Anthropic尚未正式发布的下一代旗舰模型Claude Fable 5.5近日在海外开发者社区引发关注。多名开发者在社交平台X上分享了据称来自该模型的早期测试结果。其中，一段完全通过编程生成的66秒动画短片尤为引人注目。测试者称，模型在没有调用图像或视频生成工具的情况下，连续自主编写和运行代码约2小时40分钟，完成了动画画面、配乐及音效制作。
11月2日，X平台AI开发者Chetaslua发布了一段动画短片，并声称其由Fable 5.5独立完成。按照他的介绍，短片长度为66秒，分辨率1920×804，帧率24fps，画面包含狐狸、纸鸟、云层、文字及光影效果，背景音乐和音效同样由模型编写的Python代码生成。
This is a modal window.
A network error caused the media download to fail part-way.
自动播放
与Sora、Runway等直接生成视频画面的AI工具不同，这段演示展示的是一条通过代码制作动画的技术路径。开发者称，整个过程中没有使用外部图片、素材库、音频采样或字体文件，模型自主完成了场景设计、画面渲染和声音合成，并将最终作品封装成一个HTML文件。
除了动画制作，多名海外开发者还展示了据称由Fable 5.5编写的交互式3D场景，包括古帆船、海浪和跑车等。相关演示涉及Three.js等前端图形技术，通过程序控制模型、光照和物体运动，部分场景仅需简单提示词即可生成。
在另外几组复杂动画对比测试中，开发者Jazil等人展示了不同模型生成自行车、动物角色及人物战斗场景的结果。据测试者介绍，Fable 5.5能够在约一小时内完成复杂动态场景的代码编写与运行，而部分现有Claude模型在相同任务中出现代码不完整、动画运行失败等问题。不过，这些测试尚缺乏统一的评测环境，不能直接视为模型性能排名。
事实上，Fable 5.5的发布线索此前已经多次出现。10月中旬，第三方模型工具LiteLLM的Azure目录中曾短暂出现名为claude-fable-5-5的条目，随后被移除。近日，又有开发者发现，Anthropic官网部分页面的HTML源码中曾出现Claude Mythos 5.5字样，引发外界对新模型即将发布的猜测。
按照Anthropic现有产品体系，Fable与Mythos采用相同的底层模型，但实施不同级别的安全限制。Mythos主要面向经过审核的网络安全和生命科学研究机构，Fable则通过增加安全防护措施向普通用户及开发者开放。不过，Anthropic尚未确认5.5系列是否延续这一设计。
伴随新模型曝光，Hugging Face平台近日还出现了一个声称包含Mythos模型权重的仓库，参数规模据称达到14万亿。但研究人员检查发现，部分权重数据存在明显的人工合成迹象，所谓模型泄露并无可靠证据支持。
这轮模型曝光恰逢Anthropic推进IPO的关键阶段。公司11月2日披露的招股文件显示，其今年前三季度营收达到332.42亿美元，但持续增长的算力投入和巨额亏损，也让外界更加关注下一代模型能否带来实际的技术突破与商业回报。
随着AI模型竞争从代码补全转向长时间自主执行复杂任务，Fable 5.5展示的动画制作与3D场景生成能力，反映出AI编程智能体潜在的应用空间。
截至发稿，Anthropic尚未正式确认Fable 5.5的发布时间、性能指标及开放范围，上述测试结果仍有待官方证实和独立评测。
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。
The next-generation flagship model of publish, Anthropic, Claude Fable 5.5, has not yet officially been brought to the attention of overseas developers. A number of developers shared early test results allegedly from the model on social platform X. Of these, a 66-second animated film produced exclusively through programming is particularly striking. According to the tester, the model, without calling on the image or video generation tool, had been able to develop and run the code on its own for approximately 2 hours and 40 minutes, and had completed the production of animation images, music and sound effects.
On 2 November, ChetasluaZ publish, the developer of platform X, presented an animated short film, claiming that it had been independently completed by Fable 5.5. According to his presentation, the short film length was 66 seconds, resolution 1920 x 804, with a frame rate of 24fps, with images of foxes, paper birds, clouds, text and photo-effects, as well as background music and sound effects generated by the Python code prepared by the model.
This is a modal window.
A network error caused the media download to fail part-way.
Autoplay
Unlike the AI tools that directly generate video images such as Sora, Runway and others, this presentation shows a technical path to animating through code. The developers stated that no external photographs, asset library, audio sampling or font files were used throughout the process, and that the model had independently completed the scene design, image rendering and synthesis of the sound and sealed the final work into an HTML file.
In addition to animated production, several overseas developers presented interactive 3D scenes allegedly written by Fable 5.5, including sailboats, waves and sports vehicles. The demonstration involved front-end graphic techniques such as Three.js, which could be generated only by simple prompt through program control models, light and object movements.
In several other complex sets of animated comparison tests, developers, Jazil et al. showed the results of different models that generated bicycles, animal role and character combat scenarios. According to the tester, Fable 5.5 was able to complete code writing and running of complex dynamic scenarios within approximately one hour, while some of the existing Claude models had problems with incomplete codes and failed animation in the same tasks. However, these tests lack a uniform assessment environment and cannot be considered directly as model performance rankings.
In fact, Fable 5.5 ' s publish leads have appeared several times before. In mid-October, the Azure directory of the third-party model tool LiteLM briefly appeared and was subsequently removed. In recent days, another developer has discovered that Claude Mythos 5.5 was found in the HTML source code on part of the Anthropic official web page, triggering speculation about the new model's upcoming publish.
According to the current product system of Anthropic, Fable and Mythos use the same bottom model, with different levels of safety restrictions. Mythos is primarily oriented towards the audited institutions for cybersecurity and life science research, while Fable is open to ordinary users and developers through increased safety protection measures. However, Anthropic has not confirmed the continuation of the 5.5 series.
With the exposure of the new model, the Hugging Face platform has in recent days also seen the emergence of a warehouse claiming to contain the weight of the Mythos model, the size of which is said to be 14 trillion. However, a search by researchers found that there were clear signs of synthetic partial weight data and that the so-called model leak was not supported by reliable evidence.
The exposure of this round coincided with the critical phase of Anthropic pushing IPO. The company ' s disclosure on 2 November showed that it had collected $33,242 million in revenue in the first three quarters of the year, but the steady growth of the calculus inputs and the huge losses had also led to a greater focus on whether the next generation model would bring real technological breakthroughs and commercial returns.
As the competition for the AI model moved from code completion to long, autonomous, complex missions, the animation of Fable 5.5 and 3D scenario generation capabilities reflected potential application space for AI programming AI agent.
At the time of the release, Anthropic had not formally confirmed Fable 5.5 ' s publish time, performance indicators and open scope, and the results of the above-mentioned tests were still pending official confirmation and independent evaluation.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.