Claude Fable 5 Leaves GPT-5.5 and Chinese Rivals in the Dust on a Brutal Frontend Test

Hahahahaha!

Claude's strongest model has arrived, and I couldn't help running a comparison test. It cracked me up.

What surprised me wasn't how strong Claude Fable 5 is, but how incredibly weak the other models are!

Of course, conversely, Fable 5 is just too strong. Actually, Opus 4.8 was already enough to beat everyone else.

I tested four platforms in total: Fable 5, and three others for reference, including both international and domestic ones!

I'll try to keep my "sarcasm skill" in check! No malice intended. I trust the big companies aren't that "petty."

First, let's look at what everyone cares about most: Fable 5.

Fable

The Chinese meaning of Fable is roughly "寓言" (allegory/fable). Let's first see how this "fable" performs.

Before starting, let me explain my testing environment.

Because the network has been quite unstable lately, models like Fable hang for a long time when thinking and get interrupted directly. I won't use CC for testing.

I'll just use the web version. Generally, using an agent is definitely stronger than the web version, so to be fair, everyone gets tested with the web version!

Then, our test prompt is this one:

I've previously tested 14 models with this, so the reference data is very rich.

The complete requirements for this test are as follows:

Please do not draw a picture directly. Instead, write the code for a single HTML file. When I open it in a browser, I want to see a dynamic, cyberpunk-style long scroll of "Along the River During the Qingming Festival." It needs to fully combine the ancient style of the Qingming scroll with cyberpunk elements.

Make the scroll content rich in detail.
Requirements:

The image must automatically scroll slowly from right to left.
It must contain at least 50 dynamic elements: such as flashing neon signs, flying cars, holographic advertisements, and mechanical prosthetic pedestrians on the street.
When the mouse hovers over any shop, a cyberpunk-style info card should pop up (e.g., "Old Wang's Prosthetic Repair Shop - 98% Positive Rating").

Key Technical Points:

SVG/Canvas drawing programming ability
CSS animation logic
Mouse interaction event handling
Aesthetic design and visual presentation

I gave them all the prompt requirements and key technical points. It's an open-book test to see the upper limits of their capabilities!

Don't underestimate this prompt; it's not simple, as you'll see later.

I'll just go straight to the results:

This is the result Claude Fable 5 produced directly in the web version, written purely in code without referencing any third-party resources! Remember this image; it will be useful later.

Then I'll record the testing process. If you don't care about the process, you can skip to the next chapter!

Here, I must first swallow back the words "great value" I said this morning! Although, based on data comparison, it's only twice as expensive as Opus 4.8, the actual consumption is terrifying.

Look, I started the test with a full quota:

Because Fable launched, Claude officially reset everyone's quota.

Then, I happily started testing:

The whole process was somewhat lengthy, taking about 23 minutes!!!

It even compressed the conversation once midway. You can see its thinking context is very, very long. Of course, this might also be related to the web version.

After it finished running, I went to check the quota, and my world collapsed:

One webpage directly consumed 63%.

When I wanted to test further 🤣... I immediately saw the following scene:

I have no words for this situation.

I checked via the terminal, and the quota was pushed to 112%.

Claude's web version is now extremely powerful, definitely running a lightweight agent. The entire thinking process outputs a massive amount of content.

Expanding the whole process is very long; I counted roughly 38 nodes!

Even the web version is this terrifying. From now on, even in the chat function on the web, I won't dare casually use the advanced model; the consumption is too horrifying.

Of course, the result was also very good:

Just the opening "Along the River During the Qingming Festival," Cyberpunk Bianjing, plus the red character "汴" (Bian), already beats everyone else.

It even drew the Rainbow Bridge:

On the bridge are flying vehicles and delivery drones; under the bridge are boats and reflections; behind are skyscrapers, a hint of distant mountains, and stars dotting the sky.

Various shops:

Then it also created various shops for you, like "Old Wang's Prosthetic Repair Shop, 98% positive rating, 40-year-old shop, open past midnight, trade-in accepted." The slogan is "Left arm replacement, while you wait; military-grade prosthetic eyes, fair prices for all."

There's also "Wang's Paper Offerings - 'Digital Reincarnation, One-Click Salvation'" and "High-Price Memory Buyback"!

Anyway, the shops, billboards, and flying vehicles from the requirements are all there. The flying vehicles come in two types, both very delicately made. One type looks like it can carry people, and the other is a delivery drone.

Notice their lines are all very smooth, and the shapes are very vivid. I even saw one pulling a donkey—no, a donkey pulling goods!

Overall, it can't be called perfect, but it's very good. It did everything it was asked to do. Of course, whether it's done well depends partly on itself and partly on the competition!

Arch-Nemesis

As everyone knows, Anthropic and OpenAI are the world's two strongest AI companies. They should IPO in the second half of the year, with estimated market caps starting in the trillions. Recently, Anthropic has overtaken OpenAI. They say colleagues are rivals; these two are definitely arch-nemeses. Logically, Altman should have unleashed a big move to counter this by now.

Based on this situation, we must also test GPT-5.5!

Below is the test situation on ChatGPT's official website:

Although GPT-5.5 had thinking mode turned on at medium intensity, it was clearly very perfunctory on the web. It finished this webpage in just 17 seconds.

The effect looks like this:

This is really a bit abstract. Is that oval above supposed to be a flying vehicle? And is that black blob a person? Is that small thing a sign or what? The colors are quite red and green, very cyberpunk, but this thing is just too abstract!

This result was truly unbearable to look at, so I made an exception and ran it again using Codex.

It took about ten minutes or so, and finally finished:

The elements seem much richer than before, but the overall lines and design sense are still very rough. People are in the water, and the bridge is on land!

This is probably just the model's inherent capability; no matter how you prompt it, it's useless.

If you've used GPT-5.5's frontend capabilities, you should know what I mean!

My feeling is that Fable 5 is somewhat like a master's work, while GPT-5.5 is somewhat like a cartoon.

We'll stop looking at GPT-5.5 here.

Let's find a few more test subjects below.

What everyone loves discussing most is whether the gap between domestic models and top international models is shrinking or widening. No one can say for sure.

The capabilities of domestic models basically exist in "benchmark" data; they all look quite fierce, and their marketing is quite fierce too.

Usually, they are all benchmarking, approaching, surpassing. Today, let's pull them out... for a head-to-head PK.

I'll pick the two with the most aggressive marketing recently, both abbreviated as MM. To avoid triggering keywords, I'll just call them "Wolong" and "Fengchu"!

Wolong

Wolong's most powerful feature is its Credits, which can be given out in the tens of billions at once. Recently, it's been heavily pushing an extreme-speed model at 1000 tokens/s, and I've already seen all sorts of awesome posts about it.

Anyway, the numbers are one more awesome than the next. But as for actual capability:

Hahahaha!

Now you know why I laughed so loudly at the beginning!

This is absolutely a real test, done on their official website AIStudio:

It ran for about 334.7 seconds before starting to spit out code.

Then it kept spitting out code for quite a while. Time-wise, it shouldn't be considered long; I estimate less than ten minutes.

This time is actually okay, but the result completely failed to display content. This problem seems to have occurred several times in previous tests. For example, that "Master's Diary" also didn't display content.

At that time, there was a low-level JS error.

Then I looked at this example, and it was also a JS error:

Uncaught SyntaxError: Invalid shorthand property initializer

This error isn't much more advanced:

This is a JavaScript syntax error, meaning:

An illegal property initialization method was written in an object literal.

The most common reason is that you wrote `=` instead of `:` inside an object `{}`.

This matter, sigh! Even a junior programmer shouldn't make this kind of mistake.

Alright, I won't say more. Everyone can judge for themselves.

Fengchu

Fengchu is also a very formidable player. Back in the "previous-previous generation" model, it already proclaimed itself the strongest coding agent. Recently, it released the upgraded Model3, which is even more "incredible." You've probably seen many of the promotional articles, hyping it up to godlike levels.

Logically speaking, if the grandfather was already the strongest, then the grandson must be absolutely invincible.

So I went to experience their strongest Agent Team.

Then it just wouldn't send no matter what. Clicking the button just gave a shake, and nothing happened. Alright, maybe it's my problem!

No choice, had to switch to their older version Agent.

This time it performed well; the button was clickable!

Then, "Wolong" had already finished, but "Fengchu" still hadn't figured it out:

From morning until afternoon, after finishing lunch. Finally, there was a result:

The result was "API Error: Connection error."

Hahahaha, now you know why I laughed so loudly at the beginning!

This must be my network problem. Let's try again.

This time the posture was very impressive; a Todo list had already appeared on the right side. I felt it was a sure thing.

Ten minutes later, it told me "System restarted..."!

Uh, did it crash the system???

It looks incredibly fierce, with Agents and Agent Teams! But when it comes to doing work, it's either interrupted or restarted.

What can I say? You call it an agent team, but if a single agent can't get the job done, what's the use of a whole team of agents!

Besides looking good in marketing materials, I really can't think of any use!

I've been running this intermittently for several hours now, so I won't waste any more of my life.

I'll just paste a result it produced on CC before:

It's roughly at this level!

All elements are very primitive. Lanterns are hanging on the ground, people are flying in the air. As for the style, if you don't look at the name "Along the River During the Qingming Festival," it has absolutely nothing to do with the Qingming scroll.

Honestly, this test was quite surprising to me.

When I first saw Fable's result, I thought, "Eh, it's okay, just average." But as I slowly tested further, suddenly, my admiration for Fable 5 surged like a torrential river, endless and continuous.

And my test here isn't even Fable 5's strongest area. I specifically chose areas they haven't trained on to test. If you've read other articles or some posts from abroad, you can see it's incredibly powerful in game development.

Actually, it's okay if the results aren't good. Everyone just be a bit more sincere and honest!

Alright, that's enough. My quota is back, so I'll continue testing the rest!

All my screenshots, all examples, all content here are completely real, and all are single-run tests. My account has the complete conversation records!

A casual test, full of fun!

I've put all the examples here:

https://topai.jarvisuni.com/

It also includes the complete prompts!

Finally, I also had it make a version that can switch between distant, near, and mid-range views. The distant view should be impressionistic!

image-20260610161349928

Everyone can also try using my set of prompts. Currently, this example probably hasn't been optimized/trained on yet. Those who do well truly do well, and those who don't truly don't.

Alright, that's it for this example.

The conclusion is: Fable5 is truly far ahead!

If this article is useful for your model selection, or gives you a "good laugh."

Remember to hit like!

Liking costs nothing, but it's a huge encouragement to me.