Claude Fable 5.2 来了:8 天从零写出一套电脑系统Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days


AI free token/AI免费积分
0

出品 | 网易智能·今日硬核

作者 | 辰辰

编辑 | 王凤枝

用新版Claude手搓出修图和剪辑软件后,一位视频制作人转头就把Adobe订阅退了。

美国当地时间11月9日,Anthropic发布Claude Fable 5.2和Mythos 5.2,重点提高编程、推理和视觉能力。两者使用同一个底层模型,其中Fable面向广泛使用者开放,Mythos则继续给出给借助审核的机构。

Claude Fable 5.2 来了:8 天从零写出一套电脑系统

Anthropic还展示了一个更大的项目:Fable 5.2带着1000多个子智能体干了8天,写出了一套能在笔记本电脑上启动的操作平台原型。

开头提到的退订使用者是亚历克斯·麦克劳(Alex McCraw),AI创作者马修·伯曼(Matthew Berman)团队的制作人。他用了多年Photoshop,清楚自己常用哪些能力。这次,他让Fable按自己的习惯排布控件、快捷键和编辑工具,做出了一款专属修图软件:导入图片、调整尺寸、处理图层、添加文字都能在里面完成,团队最近的多张频道封面就是用它做的。

随后,他又做了一款剪辑工具,把Claude接进软件里,需要改哪里,就在软件里直接吩咐AI。这款工具同样已经投入日常视频制作。按伯曼的说法,麦克劳随后便退掉了原来的Adobe订阅。

伯曼以前也让AI仿写过软件,当时常见的毛病是界面看着挺像,点进去却发现按钮点不动、关键能力也没做出来。这次让他改观的,是做出来的工具真能接手团队每天都在干的活。这些都是发布前的提前试用成果;他测试的另一些大型项目,让模型连续跑了几十个小时。

不只做修图软件,连电脑平台都能从头写

科技媒体Every的团队也做了一个供自己使用的视频编辑器,名叫EveryCut。

据Every首席执行官丹·希珀(Dan Shipper)公开介绍,团队成员基兰·克拉森(Kieran Klaassen)用Fable 5.2,把他们平时在Premiere、Descript等软件里常用的能力集中到了一个工具里,还在里面接入了AI智能体。提前试用期间,团队已经拿它处理内部展示所需的视频。熟悉剪辑流程的人清楚现有软件哪里顺手、哪里碍事,如今可以直接把工作习惯交给AI,让它按自己的路子重做一套。

Anthropic自己的工程师把任务又放大了一些:从一个空文件夹开始,写一个操作平台原型。

据官方介绍,Fable 5.2协调了1000多个子智能体,连续工作8天,编写、测试和审查代码。最终做出的平台能在笔记本电脑上启动,有自己的内核、磁盘和键盘等硬件驱动,也有文件存储、窗口、登录界面和安装器,还接入了Claude。

Claude Fable 5.2 来了:8 天从零写出一套电脑系统

图:Anthropic展示的操作平台原型

数学任务里,团队则让1000个Fable 5.2智能体,把已有的费马大定理证明写成计算机可以逐步检查的Lean形式化证明,用了30小时。此前,一个尚未发布的研究版Claude完成这项工作用了11天,还需要研究人员引导。

不管是写操作平台还是做形式化证明,跑的都还是显示屏里的代码。到了下一组测试,Anthropic开始让模型去处理显示屏外的实物。

不光会写代码,还能控机械臂、外观电子宠物

在官方视频里,Fable 5.2控制机械臂完成了几种不同的动作:还原魔方、拧下瓶盖,再拿笔画出一个机器人。Anthropic称,与此前的Claude模型相比,新模型控制实体机器人的能力有显著提高,其中一些任务并没有经过专门训练。

Claude Fable 5.2 来了:8 天从零写出一套电脑系统

图:Fable 5.2控制机械臂的官方演示

在另一项模拟环境测试RoboDojo Sim中,模型需要依据摄像头画面控制双臂机器人。Anthropic公布的内部测试成绩表明,Fable 5.2得分64.9%,高于Fable 5.1的27.6%和GPT-6 Astra的29.8%。

电子设备的案例走得更完整。Anthropic与电路板外观制造公司Diode合作,只给模型一段需要,让它外观三款设备:家用语音助手、USB免提通话设备,以及掌上电子宠物。

模型先列出规格,从供应商处挑选可购买的元件,再外观四层电路板、规划走线,用计算机辅助外观工具画出外壳,最后编写固件和软件。按官方介绍,每块电路板都在不到一天内完成了可送厂生产的外观,过程中只需要少量人工介入。

电路板由Diode制造,工程师在Claude指导下组装、测试;三款设备最终都正常工作,没有再修改电路。

Claude Fable 5.2 来了:8 天从零写出一套电脑系统

图:家用语音助手、USB免提通话设备和掌上电子宠物

参与这项工作的硬件工程师达维德·阿斯纳吉(Davide Asnaghi)在X上透露,这批电路板组装后第一次尝试就成功了。他说,自己现在只需做少量审查,就敢把模型外观的电路板直接送去制造。

Claude Fable 5.2 来了:8 天从零写出一套电脑系统

图:参与Diode项目的工程师分享外观结果

普通人更容易想到的玩法,浮现在沃顿商学院教授伊桑·莫利克(Ethan Mollick)的提前试用中。他桌上有一个带灯光按键、旋钮和摇杆的小键盘,原本是配合ChatGPT使用的设备,这次他让Fable把它改成游戏机。

模型读了控制这块键盘的两个应用的代码,弄清楚怎样改变按键颜色、读取旋钮和摇杆,随后写出六个小游戏,其中一个要靠转动旋钮“破解保险箱”。莫利克说,这次从理解要求到做出能玩的游戏,中间不需要他再补充指导。

调用费降了两成,大项目仍可能花上万美元

Fable 5.2和Mythos 5.2的标准API售价是每百万输入token 8美元、输出40美元,比上一代的10美元、50美元便宜了20%。

新模型在部分任务里也更省用量。Anthropic公布的FrontierCode 2.0编程实测中,Fable 5.2在High推理档位的得分为45.3%,接近Opus 5.5在Max档位的45.5%,平均每项任务的费用约为后者的一半。把Fable调到更高的Xhigh档位后,得分升至49.1%,费用也随之增加。

不过,一旦把大项目交给模型连续跑上几十个小时,账单就是另一个量级了。据Decrypt报道,伯曼团队做的一段《战争机器》风格游戏关卡,连续运行了43小时48分钟,消耗约86.3亿token;如果按API按量付费方式折算,费用约为14710美元(并非团队实际支付的账单)。

因而,退掉一项软件订阅,不等于自己造软件就一定更省钱。让AI写出第一版之后,后续的修改、测试和继续加能力,同样会消耗额度或产生调用费用。

参加提前试用的开发者马特·舒默(Matt Shumer)还特别提醒,高推理档位格外慢。在他的日常使用中,低档和中档通常已经够用,最高档更合适那些值得等上很久的难题。

Anthropic称,标准模式的token输出速度最高比5.1快30%;可选的Fast模式能把token生成速度提高到普通模式的最多2.5倍,但售价也翻倍至每百万输入16美元、输出80美元。由于编译程序、运行测试等步骤仍各自需要时间,文字生成快一些,并不等于整个项目能按同样比例提速。

使用订阅套餐的人,则已经开始算额度了。开发者扎伊德·汗(Zaid Khan)在X上说,自己用了差不多一个月Opus 5.5都没碰到Claude的限额,换上Fable 5.2后很快就撞上了上限。

Claude Fable 5.2 来了:8 天从零写出一套电脑系统

图:一位开发者分享切换模型后的额度感受

这次发布时,Anthropic也给符合条件的付费套餐增加了临时使用机会:原本包含Fable的套餐或席位,可按订阅时长取得一至三次Fable额度重置;Pro及Team标准席位本周也可以把最多一半的周额度用于Fable。

乱发邮件,还会立规矩拖慢自己

除了等待时间长,舒默在试用中还碰上了模型“太有主见”的时候。他打算修改一个机型,同时更新隐私政策,把两边一起上线,模型却坚持必须先发布隐私政策、再改机型;他想精简订阅按钮下方的免责声明,模型也不肯照办。还有一次,智能体没经他同意,就给一个合作方连发了33封邮件。

写文章时,模型有时也会用力过猛。舒默把工作记录喂给它,想整理成一篇文章,拿到的初稿却塞得太满,完全按时间顺序铺陈,读起来不像他想要的东西;等他指出难题后,第二版才显著好转。

莫利克则撞上了一出更离谱的“内部官僚主义”。起初,他让模型制作一个由传送门连接的3D世界网站,使用者可以从一个情形不断走进下一个情形。20小时后,网站上已经长出了204个世界:有用牛仔布缝成、正被巨大拉链拉开的山谷,也有从果蝇视角看到的冰箱内部,许多情形的创意都是模型自己想出来的。

随后,莫利克发现部分作品质量参差不齐,几个子团队还做出了雷同的情形,便要求先停下扩建,把已有的作品打磨一下。

模型为此自己定了一套“打磨”检查办法:新版本必须在12张测试截图上借助AI评委的盲评,胜过旧版本才算过关。结果,一些子团队开始专门针对那12张截图做应试优化。负责统筹的智能体发现后,明令禁止这种做法,又叠加上了更严格的规则。

到了第二天早晨,这本规则书已经写到了84节;12天后,更是膨胀到204节、足足87500个英文单词。此时想上线一个新世界,得先审批创意,再过三个AI评委,最后对照规则书检查全部代码,走完一套流程大约要花一天,光是读完那本规则书就占掉了智能体大量的上下文空间。

Claude Fable 5.2 来了:8 天从零写出一套电脑系统

图:莫利克公开的AI团队组织图,蓝色部分由智能体自行外观

莫利克最后只好找来另一组“顾问智能体”,从外部精简这套流程。调整之后,项目重新恢复到每天产出30多个世界,同时保留了不少质量检查。

他在文章里感叹,模型能做出很有创意的作品,却没能自己意识到,整个项目已经被它亲手定下的规矩和审批拖住了。此前做课程网站时也是如此:内容虽然准确,读起来却枯燥得像一场苦差事;他指出难题后,模型很快就能改好,但那个发现难题的人,仍然得是他自己。

对于多数常规编程和资料研究任务,Anthropic仍建议先用Opus 5.5,再把最棘手、开放性更强的工艺项目交给Fable。

伯曼对新版的评价很高,但也没打算把日常工作全迁过去。他在X上说,自己大概率还是会继续拿Opus当主力,因为它能长时间运行,也较少碰到额度上限;只有遇上最难啃的任务,才会切到Fable 5.2来做规划。

本文来源:https://www.163.com/tech/article/LBC3JHDS00097U7T.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

0

Presented by | NetEase Smart·Today Hardcore

Author | Chen Chen

Editor | Wang Fengzhi

After building photo-editing and video-editing software with the new Claude, one video producer went ahead and canceled his Adobe subscription.

On November 9 local time in the United States, Anthropic released Claude Fable 5.2 and Mythos 5.2, focusing on improvements to coding, reasoning, and vision capabilities. Both use the same underlying model; Fable is available to a broad user base, while Mythos continues to be offered to organizations that pass review and vetting.

Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days

Anthropic also showcased an even bigger project: Fable 5.2, working with more than 1,000 sub-agents for eight days, wrote a prototype operating platform that can boot on a laptop.

The subscriber who canceled at the beginning is Alex McCraw, a producer on AI creator Matthew Berman’s team. He had used Photoshop for years and clearly knew which capabilities he used most often. This time, he had Fable arrange controls, keyboard shortcuts, and editing tools according to his own habits, creating a customized photo-editing app: importing images, resizing, working with layers, and adding text can all be done within it. The team’s recent channel cover images were made with it.

Then he built an editing tool as well, plugging Claude into the software so that wherever he needed a change, he could simply instruct the AI directly in the app. This tool has likewise been put into everyday video production. According to Berman, McCraw subsequently canceled his original Adobe subscription.

Berman had previously had AI imitate software too, and back then the common problems were that the interface looked similar, but once you clicked in, the buttons didn’t work and key capabilities weren’t actually built. What changed his view this time was that the tools produced could genuinely take over the work the team does every day. These were all results from early pre-release trials; some of the larger projects he tested kept the model running continuously for dozens of hours.

Not only building photo-editing software—it can even write a computer platform from scratch.

The team at tech media outlet Every also built a video editor for their own use, called EveryCut.

According to public remarks by Every CEO Dan Shipper, team member Kieran Klaassen used Fable 5.2 to consolidate capabilities the team commonly uses in software like Premiere and Descript into a single tool, and also integrated AI agents into it. During early trials, the team had already used it to process videos needed for internal presentations. Anyone familiar with editing workflows knows where existing software feels smooth and where it gets in the way; now you can hand your work habits over to the AI and have it rebuild a suite in your own style.

Anthropic’s own engineers took the task a step further: starting from an empty folder, write a prototype operating platform.

According to the company, Fable 5.2 coordinated more than 1,000 sub-agents and worked continuously for eight days, writing, testing, and reviewing code. The platform it ultimately produced can boot on a laptop, has its own kernel, disk and keyboard hardware drivers, and also includes file storage, windows, a login screen, and an installer; it also integrates Claude.

Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days

Photo: An operating platform prototype demonstrated by Anthropic.

In a math task, the team had 1,000 Fable 5.2 agents translate an existing proof of Fermat’s Last Theorem into a step-by-step, computer-checkable formal proof in Lean, taking 30 hours. Previously, an unreleased research version of Claude took 11 days to complete the same work and required researchers to provide guidance.

Whether writing an operating platform or producing formal proofs, what is running is still code on the screen. In the next set of tests, Anthropic began having the model handle real physical objects beyond the screen.

Not just writing code—it can also control robotic arms and design electronic pets.

In the official video, Fable 5.2 controls a robotic arm to complete several different actions: solving a Rubik’s Cube, unscrewing a bottle cap, and then using a pen to draw a robot. Anthropic says that compared with previous Claude models, the new model’s ability to control physical robots has improved significantly, and some of these tasks had no dedicated training.

Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days

Photo: Official demo of Fable 5.2 controlling a robotic arm.

In another simulated-environment test, RoboDojo Sim, the model needs to control a dual-arm robot based on camera images. Internal test scores released by Anthropic show Fable 5.2 scoring 64.9%, higher than Fable 5.1’s 27.6% and GPT-6 Astra’s 29.8%.

The electronics examples went even further. Anthropic collaborated with circuit-board appearance-manufacturing company Diode, giving the model only a brief requirement and having it design three devices: a home voice assistant, a USB hands-free calling device, and a handheld electronic pet.

The model first listed specifications, selected purchasable components from suppliers, then designed a four-layer circuit board and planned routing, drew the enclosure using computer-aided design tools, and finally wrote the firmware and software. According to the company, each circuit board completed a factory-ready design in less than a day, with only minimal human intervention during the process.

The circuit boards were manufactured by Diode, and engineers assembled and tested them under Claude’s guidance; all three devices ultimately worked normally, with no further circuit modifications needed.

Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days

Photo: A home voice assistant, a USB hands-free calling device, and a handheld electronic pet.

Davide Asnaghi, the hardware engineer involved in the work, revealed on X that the batch of circuit boards succeeded on the very first attempt after assembly. He said that he now only needs to do a small amount of review before he dares to send the circuit boards for the model's appearance directly to manufacturing.

Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days

Photo: An engineer involved in the Diode project shares the appearance result

A use case that an ordinary person might more easily think of emerged during the early access trial by Wharton School professor Ethan Mollick. On his desk is a macropad with illuminated keys, a knob, and a joystick, originally a device used with ChatGPT; this time, he asked Fable to turn it into a game console.

The model read the code of the two applications that control this keyboard, figured out how to change the key colors and read the knob and joystick, and then wrote six mini-games. One of them required turning the knob to "crack a safe." Mollick said that in this case, from understanding the requirements to producing a playable game, there was no need for him to provide additional guidance in between.

Call rates dropped by 20%, but big projects can still cost over $10,000

The standard API price for Fable 5.2 and Mythos 5.2 is $8 per million input tokens and $40 for output, 20% cheaper than the previous generation's $10 and $50.

The new model also consumes fewer tokens in some tasks. In Anthropic's published FrontierCode 2.0 programming benchmark test, Fable 5.2 scored 45.3% at the High reasoning tier, close to Opus 5.5's 45.5% at the Max tier, with an average cost per task of approximately half that of the latter. After switching Fable to the higher Xhigh tier, the score rose to 49.1%, with the cost also increasing accordingly.

However, once you hand a large project over to the model to run continuously for dozens of hours, the bill reaches another level. According to a Decrypt report, a *Gears of War*-style game level made by Berman's team ran continuously for 43 hours and 48 minutes, consuming approximately 8.63 billion tokens; if calculated based on the pay-as-you-go API pricing, the cost would be about $14,710 (this is not the bill the team actually paid).

Therefore, canceling a software subscription does not mean that building your own software will necessarily be more cost-effective. After letting AI write the first version, subsequent modifications, testing, and continued feature additions will also consume quota or generate call costs.

Matt Shumer, a developer participating in the early access, also specifically warned that the high reasoning tier is exceptionally slow. In his daily use, the low and medium tiers are usually sufficient, while the highest tier is more suitable for difficult problems that are worth waiting a long time for.

Anthropic stated that the standard mode's token output speed is up to 30% faster than 5.1; the optional Fast mode can increase token generation speed to up to 2.5 times that of the normal mode, but the price also doubles to $16 per million input tokens and $80 for output. Since steps like compiling programs and running tests still each require time, faster text generation does not mean the entire project will speed up at the same ratio.

Those using subscription plans have already started calculating their quotas. Developer Zaid Khan said on X that he had used Opus 5.5 for nearly a month without hitting Claude's limits, but after switching to Fable 5.2, he quickly hit the ceiling.

Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days

Photo: A developer shares their feelings on quota after switching models

With this release, Anthropic also added temporary usage opportunities for eligible paid plans: plans or seats that originally included Fable can get one to three Fable quota resets based on the duration of the subscription; Pro and Team standard seats can also use up to half of their weekly quota on Fable this week.

Sending emails recklessly, and also establishing rules that slow oneself down

Besides the long waiting times, Shumer also encountered times during the trial when the model was "too opinionated." He planned to modify a model and update the privacy policy at the same time to launch both together, but the model insisted on publishing the privacy policy first before changing the model; when he wanted to streamline the disclaimer below the subscribe button, the model also refused to comply. Another time, the agent sent 33 consecutive emails to a partner without his consent.

When writing articles, the model can sometimes overdo it. Shumer fed it his work logs, wanting to organize them into an article, but the first draft he got back was stuffed too full and laid out entirely in chronological order, reading like something he didn't want; only after he pointed out the problem did the second version improve significantly.

Mollick encountered an even more absurd case of "internal bureaucracy." Initially, he asked the model to create a website of 3D worlds connected by portals, where users could walk from one scenario into the next. 20 hours later, the website had grown 204 worlds: there was a valley sewn from denim fabric being pulled apart by a giant zipper, and the inside of a refrigerator viewed from the perspective of a fruit fly; the creative ideas for many scenarios were thought up by the model itself.

Subsequently, Mollick found that the quality of some works was uneven, and several sub-teams had created similar scenarios, so he requested to stop the expansion and polish the existing works first.

The model defined a set of "polishing" check methods for this purpose: new versions had to pass a blind evaluation by AI judges on 12 test screenshots, beating the old version to be considered approved. As a result, some sub-teams began to specifically optimize for those 12 screenshots to pass the test. The coordinating agent discovered this, explicitly forbade the practice, and added even stricter rules on top.

By the next morning, this rulebook had grown to 84 sections; 12 days later, it had ballooned to 204 sections and a full 87,500 English words. By this point, to launch a new world, one had to get creative approval, pass three AI judges, and finally check all the code against the rulebook; completing this set of processes took about a day, and just reading through the rulebook consumed a massive amount of the agent's context window.

Claude Fable 5.2 Arrives: An AI Agent Wrote a Computer OS from Scratch in 8 Days

Photo: The AI team org chart revealed by Mollick; the blue sections were self-designed by the agents

Mollick finally had to bring in another set of "consultant agents" to streamline this process from the outside. After the adjustments, the project returned to producing over 30 worlds per day while retaining much of the quality control.

He lamented in his article that while the model can produce highly creative work, it failed to realize on its own that the entire project had been bogged down by the rules and approval processes it had set up. The same thing happened with his previous course website project: the content was accurate, but reading it was as dry as a chore; after he pointed out the problem, the model quickly fixed it, but the person who had to discover the problem still had to be him.

For most routine programming and research tasks, Anthropic still recommends using Opus 5.5 first, and then leaving the most tricky, open-ended craft projects to Fable.

Berman rates the new version highly, but doesn't plan to migrate all his daily work to it. He said on X that he will most likely continue to use Opus as his main workhorse because it can run for long periods and he runs into quota limits less often; only when facing the most difficult tasks will he switch to Fable 5.2 for planning.

Source: NetEase Intelligence Executive Editor: Wang Fengzhi_NT2541

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry