keep it from doing harm. Fable needed more cautious ones than anything we'd built before. Our safety systems for Fable five automatically review requests that touch on high risk areas like cybersecurity or biology.

Those requests are then redirected to OPUS 4.8. We do that intentionally so people can continue to benefit from the capabilities of a powerful model like Fable without the cyber and biology risks that come with it.

The safeguards are broad today, but we'll keep refining them so that they're better at allowing safe requests.

We built Claude Fable five for your most ambitious work. It can stay with a problem far longer than any model before it. It's highly autonomous and can operate for days without intervention.

And it's not just coding. It can take on projects in finance, research, economics, law, complicated tasks that used to need constant supervision. So point it at something that matters.

What's the problem we'll look back on and wonder why it took so long to solve? We know what Cloud Fable five can do. The interesting part is what Very,

very, very, very, very interesting. So immediately, I don't know if you saw this, but midway through, it had this little graphic between Fable five and OPUS 4.8.

So anytime you get any questions, if you ask any questions related to cybersecurity biology and chemistry or distillation, the response isn't handled by Fable five.

In fact, it's handled by OPUS 4.8. Very, very, very interesting.

So the idea is apparently that if there's any questions that are risky, like like any exploits, I guess, until people probably break it, Then, um, uh, Fable five uses this fallback mechanism because they don't trust us, essentially.

The safeguards are tuned conservatively and over trigger, so they'll sometimes catch harmless requests. Okay.

Interesting. But I mean, I I don't know. Isn't the entire point of of this model, the whole cybersecurity component?

You know? Okay. Interesting.

So they also, for a small group of cyber defenders and critical infrastructure providers, are also releasing Claude Mythos five. So Claude Mythos five is the full broader model with the extensive cybersecurity work and biomedical research.

So, you know, they're not giving us the full thing, but they're giving us some of it. So what are the specifications surrounding it?

What do we have here? We've got agentic coding, you know, a lot better, actually. We have agentic coding a lot a lot better on extra.

Wow. That is so much better on extra high. Look at that.

Knowledge work is also okay. Who cares about that? Spatial reasoning seems to be better.

Tool use seems to be marginally better. Computer use is, you know, a little bit better. Legal is a little bit better.

Multidisciplinary reasoning seems to be quite a bit better.

Cool. Biology seems to be a little bit better. Identic coding seems to be quite a bit better.

Cybersecurity is really good. And then health is good. So this is, of course, the cybersecurity slash health focused model.

Now keep in mind, this model is also don't know if you were looking at what I said before, but, um, limited in capabilities in specifically cybersecurity and health.

So I've gotta be honest. Uh, first look in, it doesn't seem like the most productive model in the world. It doesn't seem like the most productive model in the world.

It's impossible to test agentic coating at the moment because it's not available on the platform. The pricing is pretty high. The pricing is pretty high here.

Now just looking into their actual news release here on the cost for everything, you can see here, during their early testing, Stripe reported that Fable five compressed months of engineering in two days In a 50,000,000 line Ruby code base, the model performed a code base wide migration in a day that would have otherwise have taken a whole team over two months by hand.

So it seems like for long context tasks, so longer term tasks, this is going to be a pretty solid model.

Also, apparently, for agentic coding, this is specifically going to be good. So for anything that seems like not simple, anything that seems complex, seems like it'll be done using Fable five.

For cybersecurity stuff, for health related tasks, you probably shouldn't use the Fable five model because it's just gonna fall back, you know, unless you can kinda play around with it and see if you can find a way to break through this harness. You know, it's gonna be annoying. You know, I'm sure people are probably gonna figure out a way to break through this.

For memory long context, Fable five c is focused across millions of tokens. It's a very expensive model. So this model is a a very expensive model.

It is $10 per per million input tokens and $50 per million output tokens, less than half the price of Claude Mythos preview. Less than half the price, guys.

Guys, the price. So that means Mythos is gonna be way more expensive.

Okay. You know, this is for those that don't have a budget. Okay.

Unlimited budget. They just want the best model. Well, it's gonna be expensive.

So Fable's double the cost of Opus 4.8. So if you were having trouble with Opus, you're gonna have a lot more trouble with Fable. To be fair, on CoWork specifically, I think they've doubled the usage for the next couple of days.

So that that's cool. That's very cool. Sorry, guys.

My my dog came in here. You know, hello to the dog. Right?

She's little dogger. She's cute. But, you know, in general, this is going to be a big deal for anybody that's coding probably.

This is marginally better model. It it catches a lot of bugs. I think, um, recently, I'm in the crypto space, and we had a $10,000,000,000 protocol expose vulnerability for infinite coins.

Infinite coins, it's been around for four years. And it was exposed by Opus four point eight. So imagine with Fable five, it's gonna get it even crazier.

It's gonna become more accelerated. I mean, I would if I were you and I had a SaaS software business, I would immediately plug in Fable five and see what exploits you can, you know, find because this is going to definitely help you out in that regard for normal conversations.

And, uh, you know, agentic bots, this could also be an interesting one too. I just think it's gonna be too expensive for most people. So for most people, this is probably not a great model for you.

I don't I don't think Opus is bad. You know, for regular tasks or just writing, you know, for regular tasks, you don't need any agentic coding. Opus is gonna be perfectly fine for you.

And then, you know, other than that, you're gonna be using Fable five for agentic coding, cybersecurity, health. This is a pretty big, you know, release. I I don't think it's gonna be that that that that big of a deal until the actual mythos comes out.

But nonetheless, we finally got at least a taste of what's to come. And I imagine everybody, especially people looking into the Anthropic IPO, are very interested in seeing just how many users this brings over to Anthropic. So thank you for watching this video.

Quick one getting it out there, so you now know about Claude Mythos for Fable five. Until next time, stay classy and leave a like for the little dogger.

The Hook

The bait, then the rug-pull.

Anthropic's most capable model ever shipped with a catch: the version that could find thousands of real security vulnerabilities never reached the public. What landed instead is Fable-5 — a safety-harnessed Mythos-class model that quietly hands off its most dangerous queries to Opus 4.8.

CTA Breakdown

How they asked for the click.

VERBAL ASK

08:13subscribe

“Thank you for watching this video. Quick one getting it out there. Until next time, stay classy and leave a like for the little dogger.”

Casual sign-off with dog reference — low-pressure, personality-driven.

MENTIONED ON CAMERA

00:00productClaude Fable-5 ↗

00:52linkAnthropic Fable-5 launch announcement ↗

Storyboard

Visual structure at a glance.

intro — host announces the release

hookintro — host announces the release00:00

Anthropic promo video reaction

contextAnthropic promo video reaction00:52

Fable 5 / Opus 4.8 routing diagram

revealFable 5 / Opus 4.8 routing diagram02:42

Anthropic benchmark page on screen

valueAnthropic benchmark page on screen05:05

pricing comparison table

ctapricing comparison table07:00

sign-off

ctasign-off08:10

Frame Gallery

Visual moments.

intro — host announces the release

Frame at 00:09 from Claude Fable-5 Is So Dangerous They Had To Lock It Down

Frame at 00:15 from Claude Fable-5 Is So Dangerous They Had To Lock It Down

Frame at 00:22 from Claude Fable-5 Is So Dangerous They Had To Lock It Down

Frame at 00:28 from Claude Fable-5 Is So Dangerous They Had To Lock It Down

Frame at 00:35 from Claude Fable-5 Is So Dangerous They Had To Lock It Down

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.

12:27

Alex Finn · Review

Claude Fable 5.1: what actually changed, and three ways to use it today

A benchmark-by-benchmark walkthrough of Anthropic's Fable 5.1 release, plus a custom test suite proving it beats GPT-5.6 Sol and demolishes Fable 5.

September 1st

09:31

Every · Review

We Tested Claude Opus 5. It's Frustrating with Flashes of Brilliance.

Every's Dan Shipper spent a week with Opus 5 and comes back with a mixed verdict: pushy, prone to quitting early, and a real pain if you built workflows around Opus 4.8.

July 24th

28:37

Theo - t3․gg · Talking Head

FABLE IS BACK! (And Sonnet 5 is here too)

A 28-minute benchmark teardown of Claude Sonnet 5, plus the government letter that brought Fable back from the dead.

July 1st

10:14

Chase AI · Tutorial

How Claude Fable 5.1's Text Watermark Actually Works (And How to Get Around It)

It's not a hidden mark in the text, it's a loaded die on word choice, and only a key Anthropic won't hand out can read it.

September 2nd

21:43

Brock Mesarich | AI for Non Techies · Review

I Tested Fable 5.1 vs Fable 5 vs Opus 5: Cost, Speed, and Design

Twelve identical builds against one prompt, scored blind, to find out whether the newest model's cheaper caching actually lowers what you pay.

September 3rd

13:15

Ben AI · Tutorial

Anthropic's 7 New Rules for Prompting Claude 5 Models

A YouTuber pulls Anthropic's own internal prompting guides for Opus 5 and Fable 5, cross-references them against Claude Code creator Boris Cherny's Y Combinator talk, and distills them into seven concrete habits.

September 2nd