Practical test of Claude Fable 5.2: From coding to modeling to making videos, it has also taken on creative work

As reasoning, programming, and mathematical abilities continue to break through, people have gradually become accustomed to AI setting new records on various lists. But this time, Claude Fable 5.2, released late at night by Anthropic, brought another kind of more intuitive shock. It can be seen that AI's aesthetics, creativity, and attention to detail are undergoing astonishing leaps.

A single prompt independently completed a full video that included data collection, visual design, original music, and editing. It spent over two hours researching hundreds of reference photos, using code to recreate the architectural space of the Apple Store in Jing'an, Shanghai… What’s even more noteworthy is that the model has begun to exhibit increasingly strong autonomous judgment capabilities. It actively researches materials, selects design schemes, overturns unreasonable technical paths, and even forms its own quality requirements for the final work.

This may indicate that large models have transitioned from following instructions to independent creation. When AI begins to possess technical capabilities, aesthetic judgment, and complete delivery capabilities simultaneously, the future division of creative labor between humans and machines will also usher in new changes. Below is an in-depth test of Claude Fable 5.2. From video production to 3D modeling, let’s see what surprises this release has brought. Here, enjoy:

Source | Digital Life Kazik

Author | Kazik, AIZ Xiao Zhu

At 3 AM, Claude officially released Claude Fable 5.2.

Meanwhile, OpenAI is selling a handheld device for $200 tonight.

Currently, Claude Fable 5.2 has officially launched across all channels, and it can already be seen in the Claude desktop version.

At the same time, Anthropic has also opened temporary access to Fable 5.2 for Pro members at $20, but only for this week.

Additionally, for Claude's $200 Ultra subscription users, they are offering a reset opportunity for Fable's quota, based on the subscription time, valid until March 8 of next year. To be fair, sometimes Anthropic does have a conscience…

I am very happy that after enjoying Claude Opus 5.5 for nearly a month and a half, I finally awaited the new Fable model, although it is not the rumored Claude Fable 5.5; Anthropic named it Claude Fable 5.2.

This is actually normal. It was first reported on September 9 that someone leaked that Anthropic was running a new model, and the temporary code name given was indeed Fable 5.2. However, later Opus 5.5 and Sonnet 5.5 were released one after another, and everyone naturally assumed that the next Fable would also be called 5.5, but in the end, it was 5.2. So, it is highly likely that Anthropic is still hiding a stronger Fable 5.5.

First, let’s look at the basic situation of Fable 5.2.

Claude Fable 5.2 still has 1M, which is 1 million tokens of context, with a maximum output of 128K tokens, and knowledge cutoff is set to June 2026.

This time, the price has actually decreased. Now everyone knows that tokens are too expensive and have exceeded the acceptable range for all humanity, so after the price increase trend, prices are now continuously decreasing.

Fable 5.1 was originally $10 for every million tokens input and $50 output.

Fable 5.2 has changed to $8 input and $40 output, with caching being similar, an overall price reduction of 20%.

The price comparison of several main models from Anthropic is roughly as follows.

Then, the speed has also improved somewhat; the official statement says it is up to 30% faster than Fable 5.1, which is not particularly significant, but it is better than nothing.

Moreover, the number of false positives has decreased. The safety barriers of Fable have always been criticized; when it was first released in June, many people complained that it treated biological assignments as security threats. This time, the official statement said that false interceptions were about 55% less than 5.1, with a 60% reduction in cybersecurity-related issues and a 40% reduction in biological issues.

Another point that I particularly like is that it has inherited the advantages of Opus 5.5, which is that it finally started to speak more human-like. The output style of Fable 5.1 was quite similar to that of Opus 5, which was filled with jargon that I particularly disliked and could not understand at all. Later, with Opus 5.5, Anthropic finally realized that the model is meant for human use, so communication and speaking human language are particularly important. They made significant adjustments, and the reputation instantly improved.

This time, Fable 5.2 also inherits the advantages of Opus 5.5, with a similar speaking style.

However, from this response, it can also be seen that the sense of superiority of the Claude series models has been further intensified due to the enhanced capabilities of Fable.

Recently, there have indeed been too many things done, some involving underlying infrastructure, which is a bit chaotic. So this time, it directly overturned my ideas and told me that there is a better and more suitable approach, which should directly integrate the things I modified into the general middle platform plugin, rather than still modifying the previous broken CI middle platform.

I previously posted a rather popular joke about the personality differences between Claude and GPT models.

"I: I have to download 20 Excel files from the backend every week and manually copy and paste them into a summary sheet. It's too troublesome; help me write a script to automate mouse operations.

GPT: Sure, boss! I've already written it for you. It automatically clicks, downloads, copies, and pastes, and even includes exception handling and failure retries. From now on, you just need to click run!

Claude: No, buddy, aren’t those 20 Excel files all exported from the same database? Why are you downloading them? I’ve directly written an SQL for you, connecting to your backend data source, so that the final report is automatically generated every week. Why are you downloading Excel files for no reason…"

For those who hope for stronger model capabilities and want to hear things they don't know, Claude might actually be more suitable for me. Some call it wishful programming; of course, it can also be called conversing with a deity.

In terms of scoring, there’s no need to think; the Fable series today is basically guaranteed to be SOTA.

It topped the AA index with a score of 65.

The official scores are also overwhelmingly superior, and mainly Anthropic's scores can be trusted.

Anthropic also conducted a very interesting experiment, assigning Fable 5.2 the task of designing a prototype operating system, starting from an empty folder, coordinating over 1,000 sub-intelligences, and continuously writing, testing, and reviewing code for eight days and nights. Then, it produced a runnable operating system that can boot on a laptop, with its own kernel, disk, keyboard and network drivers, permanent file storage, windows, login interface, installation program, and built-in Claude features…

Such a hardcore test is certainly beyond our capabilities.

With limited quotas, I still focused on the longest board of Opus 5.5, video effects, and front-end aesthetics.

(PS: To clarify, all the following cases are direct outputs; I did not engage in any secondary dialogue or post-processing. All were generated from a single prompt, which ran for 2 hours to give me the results, and I did not make any secondary adjustments.)

First, what I thought of was a case originally prepared for GPT-6.2 Sol, as it is almost ChatGPT's fourth anniversary, so I also wanted to make a video commemorating ChatGPT's four years of history. This time, I prepared to let Claude Fable 5.2 produce it directly, one to check the hallucination rate, two to check the aesthetics, three to check the accuracy of the code, and four because the NTR plot is just too good.

My prompt was very simple:

"Create a 40-second to 1-minute animated video to commemorate ChatGPT's fourth anniversary (November 30, 2022 - November 30, 2026), in 16:9 landscape format, with no narration and original music. The content should be verified online, covering the key versions, milestones, and numbers from these four years, without fabricating any information. Before starting, observe the visual style from OpenAI's official website, blog, and recent videos, and refine it as the visual tone of this piece. The pace should be fast, with dense animations and full visuals, synchronized with the music. The positioning of the piece is as a third-party tribute, so it should not look like an official OpenAI production."

Then, it went to work. After nearly an hour, Claude gave me a video like this…

To be honest, I momentarily didn't know what to say.

I did verify the information inside, piece by piece, and sadly found that I didn't identify any factual errors.

The information was sourced by Fable 5.2 itself, the graphics were created by it, the music was composed by it, the timing was set by it, the animations were made by it, and the editing was done by it.

Honestly, at the last scene, I really couldn't hold back.

Watching this video, I truly fell into deep contemplation…

Then, I had it create an explanatory video based on the article about the Super-agent from DeepSeek.

Perhaps because DeepSeek's main color falls into the blue commonly used in AI, this video still has some AI feel, but in terms of the finesse of the animations, I can only say that compared to Opus 5.5, there has indeed been significant improvement. Of course, Opus 5.5 could also produce such effects, but it required a lot of effort and multiple discussions. However, now, it’s direct output…

The moment the text for Super-Agent came out, I was truly amazed.

This rhythm, ah, is really already quite skilled.

Then I tried a 3D scene using Three.js.

The prompt was like this.

Then Fable 5.2 started working. After running for dozens of minutes, I saw it contemplating the image previews in the summary, and at that moment, I felt that the aesthetic of Claude's model is really strong…

This is a static frame rendered in the webpage purely with code; everyone can feel it.

At this point, I didn't know what else to say.

After that, Fable 5.2 finished and gave me the webpage.

I also recorded a video.

It also created a 3D voxel floating island.

Of course, the most outrageous is still the architecture of the Apple Store in Jing'an, Shanghai…

The prompt I provided was like this; I wanted to create a 3D webpage.

I didn't even expect that this case would take me nearly two and a half hours.

Fable 5.2, the first thing it did was search for reports about the Apple Store in Jing'an, which was no problem.

Then, it started looking for real photos to use as references…

It even zoomed in to look at the details.

In total, it looked at nearly a hundred images for almost an hour.

After finishing, it finally felt like it was ready to start modeling.

This took nearly another hour.

I was actually quite anxious inside because I didn't know what level of output it could achieve.

Because this case had been run before when GPT-6 Astra had just come out, and at that time, achieving such effects was already very impressive.

I didn't know what Claude Fable 5.2 could produce.

Two and a half hours later, Fable 5.2 submitted its work.

The static images looked like this.

The video presentation looked like this.

Because I found that it was indeed seriously modeling, I even specifically found an indoor effect picture of the Apple Store in Jing'an for comparison.

Note that this is a normal perspective, and one can actually be considered wide-angle.

Although it can't be said to be completely identical, in terms of the material representation in the modeling, the restoration of space, and the general proportional relationships, it was done extremely well.

In addition to the core material representation and the restoration of the internal space being very good, Fable 5.2 also extended the details as much as possible to the outdoor circular sunken plaza:

I can only say that Fable 5.2 is indeed incredibly strong in aesthetics and details.

I ran about 10 tasks, basically all Fable 5.2 Xhigh. Do you all remember what I said at the beginning about how much quota I had left before running the tests?

At that time, I had 30% left, and I used 70%.

And now, it is stuck at 94%.

Nearly four hours, 10 tasks, consumed about 24% of my quota, and the usage of Fable was almost 50% used up.

To be honest, it is indeed quite expensive.

So although the Fable model is SOTA and its capabilities are indeed strong, the price is a factor, so in daily use, it will likely still be Opus 5.5 and the soon-to-be-released GPT-6.2 Sol (which I have already experienced for two days and has a very high cost-performance ratio, comparable to Opus 5.5) as the main options. Fable 5.2 is only suitable for the largest diagnostic planning and command-type models, while aesthetics and front-end related tasks should be handled by Opus 5.5, and algorithms and back-end related tasks should be handled by GPT-6.2 Sol. This might be what I consider the optimal solution.

Honestly, models like Fable 5.2 are the best for personal agents, just like you are the emperor, and it is the chancellor. It guesses your thoughts based on your preferences and goals, and directs other ministers to work. This is just perfect, and the quota won't be burned too much.

That's about it.

Due to quota and time constraints, there are still many, many things I want to test that I haven't tested.

For example, really creating a game.

For example, starting to try to draw using code.

There are so many.

Even more inclined towards hardware and scientific research.

It feels like every advancement in large models brings more imagination to everyone.

Yes, there are still so many fun things in this world that have not been explored.

Creation, and AI.

It's really fascinating.

It's another morning, wishing everyone good morning, good afternoon, and good night.

Today, let's be full of curiosity about the world~