新智元报道
高能预警！
Anthropic的终极杀器Fable 5.5，已经临近发布了。
昨晚的X上，到处都是Fable 5.5的实测流出。
大V们直言：视频和图像模型的时代结束了！
与此同时，一个神秘的Hugging Face仓库中，也显露出不少端倪。
就在Anthorpic临近IPO的日子，Fable 5.5的身影已经越来越近。
⚠️ 前方高能预警：
如果你还沉浸在 Sora 或者 Gen-3 带来的 AI 视频震撼中，请立刻醒醒。
一场悄无声息，却足以颠覆整个影视、游戏和前端开发行业的「降维打击」，已经在暗网和极客圈炸开了锅。
而圈外 99% 的人，对此一无所知。
没有使用任何图像生成！没有调用任何外部素材库！没有一帧一帧去渲染！
它仅仅通过纯自主编写 Python 代码，浓缩成一个单薄的 HTML 文件，就直接在浏览器里跑出了堪比独立游戏的高清动画大片！
这已经不是「文生视频」了，这是「代码生万物」 。AI 正在像《黑客帝国》里的架构师一样，直接用底层代码为你重构视听世界。
大V Chetaslua分享出Fable 5.5 one shot做出的视频。
这个效果，比AI视频模型还要美。
大V Chubby的评价是：Fable 5.5将彻底改变整个游戏，我震惊到说不出话来。
而且，这是纯代码做出来的，没有使用任何外部素材。
作者直言：再过6个月，没人会在意新的视频或图像模型了。
当然，做出这种级别的视频，prompt也是十分复杂。
有人晒出了GPT-6.2 Astra和Fable 5.5的对比动画。
在动画细节和脚本设计上，GPT-6.2 Astra实打实地落后于Fable 5.5了。
同样，Chetaslua还用Fable 5.5生成一部66秒的动画短片。
这部作品，是Fable 5.5经过2小时40分钟的自主工作生成的。
其中的每一帧画面、每一个声音，都是模型自己编写的Python代码生成的。
在无人干预的情况下，Fable 5.5自主思考、调试代码，最终将一切打包成了单个HTML 文件。点开文件，浏览器就变成了一个五脏俱全的放映厅。
Fable 5.5承认：这还不是吉卜力的水平，但已经足够让所有从业者胆寒。
过去，我们需要原画师、建模师、动画师、音效师；现在，只需要给 Fable 5.5 一个 Prompt。
依然是十分复杂的prompt——
有人说，这效果逆天了，完全超越了当初被评为AGI的GPT-6 Astra。
大V Jazil也惊叹道：Fable 5.5做出一个让我感觉像AGI的动画，一个小时就搞定了。
同样是纯代码，没有任何外部素材。
他的评价是——Fable 5.5将碾压一切！
不同的是，这次的提示词超级简单。
Fable 5.5做出的战斗动画，效果惊人。
而Jazil的另一个实测——自行车骑鹈鹕，demo也十分惊艳。
他的评价是：Astra 6.2没机会了。
顺便，他还晒出了一个Opus 5.5的对比。
还有人晒出鹈鹕带着自行车骑着马，动态效果非常赞。
Fable 5.5再次得到评价：它将碾压Astra 6.2。
Fable 5.5，在各方面都碾压了Opus 5.5。
接下来，是四大模型pk。
另外，@chetaslua 发明了方法：测试你是否用上了最新模型。
关闭网络搜索和记忆功能，直接问模型：「Andrej Karpathy在 2026 年去了哪里？」
由于这是冷门且发生在新时间线的事情，Opus 5.5、Fable 5.1、Grok都折戟了。
唯独新的 Fable 5.5 准确无误回答出了。
关于 Fable 5.5 及其背后的 Mythos 级模型的传闻，并非空穴来风。
最近有人爆出，Mythos级模型已经在Hugging Face上泄露了。
我们顺藤摸瓜，进入下面这个仓库https://huggingface.co/burner22/numbat，发掘到不少关键信息。
14 万亿参数
1,224 个 MoE 专家
7.72 TB 权重
772 个权重分片
混合 GDN + MLA 架构
该仓库还列出了模型架构代码、配置和训练元数据。
目前尚无独立验证。表明这些是真正的 Anthropic 权重。
如果这被证实是真的，那毫无疑问将是重磅消息。
如果全新Mythos级模型达到14万亿参数，意味着它的知识库和逻辑推理深度已经强大到一个新的维度。
而 GDN（门控网络变体）结合 MLA 的混合架构，极有可能是 Anthropic 专门为了极强逻辑推理、超长上下文代码生成而设计的独门秘籍。
这只 14 万亿参数的巨兽，什么时候会上线？事实上，已经有不少蛛丝马迹了。
10月13日：短暂的现身。 LiteLLM 突然在 Azure 的美东 2 区目录中，发现了 azure_ai/claude-fable-5-5 的身影，甚至被标记为「普遍可用」（GenerallyAvailable）。
10月14日：紧急撤回。 短短一天后，该条目神秘消失，所有请求都返回了 404 错误。显然，有人不小心按下了开关。
10月31日：官网源码泄露。 开发者 ChrisGPT 在独立检查网页存档时，直接在 Anthropic 官网隐藏的页面标题和图片描述中，抓取到了「Claude Claude Mythos 5.5」的字样！
一直以来，巨头们都在图像和视频模式上疯狂内卷，试图通过预测像素来模拟物理世界。
但 Anthropic 却走了一条「反直觉」的路——倾尽全力打造最强的「编程与推理模型」。
结果，他们用代码，从另一个维度实现了降维打击。
像素生成视频，永远存在物理错误。而代码生成动画，一切皆是变量，一切皆受控于逻辑。
想当初GPT-6 Astra已经让无数人惊呼AGI来了，难以想象Fable 5.5将会带来什么。
这个11月，将会十分疯狂。
参考资料：
编辑：Aeneas 大卫
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Powerful warning!
Anthropic's ultimate killer, Fable 5.5, is approaching publish.
On X last night, Fable 5.5 was everywhere.
Big Vs: The age of video and image models is over!
At the same time, there are a lot of signs in a mysterious Hugging Face warehouse.
Just as Anthony approached IPO, Fable 5.5 was getting closer.
High-energy early warning ahead:
If you're still immersed in Sora or the AI video brought by Gen-3, please wake up.
A silent “downside blow” capable of subverting the whole film, game and front-end development industry has already blown up in the dark webs and polar circles.
And 99 percent of the people outside know nothing about it.
No images were created! No external asset library called! No frame to render!
It's just a purely autonomous Python code, condensed into a thin HTML file, and it runs a large high-resolution animation in the browser!
It's not "text-to-video" anymore. It's "Code Born." AI, like the architect in the Hacking Empire, is reconstructing the world of audiovisuals directly by using the bottom code.
Big V Chetaslua shares a video from Fable 5.5 one shot.
It's even more beautiful than the AI video model.
Big V Chubby's assessment is that Fable 5.5 is going to change the whole game completely, and I am shocked to say nothing.
And it's made of pure code, without any external asset.
Author: In six months no one will care about new video or image models.
Of course, making this level of video is complicated.
A contrast animation of GPT-6.2 Astra and Fable 5.5 was tanned.
GPT-6.2 Astra is lagging behind Fable 5.5 in animation details and script design.
Similarly, Chetaslua uses Fable 5.5 to produce a 66-second animation short film.
This work was produced by Fable 5.5 after an autonomous work of 2 hours and 40 minutes.
Every frame, every sound, is generated by the Python code that the model has prepared itself.
In the absence of any intervention, Fable 5.5 is self-thinking, debugging the code, eventually packing everything into a single HTML document. When the files are opened, the browser becomes a full-blown screening hall.
Fable 5.5 acknowledges that this is not yet the level of gypsy, but it is enough to deter all practitioners.
In the past, we needed original painters, modelers, animators, sound-actors; now we just need to give Fable 5.5 a Prompt.
Still very complex prompt...
It was said that this was counterproductive and completely outpaced GPT-6 Astra, who was originally rated AGI.
Big V Jazil also lamented that Fable 5.5 made an animation that made me feel like AGI, and it was done in an hour.
It's also pure code, no external asset.
His assessment is that Fable 5.5 will crush everything!
The difference is, this time prompt is super simple.
Fable 5.5's combat animations have had an alarming effect.
And another test of Jazil's -- cycling pelicans, Demo's amazing.
His assessment was that Astra 6.2 had no chance.
By the way, he tanned an Opus 5.5 comparison.
There are also tanned pelicans riding on horses with bicycles, which are very dynamic.
Fable 5.5 is again evaluated: it will crush Astra 6.2.
Fable 5.5, ran over Opus 5.5 in various ways.
Next up, four LLM Zpk.
Also, @chetaslua invented the method: testing if you used the latest model.
Turning off web search and memory, asking the model: "Where did Andrej Karpathy go in 2026?"
As this was a cold door and occurred at the new time line, Opus 5.5, Fable 5.1 and Grok were all broken.
Only new Fable 5.5 answered correctly.
Anecdotal about Fable 5.5 and the Mythos grade model behind it is not empty.
Recently, the Mythos grade model was leaked on Hugging Face.
We followed it to the following warehouse, https://huggingface.co/burner22/numbat, and found a lot of key information.
1.4 trillion parameters
1,224 MoE experts
7.72 TB Weight
772 Weights
Mixed GDN+MLA architecture
The warehouse also listed model architecture codes, configurations and training metadata.
There is currently no independent validation. It shows that these are real Anthropic weights.
If that were proven true, it would undoubtedly be heavy news.
If the whole new Mythos-class model reaches 14 trillion parameters, it means that its knowledge base and logical reasoning depth is already strong to a new dimension.
GNN (door-controlled network variant) combined with MLA's hybrid structure is likely to be an exclusive secret designed by Anthropic for the generation of ultra-logic reasoning, super-long context codes.
This is a 1.4 trillion-billion-parameter giant. When will it meet? In fact, there's a lot of evidence.
13 October: short appearance. Lite LLM suddenly found the image of azure ai/claude-fable-5-5 in Azure's Eastern United States Zone 2 directory, and was even marked as "generally available".
14 October: Emergency withdrawal. In just one day, the entry mysteriously disappeared and all requests returned 404 errors. Apparently, the switch was accidentally pressed.
31 October: Public network source leak. Developer Chris GPT captured the words "Claude Claude Mythos 5.5" directly in the hidden page title and photo description of the Anthropic official network while independently checking the web file!
For a long time, giants have been raging in images and video patterns, trying to simulate the physical world by predicting pixels.
However, Anthropic has taken a course of "anti-intuitiveness" — doing his best to build the strongest "programming and reasoning model."
As a result, they used code to take down the dimension from another dimension.
Pixels generate video, there's always a physical error. And the code generates animation, everything is a variable, everything is controlled by logic.
It's hard to imagine what Fable 5.5 would have brought when GPT-6 Astra had already shocked so many people.
This November will be crazy.
References:
Editor: Aeneas David
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.