Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
Fable 5.1's caching reduces input token cost by 75% compared to Fable 5, but in a practical test generating animated websites, the cost and design quality varied unpredictably across models and effort levels, with no clear winner.
Key points
Fable 5.1 High produced the best design (8/10) at $8.78.
Fable 5 High was the cheapest at $5.78.
Opus 5 Low was the worst design (2/10) at $6.58.
Fable 5.1 Max cost $10.67 and took 27 minutes, but design was 7/10.
Caching for Fable 5.1 is $0.25 per million tokens vs $1 for Fable 5.
Tools mentioned
Techniques
- Prompt caching
- Model Context Protocol (MCP)
- Effort levels (low, high, max)
- Blind testing
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
Enthropic just dropped Claude Fable 5.1 and everybody is losing their minds about it. So instead of just hyping it up, I decided to test it with all the other models and compare the outputs and
more importantly the cost. For this specific use case, I tested it by using the exact same prompt for all of the models in order to generate an animated website. So by the end of this video,
you can see the capabilities of each model and a rough estimate of the cost and most of all whether or not it is worth it to use Fable 5.1 for your everyday tasks. So without further ado,
let's dive right in. All right, so here is the scorecard that we're going to go by. So by the end of this video, I'm going to fill out this entire chart here. So we can have a sense of which is
the fastest model when it comes to generating this, which one looks the best, which is kind of up to my personal opinion, how long it takes, and then the most important part probably the cost.
How much do each of these different models cost at each of these different effort levels. So, to give you more of a breakdown, not only am I going to test each of the models, Fable 5.1, Fable 5,
Opus 5, and Codeex Soul, but we're also going to be testing the effort level as well. We're going to be testing three effort levels per model. So, we're going to test the low effort level, the high,
and the max in order to see how expensive it is, how long it takes, and how good it looks. Now, before we dive into this, we need to quickly talk about why people are saying that this is 75%
cheaper than other models like Claude Fable 5. And that has to do with something called caching. If you don't know what caching is, let me explain it very simply right now with an analogy so
that way anybody can probably understand this. So think of caching like a bookmark that we are able to give an AI. So this all has to do with the amount of input tokens that we are consuming when
using Fable 5.1 compared to the other previous models. So originally input tokens are $10 per million output tokens with both Fable 5.1 and Fable 5. Well, when we use caching, it is 25 cents per
million tokens. So, basically, when Claude remembers something that you previously talked about earlier on in a specific conversation or a thread inside of Claude Code, for example, it's not
going to charge you $10 per million input tokens, it's going to then charge you the 25 rather. Now, the key point that we need to understand right now and why people are saying this is 75%
cheaper than Fable 5 is because Fable 5 when it comes to cashing was $1 per million tokens. Now, Fable 5.1 is 25. So, this is essentially a 75% discount from the previous model. Now, this all
sounds great in theory, but the question is, does this really save us money when we go and build a website, for example? And that's exactly what we're going to test in this video. Now, first things
first, I'm inside of the Claude desktop app. And as you can see on the very bottom of my screen, we now have Fable 5.1 selected. We could also select between Fable 5 in the more models
section right here. So, we're going to be testing both Fable 5.1, Fable 5, and Opus 5 here in this video. But before we do this, we need to actually generate the visual assets that we're going to be
using for our websites. So, I'm going to be using a MCP server, and I'm going to be using a specific connector called Higsfield in order to do this. Now, if you don't know what Higsfield is, it's
basically an all-in-one Creative AI studio that allows you to use all of the top AI image and video models all inside of one platform. And the best part is is when we plug it directly into Claude
with a connector, basically we don't even have to go inside of Hicksfield to generate these image and video assets, we could just have Claude do it directly for us and add this to our website
without us needing to go into this platform. Now, I'm going to show you real quickly how to set this up. It's going to take maybe 30 seconds. So, first of all, there's going to be a link
in the description to access the Higsfield MCP. So, make sure to click on that link. And from there, you're going to then see this right here. Once you have an account, all you're going to do
is come to the Higsfield MCP section and we're going to copy this right here where it says MCP URL. Then we're going to come back to Claude. Simply come over to the lefth hand side where it says
customize. Then we're going to go down to connectors. This is basically how we're able to connect different applications to Claude without us needing to leave the Claude interface.
What I'm going to do is I'm going to come over and click add and I'm going to go to the bottom where it says add custom connector right here. Personally, I'm just going to call this Higsfield.
And then I'm going to paste in that remote MCP server that I just copied right here. Next, you're going to see this checklist right here. I'm just going to move forward with this and
click on add. And just like that, you can now see that Higfield is connected into Claw. Now, for the fun part. This is when we're going to start generating these videos that we're going to use in
this animated website. So, what I'm going to do is I'm just going to paste in this very simple prompt right here. It's essentially saying, use the Higsfield MCP to generate five photoreal
videos of the Falcon 9 from SpaceX. It breaks down the aspect ratio, the resolution, and how long that these videos should be. And it kind of breaks down these three scenes that we want to
generate right here. So, this is a pretty basic prompt. I could get more specific with it if I wanted to, but Higsfield is going to actually optimize this prompt in order to generate these
videos. Now, what I'm going to do is at the bottom right here, you could see I have Fable 5.1 enabled. I'm just going to toggle this on to Opus 5 because for generating these video and image assets
with Higsfield, using Fable 5.1 isn't necessarily going to do anything. and it's probably just going to waste some of my token usage here. So, I'm just going to go ahead click on Opus 5 since
it is cheaper and I'm going to send this off to Claude. One thing I do want to point out is you could see right here that it's using the seed dance 2.5 model which is probably the best video model
at the time right now. And one of my favorite parts about using Higsfield is I don't need to go and think which model should I be using if I'm generating images and videos. It automatically
knows which model to route it to for any given task. All right, so that took probably like 10 or 12 minutes and now we have these five different video clips that were generated. So, let's take a
look at these. They're pretty simple, obviously. It's just the SpaceX Falcon 9 taking off in this clip. That looks pretty cool with the clouds. This right here is the Falcon 9 landing in the
middle of the ocean. I mean, that I think that'll make for a cool visual on the website. Then we have this one of the Falcon 9 kind of just floating in space, which looks pretty cool. As well
as we could always come inside of Higsfield, click on assets, and then we could always view the videos that were generated directly right here. So, this is another one that was generated of the
Falcon 9 kind of like going into orbit. All right. So, now that we generated these video assets here, I'm going to give this exact same prompt to all the different models that we're going to
test this with. So, here is the prompt. It's very, very simple. It simply says, generate a stunning animated landing page for SpaceX. I attach the five clips that I want you to include in our
animated website. And then I did specify not to use the animated website skill because if I were to use this skill that I've created in Claude, it's going to make them all look the same and it
really won't matter what model we're using. So this first one, I'm going to use Fable 5.1. I'm going to put this on low effort. I'm going to send that off. All right. So I have all of these
running inside of Claude. And I'm not going to lie. I have no clue how much this is going to cost. This is pretty loaded. We'll see if by the end of this video I'll even have any usage limits
left. But if you guys are enjoying this video and you're enjoying watching me burn all these tokens, make sure to subscribe to the channel for more content like this. Now, I need to do the
same thing for chatbt 5.6 soul. So, let's go ahead and do that right now. All right, so I let Claude and Chachabt do its thing for about 30 or 40 minutes. As you can see, it's actually dark out
now. When I was recording earlier, it was still light out, so this took quite a while to go and build all of these different websites and designs. So, now we're going to go through this
scorecard. So, what I'm going to do is I'm going to go through each of these websites and I'm going to look at the design, basically how it flows with these animations that we created, and
I'm going to score it on a couple of different things. I'm going to score on the look. I'm going to write down exactly how long it took for each of these to generate. I'm going to break
down the cost, how much it costed to run this entire task with this simple one prompt that we used for all of them. So that way you can understand not only the design outputs of these models but how
much time it takes and how much it actually costs which at the end of the day is probably what you are worried about. Now in order to make this more exciting for you guys I'm going to do a
blind testing. So I'm not going to explain what model generated these websites. I'm going to leave it up to you to guess yourself and then I will share once I give it a rating. All
right. So this is the first one that we're looking at. As you can see the asset that we generated with Higsfield is embedded. This looks pretty cool. There's a rocket taking off. We have
certain text going through this and I'm scrolling at a moderate speed here. It looks like that, you know, Starship is landing right here. I think this is solid, but I'm not going to lie, some of
the graphics aren't flowing basically how we'd like them to. Oh, I think this is pretty cool. Kind of uh, you know, showing this little graphic. I didn't instruct for this to generate it like
that, but it did. So, I think this one overall is pretty good. Again, I think the animations don't flow as good as they should. Like, for example, this one shoots up. I think that looks pretty
cool. Now, before I give my ranking out of 10, I do want to preface and say this is obviously completely subjective. This is just off the first glance of how I think that the website flows. And if I
am stunned by the design because I've created countless amounts of websites like this, so I can kind of see when it's more of a basic vibecoded website design and a couple of these outputs
blew me away with the creativity. So, I'm just going to simply make my judgments on what I find interesting. Now, if I had to rate this one, I'd probably give this maybe a solid 6 out
of 10. Uh, I kind of like this right here. It's pretty creative. So, I'm going to give that one a six out of 10. And the model that generated that output was Opus 5 max. So, I'm just going to
come here and just say 6 out of 10. Now, let me go ahead and check the usage. Now, if you are inside of Claude and you want to do the exact same thing. All you have to do is forward slashcost or usage
and it will automatically just pull up this table here. It'll show you your 5 hour limit, your weekly limit on all models. It'll show your weekly fable limit and will show the cost for this
session based on all the models you use. So for this session, the cost was $10.70, which isn't terrible. And the API was being used for 14 minutes. What we're going to focus on is active right
here. So 17 minutes and 57 seconds. All right, so we have the first one logged. We have Opus 5 Max. And I'm going to go through these pretty quickly. We don't need to go into all the details. Just at
first glance, I'm going to rate these, dive into the time and the cost so we could go through this quickly. All right, so here is the next one. And we have our Falcon 9. Again, these are
going to be the exact same animation shots that we generated with Higsfield. I mean, this one's pretty solid. Nothing crazy. It's got like all this texture, which the other one didn't necessarily
have. I mean, I think this one's good. I don't think it's anything crazy. I'd say this one's probably not as exciting as the last one. So, I'm going to give this one
a five. And the model that generated this one was Opus 5 High. So, I'm just going to rank this five out of 10. And let's go into the other statistics here. So, for this one, we have a cost of $11,
which is pretty similar to the previous one that we just showed. This one was active for 22 minutes and 6 seconds. Moving on to the next one. Let's take a look at this one. So, we got Falcon 9.
Obviously, we have a little final countdown. I think that's a nice cool addition. I think that's pretty creative. I scroll down, it shows liftoff. One thing that I think is cool
about this one is it shows this little like text overlay here where it shows a number. So, it shows number one, and as I scroll, it shows number two behind the main text, number three for the landing,
and then built again to fly. I think this one's pretty solid. This text looks solid. And then down here, this text I can't really see for Falcon 9. If I had to rate this one, I'd probably give this
a solid like six out of 10. And that was Fable 5 Max. So, this is getting to the better models. Obviously, the two previous ones were Opus 5, which were the cheapest models, and this is on the
max effort level as well. So, we're starting to scale up in terms of like the amount of money we're going to be spending with the models. And for me personally, so far, I'm not seeing a
massive difference in terms of the quality. I think they are all pretty solid. So, I believe I said six out of 10. And now, let's go take a look at the price and the time it took. All right,
so now is when we're going to start seeing the discrepancy between the cheaper and the more expensive models since this was Fable 5, which is obviously more expensive than Opus, and
it was the max effort level. So, this one for this session was $24 as opposed to our previous two that were like around the $10 mark. And then this was active for a total of 26 minutes and 41
seconds. On to the next one. And this one is a bit different than the previous ones we generated. So it looks like each of the videos are kind of static. Like of course they're moving, but it doesn't
move as I scroll through the website, if that makes any sense, which I think could be pretty pleasing to the eye. It's not as creative in my opinion, which is kind of what I'm judging the
outputs on. But I think that this is the most like real life website I think that you would see cuz you know it's very easy to read what is on the screen here along with like a cool little video
behind it. So I think this one is actually pretty solid. I'm going to give this one I'm going to give this one a 7 out of 10. And the model that generated that one was Codex 5.6 Terra Light,
which is actually pretty surprising since this is one of the cheaper models from Codeex. This is the first model that we've tested that is not a clawed model. But I guess this actually makes
sense since the design style was a little bit different than the claw models since it was a little bit more static. So I guess that's kind of interesting. So 7 out of 10 on the look
or the design. As far as the time, this one took about 10 minutes and 53 seconds, which is on par with some of the lower Claude models. Now, unfortunately for the cost, it wasn't as
simple as Claude was, where I could just type slash usage and it would bring up that card in order to show me how much this costed for this specific session. So, unfortunately, I'm not going to be
able to track this for both of the Chatravt models. But hopefully, you can get a sense of, you know, the design that it can create and how long it will take compared to the Claw models. Here
is the next one. I kind of like how this has pretty big text here. We got Falcon 9. As I scroll, it's launching. We have liftoff. Here it is landing. I think this one's actually pretty solid. I'm
not going to lie. I think it's clean. I like the big text and typography. And then I honestly really like this section as well as well as this. This is probably, I'm not going to lie, my
favorite one that I've viewed so far. I'm going to give this one a solid 8 out of 10. And I'm curious if you had to guess, what do you think that this one is going to be? I'll give you a couple
seconds to think about it. And this was Fable 5.1 high, which I guess makes sense since it was my favorite design. All right, I think I gave this one actually an 8 out of 10, which is the
highest ranking so far. Now, let's look at time and cost. Now, this one is very surprising to me. The cost for this was actually only $8.78, which is the cheapest output that we've
gotten, as well as it was active for 18 minutes, which isn't anything crazy. We've had some that were up to like 26 minutes so far, I believe. So, this is pretty surprising. And if we're learning
anything here, at least for me, it's that sometimes a model doesn't necessarily matter that we are using. Now, let's see if we can find a new favorite. This one. Let's go through
this one. Okay, so so far I think this one's kind of basic. I'm not a fan of this text here. Um, scrolling again. Nothing too fancy. Probably getting tired of watching the
same visuals over and over again. As am I. I think this one is okay. I'm not going to lie. I'll probably say this is my least favorite one so far. I'm going to give this one a four out of 10. All
right, next on the list, and I'm going to try to speed this up because I know this is taking quite a while, but hopefully you guys are enjoying this. So, this one is pretty simple, and if
you remember previously, this might look similar to one that we generated earlier on. I think if you like simple websites, this might be like really enticing to you. You'll probably really like this
one. I think it's clean. But if I had to judge this in terms of the creativity, I'd probably have to give this a six cuz it's nothing, you know, super creative. But I do like the way this one looks.
And if you remember the previous ChachevT model, you might have guessed that this was also generated by ChachevT, and that is correct. This was the Codex 5.6 Soul High. Again, there's
no way for me to check how much this costed, unfortunately. But this did work for 13 minutes and 7 seconds, which is actually pretty surprisingly low. Moving on, and this one is pretty unique. So,
first of all, as I scroll, nothing crazy. And then boom, we have this like explosion as there's liftoff on the rocket. I think that's pretty cool. And then as this lands as well, it also
does this like shake. So, I think this one's cool. I think this one's pretty creative. As long as I like the big text down here. This one might be near the top for me. I'm not going to lie, just
cuz it's creative with, you know, the kind of the the shaking of the frames here. I haven't seen any of the other ones do something similar to this, so I might have to give this one a solid
seven. All right. So, the website that you just saw generated is Fable 5.1 Max. And I'm going to give this a seven out of 10. Let's calculate the time and cost now. All right, so let's now break down
the cost of Fable 5.1 Max to run this generation here. So costed about $10.67, which honestly is incredibly surprising since the max effort level on Fable 5.1 is supposed to be pretty high based on
everything that I've seen online so far since it's released. And in terms of how long this took, it took 26.9 minutes. So basically 27 minutes. That is pretty long. I think this is like the longest
generation out of all these models. Now, we only have three more websites to test. So, let's move on to the next one. And I'm going to make this quick. So, here is this one here. This one is
honestly pretty terrible. There's no video that is playing for these animations. The text kind it just looks pretty bland. Yeah, this one is not good at all. I'll give this one a 2 out of
10. Now, this one makes sense, I guess, because this was Opus 5 on the loweffort level. So, this was the worst model and the cheapest model. And me personally, I could tell the difference between the
other outputs. And I'd assume you probably think that one was pretty bad, too. So, that was a 2 out of 10 in my book. And in terms of pricing, this was only $6.58, which I think is probably
the cheapest, which makes complete sense. And then this ran for 10 minutes and 16 seconds. All right, moving on to the next one. I think we only have two left after this. So, we're getting close
to the end. This one's cool. I think the typography is cool. Obviously, we're using the same graphics here in the background. This one is actually moving. I think this one's solid, but again,
it's nothing crazy. If I had to give this a rating out of 10, I'd give this one maybe a six out of 10. This was Fable 5 High, giving it a six out of 10. And this one was $5.78 to run, which is
the cheapest so far. And this ran for only 2 minutes and 11 seconds, which is pretty surprising. This is kind of interesting. All right, so one more after this. Let's take a look at this
one. Falcon 9. This one's good. It's got some pretty pretty big typography. I think it takes up a little bit too much on the screen here. I'm not a huge fan of this text right here. um seems like a
bland kind of space in our website here compared to some of the other ones. But again, it's nothing bad. It's nothing great. If I had to give this one a rating, I'd give this one maybe a five.
And finally, this is the last website. This is Claude Fable 5 Low. And I think this one's pretty solid. It's pretty basic. It kind of reminds me of the Codeex ones that we created, but I think
this one's solid. I like the orange accents. I like this cool little thing, cool bar going across my screen. I think this one's pretty creative. I'd honestly give this one probably like a 7 out of
10. Now, for Fable 5 Low, in terms of the time and the cost, this was actually pretty good. So, it only took about five minutes and 4 seconds, and it was $9.70, or I guess it was actually $9.76
in order to generate it. Now, here is the final scorecard. Now, for going off the model that was ranked the best in looks in my humble opinion, of course, this is completely subjective. The one
that wins is Fable 5.1 High, which ranked an 8 out of 10. Now, here is the final ranking scoreboard here broken down in two categories. First, let's go through the look and then we're going to
go through the cost right here. So, in terms of the model that got the highest look, that is Fable 5.1 High, which got 8 out of 10. Next, below that is Fable 5 low, which got a 7 out of 10, which is
pretty interesting because there's a big discrepancy there. We have Fable 5, the previous model on the low effort level, compared to Fable 5.1, which is this new hyped up model that is on the high
effort level, which makes sense. But the fact that these are very close and in terms of the design output, like honestly, I probably wouldn't be able to tell the difference between the two is
pretty interesting. Next, the model that you would think would be at the very top, Fable 5.1 Max. I have this tied in second place, 7 out of 10. Next, Codex Soul Light, which is 7 out of 10, which
is pretty interesting. Completely different model from OpenI. And then I'm going to go down to the bottom. We have Opus 5 Low, which honestly makes complete sense. It's the worst model.
It's cheapest model. And I gave that a 2 out of 10. Next up, above that, I have Fable 5.1 Low, which is a 5 out of 10. This is pretty interesting. If anything, this tells us that the loweffort level
models obviously aren't very good when it comes to design and specific complex tasks like this. And that goes for even 5.1, which when I started this, I honestly thought this would be ranked
closer to the highest. Now, in terms of cost, let's break this down cuz this is pretty interesting. Fable 5 High was the cheapest. It was $5.78. Next we have Opus 5 low, which obviously makes
complete sense. It should be the cheapest. Then Fable 5.1 high got $878, which seems pretty interesting. You would think that this would be quite a bit higher. Fable 5.1 Max. This is
probably the most like random cost out of all these. It is $1067 ranking in spot number five, which you would think this by far would probably be the most expensive model, but
according to this test, it simply was not. Next, the most expensive model was Fable 5 Max, which I guess makes sense since it is one of the higher effort level models. And if I walk away from
this test learning anything, it is that these models fluctuate in the price as well as the quality of their outputs. Naturally, when Fable 5.1 released, I was thinking of making a video breaking
down how to make insane websites with this new model. But I realized that the biggest thing that people using these models are starting to realize is that it's not necessarily the outputs that
we're getting from these models that is very interesting. It's how much does it actually cost to get a decent output. I really do feel like if you are a average consumer of AI, using a model like Fable
5.1 Max at the highest effort level is probably overboard for whatever it is you're trying to do. Like for example, in this video I'm generating animated websites. That would usually be
something you'd use a top model for. But when it comes to the designs, there's really not a huge discrepancy between the two. As well as for the pricing, it could be kind of hit or miss depending
on what it is you generate. So anyways, I know this video is probably all over the place. I wanted to do this so you could just get a sense of the design outputs from these different models as
well as the cost breakdown. Since I know a lot of you guys are asking which model you should be using and how to optimize your token usage. If you guys enjoyed this video, make sure to leave a like,
subscribe to the channel, and honestly just let me know your thoughts about this in the comments below. I'm interested to hear if you're feeling like model fatigue. I'm getting really
tired of new models dropping and feeling like I'm missing out by not using them. But genuinely, I feel like there's not a massive difference as of right now. Thanks for staying the end and I'll see
you in the next video. Peace.