Claude Fable 5: No Hype Full Breakdown & Testing

All right, well today Anthropic released
apparently the most powerful AI model
the public has ever been allowed to
touch. It's called Fable 5 and it is a
version of Mythos, the model that
Anthropic said was too dangerous to
release to the public back in April. AI
internet is probably about to blow up
with these this changes everything type
post, but for most of us, spoiler, it's
probably total overkill. So in this
video I'm going to show you what Fable 5
is, how it's different from Opus 4 8,
and whether it's worth your time before
it stops being free on June 22nd. And
then finally at the end we'll do a live
build to see how good Fable 5 really is
in action. Not wasting any more time,
let's dive right in and rather than
scroll through Anthropic's announcement
line by line, I put together a little
interactive presentation of what caught
my attention and what's actually
relevant to you, which by the way I had
Fable 5 design this. So if anything
looks like AI slop, just blame Fable 5,
obviously not my fault. But back in
April, Anthropic launched a model called
Mythos preview, which you're probably
familiar with, and they only made it
available to a handful of handpicked
partners through a project called
Project Glasswing, and they showed off
all these demos of how incredibly
capable was, mainly in finding security
vulnerabilities, but they said it was
too dangerous to release the public. You
like this too dangerous here by the way,
this nice little Fable 5 really wanted
to emphasize that. Honestly, this is a
one-shot output too of this entire
presentation. So it's it's did a pretty
good job. It did much better than Opus 4
8 just in terms of presentation design.
We'll get into all the use cases in a
second. Anyway, back to the comparison
of Mythos 5 and Fable 5. So today these
Mythos 5 restrictions to the public have
finally changed, sort of. So Anthropic
just dropped two new models in that same
Mythos tier. The first is Mythos 5,
which is an even bigger step up from
Mythos preview, which again was released
in April, and it's still locked to those
Project Glasswing partners. So that's
not available to the public. The second
one is Fable 5, which is available to
us. So Fable 5 is that same Mythos model
just with safety guardrails bolted on,
which are basically filters that stop it
from helping with things like cyber
attacks. So Fable 5 and Mythos 5 are the
same model underneath more or less. That
is why Anthropic is listing them
together. On their own benchmarks, We
look at those. With those fives, Fable
five. To connect to yourself, it's super
easy to do. So, let's go ahead and just
do that now. I have cursor open here.
You can see the announcement, meet Fable
five, our newest model for complex long
running work. Try anytime with {slash}
model. So, you just go {slash} model,
select Fable, and just like that we are
on Fable five. And it's included right
now in every paid subscription. So, pro,
max, team, enterprise. Although, given
how token intensive it is, you probably
won't get much use out of it on a pro
account. Which leads us straight into
the first big caveat. It's only
available on your subscription until
June 21st. Then it will switch to usage
credits, which I will talk more about
later on. So, is Fable five actually as
good as everyone's saying it is? Judging
by the benchmarks, yes, but with a
couple of real caveats. The first is
myself and a lot of folks just don't
trust these benchmarks as much as we
used to because they're so gameable.
Specifically, Swebench. So, I'm just
skipping right past that. But if you're
sure about the hype, here's what the
rest of this really means to you. So, by
the way, we're just going through these
benchmarks here. Just a better visual,
which by the way, Fable five, crushing
it. Crushing it so far. This is the
first time I'm seeing this. Look at
these little moving bars here. Great
job. First up, the everyday work you'd
actually do. So, knowledge work in
computer use. We just go back to the
benchmarks cuz I want to actually show
these numbers. Not much of a change
here. 1932 versus 1890. 85% in computer
use versus 83.4%.
Go back to my better looking chart here.
Now, the one place this jump is massive
is cybersecurity, which, no surprise, is
exactly what the model is built for. But
unless you're, I don't know, a
cybersecurity researcher, or if you're,
I don't know, patching security
vulnerabilities in your vibe coded app,
I don't see much use in that. Not much
would be touching your day-to-day in a
cybersecurity sense. The one real
exception that matters though is coding.
And on the harder coding task, Fable
five is roughly twice as good as Opus
four A. Like, that is just a crazy jump.
Sorry to keep going back to the numbers,
but Fable five, my designer did include
these numbers in the presentation. So, I
got to zoom in on this. But, agentic
coding frontier code 13.4% from Opus
4.8, 29.3% from Mythos. So, pretty darn
impressive of a leap, which is useful if
you're live coding a complex app, if
you're trying to one-shot something, if
you're trying to refactor code, trying
to do a massive code migration, this
could be incredibly useful, which we'll
be testing with a live demo here in a
second. But, of course, all of this
assumes you can actually afford to run
it, and that is where the shine starts
to wear off. And it wears off fast for
two reasons. First, how much it costs,
and second, how fast it actually burns
through tokens. So, Fable runs about
twice the price of Opus 4.8. So, $5 per
million input token, $10 on Fable 5, and
then 25 per million output token, and
then $50 on it Fable 5. Which, I guess,
if you're paying for twice the
capability, makes sense to be paying
twice the cost. But, that is a big, big
jump, especially after June 21st, when
you start to use this on your usage
credits. This is going to add up really,
really quick. And, you know, just a
random aside, but you really got to
respect the move by Anthropic, releasing
a model twice as good as your last one,
giving it away just long enough to get
everyone hooked, like any good drug
dealer, then start charging the minute
you're addicted. Anyway, the real
question isn't whether it's good, it's
whether is this actually worth it? And
the only honest way to answer this is
just run it once. So, let's do a real
build start to finish, and I'll show you
the token usage at the very end. And for
this demo, we're going to have fun with
it. You know, Anthropic was having fun
with their demos. I I actually really
did appreciate some of these. It wasn't
like boring tasks, which can be useful
and tactical, but they talked about it
playing like Pokémon games, it
simulating the solar system and
predicting a solar eclipse, pretty cool
little visual there. There's it playing
Pokémon. Oh, they chopped that down, not
sure why. And then they had a bunch of
other examples here, like just it
playing video games non-stop, so it
using its vision and intelligence. And
then this one was pretty cool, too. This
is a CAD, like almost Blender, it making
its own version of Blender, and then
having an ability to chat with it. It's
like a lovable and a Blender combined.
And then this I didn't know really know
what what going on with it, but it
looked it looked really cool. So, we're
going to have fun with these demos, too,
and I'm going to give Fable one prompt.
I'm not going to give any image
references or anything like that, but
what we're going to do is ask it to
build a fully playable RollerCoaster
Tycoon game in a single shot. Let's see
how it does. So, we'll go over to Cursor
and paste in this prompt, and just show
you what the prompt is. Build me a
complete playable RollerCoaster
Tycoon-style theme park management game
in a self-contained HTML file, no
external dependencies, core
requirements, grid-based park I can
build on, I can lay down paths, place
rides, shops, guests walk in this, ride
attractions, a working economy, classic
pixel sim visual style. Plan it out
first. So, let's go plan mode to really
get this thing humming. This is probably
going to take hours, but it'll be a good
test. While it builds, too, fun fact,
RollerCoaster Tycoon was built almost
entirely by one person, and it's still
considered one of the most impressive
low-level coding feats and in gaming
history. We'll see how Fable can match
up. And at about 36 minutes, there's
also about 10 minutes of planning, so a
little over 45 minutes of coding. Here
is Fable 5's one-shot output of the
RollerCoaster Tycoon demo. So, we'll go
to this link here. All right,
pixel-perfect Tycoon Pixel Park Tycoon.
Park rating grand 600. How to play? Pick
the tools, place rides and shops. The
coaster tool builds a custom roller
coaster piece by piece. Park button sets
the entry fee. Controls. All right,
let's open the park. Okay, got kind of
this infinite canvas park builder.
Pretty cool.
Where are the controls? Oh, there they
are, at the bottom here. Okay, path, I
can add a path here. Uh let's add a
burger stand. Burger stand's in there.
We just got to get the path to the
burger stand complete. Okay.
Uh what else? Let's get a Ferris wheel
in here. Nice path. I don't know where
all the people are. I guess I need to
hit this play button, maybe. Okay, there
we go. All right, now the people are
starting to come in. There they come.
Let's
put in a put in a soda stand, too.
Teacup. Wow, this is This is impressive.
Okay, let's look at the roller coaster
the actual builder. So, all right, build
out the roller coaster. Okay, so we have
this little pop up here coaster builder.
Left ahead right. So, if I just go and I
can place it again. So, I go ahead ahead
ahead and then up up up again up again.
We'll go right right down down. Oh,
can't do that. Left ahead up. Like this
is so impressive. Good God. Like it is
building a full roller coaster. I I
don't know how to complete this, but
closed loop. Oh, I need to I need to go
back. Okay, left. I got to go up. Nope,
undo. Maximum track height reached.
Down. I mean, just need to have Fable 5
actually use its vision to make this
roller coaster. But this is pretty darn
impressive. This is a full roller
coaster tycoon. Now, the the graphics
could use a little bit of work, but like
this is freaking working in literally
one shot. So, let's go back. I just want
to have Claude finish this up. Nicely
done. Now, can you use your vision to
just create the park? You can delete
some of my stuff. I don't really know
how to play this game, but I Can you Can
you create the roller coaster? And then
while it does that, I was looking at
usage earlier. So, it took about 45
minutes. And then if we look at actual
token costs and usage, so that single
session, 45 minutes, used about 13% of
my max plan usage. And the cost, I don't
know where the cost went. I saw cost
earlier at $18. I forgot I screenshotted
this usage. So, it's $19 total cost. So,
that's likely what I would have paid in
usage credits, which is quite a bit,
especially if you're token maxing, you
could easily get to hundreds of dollars
a day after this June 21st cutoff. So, I
mean, overall pretty impressive from
this demo. Also, check out what Claude
just did when I asked it to build out
further. Good vision. Made that roller
coaster humming. We're 22 grand. Wow, we
even have This is crazy. There's little
people talking. There's little messages.
Haunted house was great. Insane. Well
freaking done. So, overall, pretty darn
impressive from this one demo. I still
think that in your day-to-day to be
using Fable 5 is pretty crazy. It's just
going to be overkill, but to even see
this level of capability is pretty fun.
So, start hammering those builds now.
Get your, you know, refactors out of the
way. Start building that app you've been
sitting on because after Jan 22nd, all
of that gets a lot more expensive. And
this is just one demo, too. I will be
putting together a follow-up video where
I run Fable 5 and Opus 48 head-to-head
in a bunch of different real-world
building tests. So, be sure to subscribe
if you haven't already so you can catch
that one. And happy building for the
next 12 days. I'll see you in the next
one.