Anthropic just released their latest and
best model, Claude Fable 5. And get
this, it's even more powerful than their
Mythos preview model, which they
previously said was too dangerous to
release. So, in this video, I'm going to
go over all the impressive things that
Fable 5 can do, plus its limitations,
and also go over its specs and pricing
and benchmarks and how it compares to
the other leading models. Let's jump
right in. Thanks to HubSpot for
sponsoring this video. Instead of just
talking about benchmarks, let's start
with a really challenging test that no
AI model has gotten correct so far. I'm
going to feed Claude this image. There's
a frog hidden somewhere in this image.
If you don't want me to spoil the
answer, feel free to pause it right now
and look for the frog first. All right,
so I'm going to drop it in here and then
ask it to find and circle the frog
that's hidden in this photo. I'm going
to set the effort to fable 5 max and
then press run. You know, wouldn't it be
funny if I was just trolling and there's
actually no frog in this image, but
y'all are just staring at this trying to
find the frog? But no, there's actually
a frog in this image. All right, let me
expand its thinking process so you can
see how it did this. So, it's actually
zooming into different quadrants to
locate the frog. So, it first splits the
image into different sections and then
it's inspecting each region and then it
spotted a brown frog at exactly these
coordinates and it drew a red circle
around the frog. That's pretty much it.
So, if we expand this, you can see that
indeed it has circled the frog
correctly. Oh my goodness. Let me expand
this image for you. So, you can see that
the frog is indeed over here. Here is
its head and here are its hind legs.
This is the first model ever to actually
get this prompt correct. So, it looks
like I need to make an even harder photo
for the next model. Now, the strength of
Claude Fable and even the top Opus
models isn't just getting it to answer
simple questions in the Claude
interface, but they're actually designed
to do really well at agentic coding. So,
what I'm going to do instead of just
using the regular Claude chat, is to use
Claude code. So, this is just the Claude
Code desktop app, which you can download
for free. And let's try this prompt.
Develop a ray tracing simulation
featuring one sphere, one cube, and one
pyramid. The environment is a blue sky
with a checkered ground. Add adjustable
parameters such as position,
reflectivity, roughness, transparency,
and other material properties for each
shape. Put everything in a standalone
HTML file. But the most important part
is this. Do not use 3JS or other
external libraries. It needs to code up
this entire ray tracing simulation from
scratch itself. This tests its
underlying understanding of physics and
lighting. I'm just going to set this to
fable high. and then press enter and see
what that gives us. All right, so in one
shot it already gave me a pretty decent
looking simulation. But it was kind of
laggy, so I prompted it further. It
works well, but there's some latency.
Can you make this even more efficient on
a web browser and then it continued to
add some more latency optimizations and
that's it. In just two prompts, here's
what we got. Let me open this in a new
browser and let's test this out. So I
can indeed drag around the simulation.
And there is like a few milliseconds of
latency before it renders the full
simulation. Let's play with the sphere
first. Let's drag these position
sliders. All right. So these position
sliders work. Let's also change the
color to white. Radius also works. Very
nice. And let's change the reflectivity
to full reflectiveness. And as you can
see now it's super reflective. So that
works. Let's change the roughness. That
also works. Very nice. Let's decrease
the roughness back to zero. And let's
increase the transparency all the way to
one and reflectivity to zero. So now
this is fully transparent. Very nice.
And then let's also change the IUR which
does this. So all these settings work.
Very nice. Next, let's move on to the
cube. So let me also adjust the position
of the cube. So these settings work. And
actually what I'm going to do is for the
sphere I'm going to set this to
incredibly reflective so we can see the
reflection of the cube within the sphere
to make sure that the reflections are
also correct. All right. So let's set
the color of the cube to like red for
example. And as you can see the
reflection of the cube on the sphere is
also changed to red. Let's also increase
the reflectivity. So reflectivity works.
Let's increase the roughness. Roughness
also works. And let's set this cube to
white actually. And then let's also play
with the transparency and the IUR
settings. So that also works. Very nice.
And as you can see the changes are also
applied to the reflection on the sphere.
So everything works. Final test. Let's
also make sure that the pyramid works.
So we can change like the base width.
Very nice. As well as the height and the
rotation. And all the reflections within
the shapes also work. Very nice. And
then finally for environment, let's see
what this does. There's also all these
different settings. For example,
changing the sun's position. Very cool.
Also the checker size. Wow. And then
also the ground reflect. So, I mean, it
pretty much aced this simulation. And it
only took it like one or two prompts.
Notice that here after each round, it's
also automatically verifying its fixes
and testing it in the browser to make
sure everything works. And if it catches
anything, it's going to fix those errors
as well. So, this is incredibly robust.
However, it does tend to take up way
more tokens because it's autoverifying
everything. All right, that was too
easy. So, next, let's try this. Build a
fully interactive 3D digital twin of
Earth with the following features. Allow
users to zoom seamlessly from outer
space down into individual city streets.
Now, GPT 5.5 can already do this, so
that's not too impressive. Here are some
even more challenging features. When I
hover over a country, the country's
outline should be highlighted and there
should be a popup displaying stats like
area, population, and GDP. All right.
Plus, we got to add more features to
really test the hell out of it. So,
let's also include toggles for
atmospheric cloud cover, flight traffic,
and day and night toggles. Night mode
should show city lights. Use publicly
available assets and layers if needed.
Make sure it loads efficiently on a
regular web browser. Let's press run.
All right, that's pretty much it. In
just one prompt, I didn't really need to
prompt it further. Here's what we got.
Now, let me open this up in full screen.
Perfect. Now, if we hover over any
country, you can see the country's
outline is highlighted. Plus, it gives
us the various stats like population,
area, GDP, etc. And then here's another
example. All of these just work right
out of the box. Very nice. Now, let me
toggle off the country borders. And
next, let's toggle on cloud cover and
see what that does. As you can see, it
has added some cloud cover to our planet
Earth. Next, let's also add flight
traffic and see what that does. Wow,
this is super cool. So, notice that it's
like adding these flight trails across
the planet. And then let's also turn on
place labels so you can see the names of
all these different countries and where
the flights are coming from. So, it's
able to accurately label all the
countries. All right. Next, let me turn
off flight traffic and place labels. And
then for this day night terminator, if I
turn this on, then it basically divides
the Earth into day and night. Or if I
turn this off and I turn on night mode,
then it just turns the entire Earth into
night mode. And you can like zoom in and
even see the city lights. All right.
Next, let's zoom way into somewhere like
Tokyo, I think. And indeed, it allows me
to zoom way in to like this city street
view. Really cool. So, in just one
prompt, it's able to generate this super
realistic digital twin of Earth with
cloud cover, flight traffic, day and
night toggles, even country stats, and
labels. This is incredibly complicated
to build in just one prompt, but Fable
was able to ace this. Very impressive.
You've probably heard of how AI tools
are transforming marketing, but if
you're not sure how to actually use them
to build real campaigns without hiring
developers, this is the ebook for you.
Check out Mastercloud Code for
Marketing. Four workflows plus 12
prompts to ship campaigns fast by
HubSpot. This bundle includes four
workflows that are like a cheat code for
building and deploying entire marketing
campaigns from your terminal. You'll
learn how to conduct competitor
research, craft differentiated
positioning, build interactive lead
magnets, and deploy landing pages, all
without writing a single line of code
yourself. You get 12 readytto-use
prompts to help you unlock Claude Code's
full potential for marketing tasks. From
researching competitors to writing email
sequences that actually convert. Use
them to eliminate the guesswork and get
highquality outputs faster. My favorite
one is this oneshot landing page builder
workflow. It shows you how to go from
positioning to a fully deployed landing
page in a single session. Complete with
distinctive design, mobile
responsiveness, and clear conversion
paths. You can access all of this for
free using the link in the description
below. This resource was made by
HubSpot, the sponsor of this video. All
right, next. Let's see how good it is at
reconstructing a 3D scene from just
reference image. So, I'm going to input
this image, which is really complicated
with a ton of office desks and items and
people. None of the other top models out
there could really code this up well.
I'm going to feed this into claw code
and then get it to create a beautiful 3D
animated scene from this image. Use a
single HTML file. All right, here's what
I got. It did code up something pretty
nice, but the tables and the desks
weren't really aligned with the input
image. So, I wrote make sure the desks,
items, and people are aligned exactly
like the reference image. And then it
proceeded to code this up further. And
then it tried to verify everything
itself. And it also fixed some bugs
apparently. And that was it. So I only
used two prompts on this. And here's the
result. Now if you compare this with the
original image, then there's still some
discrepancies. So the center table with
the rows of three chairs, that's
correct. But the rest of the desks
aren't really aligned properly with the
original image. So you know, this human
should not really be here. There should
be more desks. And then it decided to
add this random space here, which is not
correct. So, it's not perfectly aligned
with what I have in the original image.
So, I would say this generation is a
fail. However, if you compare this to
the generation from GPT 5.5, then it
does look a lot better and more
coherent. For GPT 5.5, there's just a
lot of disconnected pieces, whereas this
one from Fable has overall fewer errors.
Next, here's one of my classic tests to
see how good it is at music composition.
First, I'm going to get it to make a DAW
interface with these instruments. A
piano, synth, pluck, strings, drums, and
bass. For each instrument, there should
be a piano roll interface where I can
drag and drop notes on the timeline.
Each track should have pan, volume, and
other standard settings. Add play,
pause, and other settings. So, let's
click run first. And with just one
prompt, it did give me a fully working
DAW with five tracks and all these
different settings. And then I got it to
actually compose a song. So I wrote by
default show a powerful expressive 32
bar song using all the instruments rich
in complexity that would win a Grammy
include effects automation panning and
make sure everything is mastered well.
So let's open this up in a browser and
play it. All right, here's our Grammy
Award-winning track.
[music]
[music]
[music]
Heat. Heat.
[music]
Heat. Heat.
[music]
[music]
[music]
>> [music]
>> It didn't sound too impressive. I mean,
it's just playing the same chords over
and over again with some slightly
different variation in the notes.
Definitely not enough to win a Grammy.
And, you know, interestingly, it sounds
very similar to the composition from GPT
5.5 when I tested it on the same prompt.
[music]
>> [music]
[music]
>> Also, it did not really include
automation and panning. Although it did
include some effects like delay and
reverb, everything doesn't really seem
to be mastered well. So, in terms of
music composition, I would say Claude
Fable isn't really like prolevel yet.
All right, here's a fun one. Let's see
how good Fable is at creating video
games. So, let's write create a 3D game
using 3JS. It should be a futuristic
battlefield where I control a Mecca
warrior and shoot down waves of alien
creatures attacking from the sky and
ground. Third-person shooter
perspective. Use publicly available 3D
assets. Make it look amazing like a pro
AAA video game. And in just one prompt,
here's what I got. I can move and do
stuff with all these different keys. So,
I can also shift to sprint and space to
jump or use jet thrusters. Pretty cool.
Let's deploy this and start playing. All
right, you can see that everything
works. Oh my goodness. There's like this
beetle thing attacking me. And then
there are these birds in the sky that
are also trying to attack me. Oh my
goodness. But everything works. This is
very nicely designed. This is a 3D
thirdp person shooter game. And let's
try pressing space. And indeed, I can
kind of fly up. Really cool. And
eventually my HP goes to zero and it's
game over. So everything works. Notice
that it even includes like multiple
levels or waves where it gets harder and
harder. There are more enemies. So for
example, I made it up to wave three. So
in just one prompt, this works really
well already. I didn't even need to
prompt it further. All right. Now that's
the first iteration. But remember in the
prompt I said I told it to use publicly
available 3D assets. Well, down here
it's basically being extra cautious. So
here it says it avoided any external
thirdparty models for now, but if you'd
like I can swap in some creative common
models for you to use. So I wrote, okay,
let's proceed with this and make the
game design even better. So afterwards,
let's look at the revised game. Notice
that it even added more stuff like
thruster dash with the Q key as well as
1 2 3 for field upgrade, whatever that
is. Let's click deploy and run this. And
instantly, notice how everything just
looks a lot better. Oh my god, there's
like this spider attacking me now. This
is pretty intense just for round one.
There's like wasps everywhere. Let me
just fly away. So there you go. With
just one additional prompt, it's able to
make this game design look even better
with more actions like thrusters and
other different settings. Really cool.
So, I mean, if you want to vibe code a
video game, then definitely Fable is one
of the most performant models you can
use. It has the least amount of errors
and it's just so good at understanding
my instructions. I'm sure with a few
additional prompts, you could
potentially make this into a
productionready game. Now, one of the
most powerful features of using Fable or
pretty much any other state-of-the-art
model is getting it to create
educational content for your kids or for
yourself. For example, let's get it to
make a fun educational course on high
school chemistry consisting of free
lessons. Include interactive visual
exercises, make it look fun and
appealing. So, let's press run. Right.
So, in just one prompt, here's what we
got. Let me open this up in a new tab.
So, here is the landing page. It has
three lessons. Everything is responsive
and nicely designed. Let's start with
lesson one. So, here's what it looks
like. It teaches us about protons,
neutrons, electrons. Plus, here is an
interactive exercise to build an atom.
So, we can like increase the number of
stuff here. And you know, everything is
actually very accurately coded. So, for
example, the protons and neutrons should
be at the center here. Plus indeed there
should be two electrons in the first
orbit and then I believe eight electrons
in the second orbit. So let's actually
increase this further. And indeed once
we get to electron 11 then it adds
another orbit outside of it. So
everything is actually scientifically
correct. Next, it also gives us a mini
quiz. So let's like just randomly select
some answers. And as you can see
everything is programmed and animated
very well. It gives us the final score
at the bottom. Next, let's hop on to
lesson two here, it even gives us an
interactive element explorer with a part
of the periodic table. And if we click
on each one, then indeed it does give us
very accurate information about each
element that I select. This is very
comprehensive. And then next, let's
start this game. I'm a noble gas hiding
in period 2. People bend me into glowing
signs. Who am I? Let's just randomly
select something and see if it is
correct. It's not nitrogen. Let's
actually try to solve this noble gas.
So, I think it's one of these. Maybe
it's neon. And indeed, it's neon. And
then it also gives us a mini quiz at the
end. And then next, let's move on to the
last lesson. Here is where we can
balance different molecules to create
compounds. I guess everything just
works. And then here's a quick game. So,
let me just randomly select an answer.
And then here is where we learn about
pH. It's designed and animated very
nicely. And then finally, here is a mini
quiz. So, everything just works. You can
get Fablett to create lessons for any
topic you or your kid wants to learn.
This is incredibly useful for creating
educational content. All right, now
moving back to Claw Code. Next, let's
see if it can classify different types
of cancer tumors. So, I'm going to input
this image of six different slides. Each
slide corresponds to a different type of
tumor. I'm going to plug it in here and
then write identify the types of tumors
in each of these six images, if any.
However, note that it immediately
switched to Opus 4.8 instead of Fable.
All right, I tried this several times
and it still just switched to Opus 4.8.
And if I click on Y here, it says that
Fable actually has safety measures that
flag messages on most cyber security or
biology topics, which is kind of lame.
So, if I have any like medical condition
or something that I want it to diagnose,
I most likely will not be able to use
Fable. Instead, it's just going to
switch to Opus 4.8. I also tried to get
it to do some deep scientific research.
So here the prompt is describe the
molecular drivers of this type of
leukemia. Discuss targeted therapy
evolution from etc etc include relevant
tables and visualizations. And again at
the bottom here it says switched to opus
4.8 again because this is a biology
topic. So, unfortunately, as you can see
here, if you want to get Fable 5 to
answer some biology related questions,
it's most likely not going to answer you
and then just switch to Opus. All right,
so that sums up my quick tests on Fable
5. It's definitely state-of-the-art and
noticeably better than GPT 5.5 in terms
of vibe coding different things. It
tends to verify its edits automatically,
so it's a lot less errorprone and it
just works most of the time. However,
that being said, it also wastes a ton of
tokens. We're going to go over the costs
in a second, and for most instances,
it's kind of overkill. You don't really
need to use Fable 5 to do most stuff.
Next, let's look at its specs,
performance, and benchmarks. So, across
pretty much all categories, you can see
that Fable 5 not only beats the previous
Claude Mythos preview, which they said
was too dangerous to release, but also
Claude's previous leading model, Opus
4.8, 8 as well as 5.5 and Gemini 3.1
Pro. And I mean for most of these
benchmarks, it's not just a marginal
improvement, but a huge leap. For
example, for SWEBench, this is like an
over 20% leap over GPT5. Same with this
Aentic coding benchmark as well as
computer use, Agentic Coding, and Cyber
Security. This is a huge leap compared
to GPT 5.5. Now, if you look at this
benchmark from artificial analysis,
Fable 5 is ranked number one. It has a
million token context window similar to
the other top models like GPT 5.5 or
Gemini 3.1 Pro. So this 1 million token
window has become the standard now. This
is basically how much information you
can fit into your prompt at once and a
million tokens is roughly like 700,000
words or a small and medium-sized
codebase. If you check out their
intelligence index, then you can see
that Fable 5 is indeed ranked number one
at 65. And then GPT 5.5 extra high is
only 60. And then Gemini 3.1 Pro is even
lower. Now, if you look at the speed of
this, Fable 5 is extremely slow. I often
have to wait minutes for it to even
complete just a simple prompt. And here
is the most important thing, which is
the price. Fable 5 is by far the most
expensive model out there. If we break
this down further, then it's 50 bucks
per million output tokens, which is
crazy. Notice that GPT 5.5 is only 30
bucks and then the Claude Opus models is
only 25 bucks per output tokens. This is
two times more expensive than Opus. And
if you look at the hallucination rate of
Fable, it's actually very high. This is
even worse than Llama 3.1, which is
pretty shocking. So, it has higher
hallucination than Opus 4.8 as well as
even some open models like Miniax, Mimo,
and GLM. Here's another benchmark called
Livebench by Abacus AI. And
interestingly here, Fable 5 is not
ranked number one, but actually number
four, even below GPT 5.5, 5.4, and
Gemini 3.1 Pro, which is kind of
shocking here. It doesn't seem to do
very well in coding and instruction
following. Here's another benchmark
called Deep Suite, which seems to be a
more accurate measure of how good these
models are in terms of handling coding
tasks. However, at the time of this
recording, they have not added Fable 5
yet. Here's another benchmark called
vending bench, which measures a model's
performance on running a business over a
long time. So, it's tasked with running
a simulated vending machine business
over a year. And the score is basically
their bank account balance at the end.
Did they make money and how much did
they make? Now, interestingly, Fable 5
High was not ranked number one. It's
like ranked all the way down at number
nine. It did not make the most money
whereas the older Opus models plus GPT
5.5 and Kim 2.6 six made significantly
more money. So maybe Fable 5 had too
much guard rails, it's too nerfed, or
it's just too honest. And therefore, in
terms of running a business, it's not as
profitable as the other models. Now, in
terms of debating, Claude Fable 5 High
is the current leader, much better than
GPT 5.5, Gemini, or the other open
models. Here's another leaderboard
called SimpleBench, which tests an AI
model on common sense stuff. And as you
can see, Fable is indeed ranked number
one, but not by a lot. Just like less
than two percentage points above Gemini
3.1 Pro. Apparently still below the
human baseline of 83.7.
Next, let's also discuss the specs of
this because there are a ton of really
fascinating insights from this release
page. So, first of all, they describe
Fable 5 as their most capable model
that's generally available with
state-of-the-art performance across
nearly all tested benchmarks, especially
for long and complex tasks. So, it's
really good for like aentic coding and
software engineering tasks. For example,
Stripe was able to get Fable 5 to
perform a codebasewide migration in a 50
million line codebase in just one day.
This task would have taken a team over 2
months to complete manually. It's also
highly capable of serious knowledge work
tasks, especially for finance, document
analysis, interpreting charts and
tables, and factual lookup. Its vision
capabilities are also super strong. So,
it's able to easily extract stuff from
scientific figures, or rebuild an app
just from screenshots, or here, it even
beats Pokémon Fire Red with a minimal
vision only harness. And as you can see
from my frog test, this is indeed very
performant. This is the only model that
could successfully find the frog in this
image. This is also much better at long
horizon tasks. This can stay focused
across millions of tokens as long as
well, you don't run out of credits. It's
also really good at biology. It was able
to come up with these really strong
candidates for drug design. However,
that is using Fable's raw model. The
Fable that you and I have access to is
severely nerfed and it's not going to
respond to any biology related
questions. Now, it's time to talk about
the most important part of this, which
is the nerfing and the guardrails of
this model. So, Fable 5 is basically the
same underlying model as Mythos 5, but
with extra safeguards or basically more
nerfing. As you can see from my medical
example, Fable 5 will refuse to answer
this and it'll simply fall back to opus
or a dumber model, especially if you ask
it about something related to cyber
security, biology, chemistry, or
distillation. So, if you need an AI to
help with these topics, don't get Fable
5. It's simply not going to answer you.
That being said, for regular vibe coding
stuff, as I showed you in some of my
examples, this is indeed incredibly
performant and noticeably better and
less errorprone than the other top
models. All right, so that sums up my
review of Claude Fable 5. If you ask me,
would I personally use this? Honestly,
the answer is not really because it's so
damn expensive and slow. It just burns
away my credits and tokens and for most
of the time, I don't really need Fable
5. I think the only time I would
potentially use this is if there's some
really annoying bug that's hidden in my
codebase, which none of the other top
models could successfully find. Then
maybe I would use Fable 5 as the last
resort. But honestly, for most of my
other daily tasks, GPT 5.5 in codeex
just handles it beautifully already. But
let me know in the comments what you
think. As always, I will be on the
lookout for the top AI news and tools to
share with you. So, if you enjoyed this
video, remember to like, share,
subscribe, and stay tuned for more
content. Also, there's just so much
happening in the world of AI every week,
I can't possibly cover everything on my
YouTube channel. So, to really stay up
to date with all that's going on in AI,
be sure to subscribe to my free weekly
newsletter. The link to that will be in
the description below. Thanks for
watching, and I'll see you in the next
one.