为什么Fable 5.1值得升级

Anthropic发布了他们的最新模型，Fable 5.1和Mythos 5.1。

Anthropic has released its latest models, Fable 5.1 and Mythos 5.1.

在基准测试上，它们无疑是最先进的，几乎在每个类别上都超越了现有的所有模型。

On the benchmarks, they are undeniably state of the art, outperforming everything else that exists on pretty much every category.

Anthropic还声称他们在成本上取得了重大进展，对于许多任务，包括长时间运行的代理任务，Fable 5.1的成本应该比Fable 5中相同任务低25%甚至40%。

Anthropic also claims that they've made major advances in the cost so that for many tasks including long running agentic tasks, Fable 5.1 should cost as much as 25 or even 40% less than the comparative task in Fable five.

初步反馈还不错，不过用户们在实际成本和使用限制下能用Fable 5.1走多远，体验差异挺大的。

Initial responses are pretty good, although users are getting pretty varied mileage in terms of just how much the costs actually are and how far you could even get with Fable 5.1 given usage limits.

不过，和每个新模型一样，现在的问题还是：这个模型够不够好，值不值得换过去用？

Still the question comes up, as it will now forever with every new model, is this one good enough that it's worth switching to?

但我觉得，这已经不再是正确的问题了。

Except I think that that's no longer the right question.

相反，我们应该问的是：这个模型能用来做什么？

Instead, the question should be, what can I use this model for?

它怎么融入我整体的模型组合中？

How does it fit in to my overall model stack?

我该怎么充分利用它，同时又能接受它带来的各种取舍呢？

What can I do to take most advantage of it while recognizing whatever trade offs it comes with?

这就是我们今天节目要聊的内容，那我们就开始吧。

That's what we're getting into in today's episode, so let's dive in.

《AI每日简报》是一档每日更新的播客和视频节目，内容涵盖人工智能领域最重要的新闻和讨论。

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

好了，朋友们。

Alright, friends.

在开始之前，先快速说几个通知。

Quick announcements before we dive in.

首先，感谢今天的赞助商：KPMG、Blitzy、Robots and Pencils 和 Hyperagent。

First of all, thank you to today's sponsors, KPMG, Blitzy, Robots and Pencils, and Hyperagent.

想收听无广告版本，请访问 patreon.com/aidailybrief，或者也可以在 Apple Podcasts 上订阅。

To get an ad free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts.

想了解如何赞助本节目，请发送邮件至 sponsors@aidailybrief.ai。

To learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai.

另外，我最近一直在提，我们在Super Intelligent的下一轮高管智能体领导力项目，就在劳动节之后马上开始。

Also, as I've been mentioning recently, our next set of executive agent leadership programs at Super Intelligent are coming up just after Labor Day.

你可以在training.besuper.ai上了解相关信息。

You can find out about those at training.besuper.ai.

再说一遍，所有详情都在training.bsuper.ai上。

Again, you can find out all about that at training.bsuper.ai.

今天有几条挺劲爆的消息。

We have kind of a dramatic set of headlines today.

第一条是关于OpenAI即将推出的Astra的最新进展。

The first up is an update about OpenAI's forthcoming Astra.

在周二的一篇博客文章中，OpenAI表示，他们现在认为Astra已经达到了他们准备框架中关键网络安全能力的门槛。

In a Tuesday blog post, OpenAI said that they now believe that Astra meets the critical cybersecurity capability threshold under their preparedness framework.

用大白话说，这意味着这个模型能够在没有人类指导的情况下，发现并利用以前未知的安全漏洞。

In layman's terms, that means that the model is capable of finding and exploiting previously unknown security flaws without human guidance.

在八月初的评估中，OpenAI认为Astra有可能达到这个门槛，但当时还不确定。

In their previous assessment at the beginning of August, OpenAI believed that it was possible Astra would reach the threshold but weren't sure yet.

其实，这和今年早些时候Anthropic把Mythos锁起来的原因是一样的。

Essentially, this is the same concern that saw Anthropic keep Mythos under lock and key earlier this year.

OpenAI分享了一些评估Astra能力的细节，说这个模型在Exploit Bench上拿到了100%的满分。

Sharing some details on how they assessed Astra's capabilities, OpenAI shared that the model achieved a perfect 100% score on Exploit Bench.

这个基准测试是评估模型根据已知漏洞开发攻击代码的能力。

This benchmark evaluates a model's ability to develop exploits based on known vulnerabilities.

OpenAI还更进一步，自己开发了一个内部版本的基准测试，包含20个最近披露的高严重性漏洞。

OpenAI then took it a step further and developed their own internal version of the benchmark, consisting of 20 high severity vulnerabilities that were recently disclosed.

他们的想法是，用训练数据里不可能出现的测试，来检验模型是否真的能从头创造新的攻击代码。

The idea was to test whether the model was actually capable of creating novel exploits from scratch by using tests that couldn't be in the training data.

OpenAI写道：在这个数据集上，Astra的任意代码执行率远高于GPT-5 60 SOL，而且用的输出令牌少得多。

OpenAI wrote: On this dataset, Astra achieves much higher arbitrary code execution rates than GPT-five 60 SOL using far fewer output tokens.

在评估过程中，模型甚至发现并利用了零日漏洞，作为攻击链的一部分。

During the evaluation, the model even discovered and used two zero day vulnerabilities as part of an exploit chain.

为了给这个对比加上一些数字，Astra在内部版Exploit Bench上得了30%的分数，用了4万个令牌，而GPT-5 60 SOL直到用了大约11万个令牌才取得任何显著成果。

Now to put some numbers around this comparison, Astra managed a 30% score on their internal version of Exploit Bench with 40,000 tokens used as opposed to GPT56 SOL, which wasn't capable of any significant results until it spent around 110,000 tokens.

但如果把这种能力推广到其他方面，这可能意味着这个模型在运行各种代理任务时，token效率会高得多。

But if we extrapolate out to other capabilities, this could mean the model is much more token efficient for running agents across the board.

在与专家合作伙伴的进一步测试中，OpenAI发现Astra能够设计并执行完整的漏洞利用链，从而获得对加固操作系统的根访问权限，并在加固浏览器上执行命令。

In further testing with expert partners, OpenAI found that Astra was able to design and execute full exploit chains to gain root access to a hardened operating system and execute commands on a hardened browser.

因此，OpenAI将为Astra的发布部署一系列新的安全措施。

As a result, OpenAI will deploy a series of new safeguards for Astra's release.

模型本身接受了额外的训练，以拒绝网络安全任务，拒绝率达到了91.5%，而GPT-5.6 SOL的拒绝率是59%。

The model itself has received additional training to refuse cybersecurity tasks, resulting in a 91.5% refusal rate, up from 59% for GPT-five point six sold.

OpenAI还在增加更多的分类器来检测网络滥用和越狱尝试，此外，OpenAI现在会将某些账户标记为高风险，并对这些账户实施更严格的模型行为护栏。

OpenAI is also adding more classifiers to detect cyber abuse and attempted jailbreaks, and in addition OpenAI will now be flagging certain accounts as higher risk and applying more stringent model behavior guardrails to those accounts.

OpenAI表示，他们认为Astra比之前的模型更有可能遵守安全边界，但他们仍在实施额外的思维链监控，以便及早发现并阻止不一致的行为。

OpenAI says that they believe Astra is more likely to respect security boundaries than previous models, but they're still implementing additional chain of thought monitoring to detect and stop misaligned actions early.

山姆·奥特曼在X上发了一篇异常严肃的帖子，甚至用了正确的语法和标点，他补充说：这里显然存在一种紧张关系。

In an unusually serious post on X that even used like correct grammar and punctuation, Sam Altman added: There is an obvious tension here.

一方面，Astra非常出色，我们很期待看到人们会用它能做出什么。

On one hand, Astra is very good and we are excited to see what people will build with it.

我们为自己的工作感到自豪。

We are proud of our work.

但另一方面，我们显然正处于一个需要谨慎的发展阶段，我们正在调整进度，以确保能够达到新能力水平所要求的安全标准。

On the other hand, we are clearly in a phase of development where we believe caution is warranted, and we are pacing our progress to ensure that we can meet the safety standards required by new capability levels.

Astra 的训练已经完成一段时间了，它在能力和对齐方面都迈出了重要的一步。

Astra has been done with training for a while now and is a significant step forward in both capabilities and alignment.

对于之后的模型，我们一直在根据需要放慢节奏，以确保在安全和对齐方面做足工作。

For the models after that, we have been slowing things as needed to ensure that we can do sufficient work on safety and alignment.

他暗示了 OpenAI 内部的气氛，继续说道：我们一直生活在既兴奋又焦虑的张力之中，这种矛盾感对我们来说依然存在。

Hinting at the mood inside OpenAI, he continued: We've been living with the tension between being excited and anxious about progress for some time and it is still discordant for us.

我们知道对其他人来说这种矛盾感更强烈，但我们坚信世界需要了解 AI 的发展方向以及模型在现实世界中的表现。

We know it is much more discordant for other people, and yet we believe strongly that the world needs to understand where AI is going and how models perform in the real world.

更重要的是，我们相信世界需要对齐的 AI 来管理这一转型的未来阶段。

More importantly, we believe the world will need aligned AI to manage the future phases of this transition.

社会与这项技术共同演进的迭代循环，才是让我们最大可能做对这件事的关键。

An iterative loop where society and this technology evolve together is what will lead to the highest chance of getting this right.

所以我们希望大家喜欢我们的新模型，也希望世界继续认真对待人工智能领域正在发生的事情。

So we hope you enjoy our new model, and we hope the world continues to take what's happening in AI extremely seriously.

有消息说Astra可能这周就会发布，这时间点对我来说太合适了，因为我正在旅行，本来应该录一些预发布节目，但信不信由你，关于Astra的讨论可不止这些。

Sources suggest that Astra could be coming as soon as this week, which would be perfect timing given that I'm traveling and theoretically I'm doing preload episodes, but believe it or not that is not the only Discourse going on about Astra.

周二深夜，The Information爆料了一个技术突破，让Astra在推理能力上大幅提升，而且按照一些人的说法，它可能也因此变得更加危险。

In a late night scoop on Tuesday, the information revealed a technical breakthrough that makes Astra much better at reasoning and, according to some, potentially much more dangerous.

这项技术叫做“递归深度”，它使用了一种循环变压器架构。

The technique is called recurrent depth, which uses a looped transformer.

从功能上讲，这意味着模型可以在生成输出之前，多次处理同一段文本，以改进它的回答。

Functionally, this means the model can process the same text string multiple times to improve its response before generating an output.

虽然消息人士说这项技术提升了性能并降低了成本，但最大的问题在于可观测性不足。

While sources say the technique improved performance and reduced cost, the big downside is a lack of observability.

现在部分推理过程是在模型内部完成的，不会生成任何输出。

Part of the reasoning process now takes place inside the model without generating an output.

这意味着思维链会部分隐藏起来，人类无法读取或理解。

This means chain of thought will be partially obscured and unable to be read or understood by humans.

OpenAI的消息人士说，他们在Astra中有限地使用了这项技术，以确保推理过程仍然可以被充分监控。

OpenAI sources said that they've used the technique in a limited way in Astra to ensure that reasoning can still be adequately monitored.

不过，《The Information》写道，AI研究人员担心，如果其他AI开发者采用同样的技术，可能不会像OpenAI那样施加同样的限制，而且不加限制地使用这项技术可能会导致AI失控，其行为难以监督。

However, writes the information, AI researchers worry that some AI developers may not impose the same kind of limits OpenAI did if they adopt the same technique for their own models, and that unfettered use of the technique could potentially lead to runaway AI whose actions can be hard to oversee.

现在，从事AI安全的人士已经担心智能体的可观测性会变得更难。

Now, folks working in AI safety have already been concerned about agent observability getting more difficult.

在对Hugging Face攻击的分析中，Meader指出，日志对人类来说根本无法拼凑起来，需要AI分析才能获得全貌。

In their analysis of the Hugging Face attack, Meader noted that logs were impossible for a human to piece together and required AI analysis to get the full picture.

Dwarkash Patel在本周早些时候对这次攻击的戏剧性且有争议的叙述中评论道：我不认为这是我们收到的最后一次警告，但这可能是我个人能理解的最后一次。

Dwarkash Patel, in his dramatic and controversial retelling of the attack earlier this week commented: I don't think this is the final warning shot we'll get, but it's probably the last one that I'll personally be able to understand.

在这篇报道之后，Encode AI的Nathan Calvin发帖说：今晚《The Information》的报道真是巨大且极其令人担忧。

Following the report Nathan Calvin of Encode AI posted: Really huge and extremely concerning story from the information tonight.

看起来OpenAI在Astra中利用了神经租赁的突破，这可能会摧毁思维链的脆弱性。

Looks like OpenAI utilized a breakthrough in neural lease for Astra that could destroy chain of thought vulnerability.

很有可能的是，如果OpenAI发现了这种架构并获得了性能或效率上的提升，那么其他公司很可能也会很快发现它，如果它们还没有的话，而且它们可能不会选择以牺牲效率为代价来优先考虑可监控性。

It seems quite likely that if OpenAI discovered this architecture and found performance or efficiency gains, that other companies are likely to find it soon too if they haven't already, and may not choose to prioritize monitorability at the expense of efficiency.

如果有些人真这么干，可能很难避免一场恶性竞争。

If some folks do, it may be difficult to avoid a race to the bottom.

Redwood Research 的 Ryan Greenblatt，也是之前调查 Hugging Face 事件的主要研究员之一，他写道：我最担心的是，从这里自然发展下去，可能会把这种不透明的推理扩展到极致，让模型完全或几乎完全在潜在空间里思考。

Ryan Greenblatt of Redwood Research, who was one of the lead researchers on the meter investigation of the Hugging Face incident, wrote: My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in latent space.

这很可能会彻底毁掉思维链在监控和监管方面的作用。

This would very likely destroy the usefulness of chain of thought for monitoring and oversight.

前 OpenAI 研究员 Steven Adler 说：如果这是真的，OpenAI 似乎正在触碰 AI 行业里为数不多的几条红线之一。

Former OpenAI researcher Steven Adler said: If this is true, OpenAI seems to be violating one of the few red lines that exist in the AI industry.

绝对不要这样训练你的模型。

Absolutely do not train your models like this.

这到底是怎么回事？

What is going on?

不过，也有不少人试图出来缓和一下气氛。

Still, a number of folks tried to jump in and calm down sentiment a little bit.

The Information 的 Amira Frady 再次强调，他们的报道显示 OpenAI 正在限制这项技术的使用，并努力确保思维链可见，但她担心其他 AI 开发者可能不会这么做。

Amira Frady from The Information again jumped in to reinforce the notion that their reports suggest that OpenAI is putting limits on this technique and trying to make sure chain of thought is visible, but is concerned that other AI developers may not.

OpenAI的首席科学家雅各布·帕卡奇写道，我想阻止因混乱报道而引发的不可监控性竞赛。

And OpenAI chief scientist Jacob Pachacchi wrote, I want to prevent a race into unmonitorability kicked off by confused reporting.

我们当前前沿模型（包括Astra）的计算图深度，与GPT-4相比，差距在一到两倍以内。

The depth of the computation graph for our present frontier models, including Astra, is within a factor or two of GPT-four.

OpenAI从我们最早的推理模型开始，就一直致力于保留和利用思维链监控。

OpenAI has worked to preserve and utilize chain of thought monitoring since our very first reasoning models.

我们非常重视这项技术，因为它能让我们了解模型对齐如何从其训练分布中泛化。

We care deeply about this technique as it can give us a view into how model alignment generalizes from its training distribution.

我确实认为它很脆弱，而且不幸的是，由于一些与架构变化无关的原因，它正朝着不利的方向发展，我很快会写文章详述，但我们可以采取措施来加强它，这也是我们当前研究计划的核心目标。

I do think it is fragile and unfortunately trending in a negative direction for reasons not contingent on architecture changes that I will write about soon, but there are things we can do to strengthen it and it's a core goal of our current research program.

所以，你知道，又一个毫无争议的发布要来了！

So you know, another uncontroversial release coming up!

说到发布，《华尔街日报》报道称，Gemini 3.8 Flash即将推出，可能会解决谷歌AI最长期存在的问题之一——编程。

Speaking of releases, The Wall Street Journal reports that Gemini 3.8 Flash is on the way could fix one of the longest standing problems for Google's AI which is coding.

尽管谷歌努力尝试，但从未推出过最先进的编程模型，而如今在这一关键能力上，他们已经远远落后了。

Try as they might, Google has never produced a state of the art coding model and at this point they have fallen drastically behind in this critical capability.

但《华尔街日报》报道说，在公司内部测试期间，工程师们更喜欢他们即将推出的3.8 Flash模型，而不是Anthropic的Opus。

Yet the Journal reports that during testing within the company engineers preferred their forthcoming 3.8 Flash model to Anthropix Opus.

这个模型预计本周发布，可能就在今天，所以我们很快就能看到它是否名副其实。不过，文章还提到了Gemini Pro系列幕后的一些情况。

Now the model is expected to be released this week, possibly today, so we'll soon see whether it lives up to the hype, but the article also covered what's been happening behind the scenes for the Gemini Pro series.

有消息人士称，所有原本打算作为3.5 Pro发布的内部候选模型都被砍掉了，因为它们并没有比Flash模型好多少。

Sources said that all internal candidates to be released as 3.5 Pro were scrapped because they weren't sufficiently better than the Flash models.

不过，研究人员对Gemini 4在预训练评估中的表现感到满意。

However, researchers are pleased with the performance of Gemini four during pre training evals.

这个模型还在进行后训练，也就是说离准备好还有一段时间。但也许这对那些希望看到更多竞争，而不只是OpenAI和Anthropic的人来说是个好消息。

The model is still in post training, meaning there's more time before it's ready, but perhaps some good news for those who want to see more competition than just OpenAI and Anthropic.

现在，有一个模型，如果不是因为我们今天的主要话题是Fable 5.1，它很可能就是今天的主角，那就是World Labs新发布的Atlas模型。

Now one model which, were it not for our main topic of Fable 5.1, could have easily been the entire main topic for today, is World Labs' newly released model Atlas.

他们称这是世界上第一个多模态世界模型，能够生成图像和视频帧，并实现像素级的相机控制，还能在三维空间中重建它们。

They describe it as the world's first multimodal world model that generates image and video frames with pixel perfect camera control and reconstructs them in three d.

建模世界，移动相机，模拟空间和时间。

Model the world, move the camera, and simulate space and time.

World Labs的本·米尔登霍尔写道：Atlas是一个自回归扩散模型，从头开始为下一帧预测任务而构建。

World Labs' Ben Mildenhall writes: Atlas is an autoregression diffusion model built from the ground up for the task of next frame prediction.

它同时是一种世界级的方法，用于相机控制的视频生成、新视角合成和稀疏三维重建。

It is simultaneously a world class method for camera controlled video generation, novel view synthesis, and sparse three d reconstruction.

World Labs的联合创始人李飞飞写道：Atlas能够生成具有像素级相机控制的帧，从仅一张输入图像重建大型场景，通过重新构图视频来模拟时空，从一张或多张输入图像原生输出三维空间，将多个带姿态的图像组合成一个一致的三维世界，等等。

World Labs co founder Fei Fei Li writes: Atlas is capable of generating frames with pixel perfect camera control, reconstructing large scenes from as few as one single input image, simulating spacetime by reframing videos, natively outputting three d spaces from one or more input images, composing multiple posed images into a consistent three d world, and more.

这是有史以来最好的相机条件世界模型，为从视觉特效到机器人技术等许多可能的用例打开了大门。

This is the best camera conditioned world model ever, opening doors to many possible use cases from VFX to robotics.

这个你真的得亲眼去看看，但它在真实世界环境中的可控视频，绝对是你前所未见的。

Now this is one that you really have to go see, but it's controllable video in real world environments like nothing you've ever seen.

在解释一个用例示例时，贾斯汀·瑞安写道：Atlas是一个AI模型，可以从最少三台相机重建移动的三维场景。

Explaining an example of a use case, Justin Ryan writes, Atlas is an AI model that can reconstruct moving three d scenes from as few as three cameras.

创作者可以记录一个真实瞬间，然后从从未拍摄过的相机角度观看它。

Creators can record a real moment, then view it from camera angles that were never filmed.

a16z的马丁·卡萨多写道：把它想象成一个具有完全相机控制的视频模型，场景保持近乎三维一致，构建在完全内部的基座模型之上。

A 16 z's Martin Casado writes, think of it as a video model with full camera control, and the scene remains nearly three d consistent, built on a fully internal base model.

从视频编辑到三维构建再到机器人技术，应用场景非常多。

There are many use cases, from video editing to three d construction to robotics.

彼得·杨总结了很多人的感受，他写道：Fable 5.1确实很酷，但这个简直太疯狂了。

Peter Yang summed up the feeling of more than a few when he wrote, yeah, Fable 5.1 is really cool, but this is bonkers.

对于X上的Elvis来说，这个比Fable 5.1更酷的不止一点点。

And for Elvis on X, it's more than just Fable 5.1 that this is cooler than.

他写道：全模态模型是下一个前沿，简单来说，这是我今年见过最令人兴奋的发布。

He writes, omnimodels are the next frontier and simply put, this is the most exciting release I've seen this year.

不过目前对大多数人来说，日常使用中更重要的还是Fable 5.1。

Alas for now, for most people when it comes to our day to day use cases, the bigger topic is indeed Fable 5.1.

那么，我们就结束头条新闻，进入本期正片。

So with that, we will close the headlines and move on to the main episode.

大家好。

Hello, everyone.

AI带来的一个重大变化是，我们的思维从如何给网页排名，转变为如何成为AI信任并用来回答问题的信息来源。

One big change around AI is we've shifted our thinking from how we rank our pages to how do we become the source that AI trusts enough to answer with.

在毕马威，他们亲眼看到了这一点。

At KPMG, they're seeing this firsthand.

AI生成的结果现在直接给出答案，往往连一次点击都不需要。

AI generated results now surface answers directly, often without a single click.

所以他们越来越关注生成式引擎优化，也就是GEO，通过结构化内容，让AI系统能够检索、理解并引用这些内容，作为可信的权威来源。

That's why they are increasingly focused on generative engine optimization or GEO, structuring content so AI systems can retrieve it, understand it, and cite it as trusted authority.

这不仅仅是SEO的演变，更是对可见性的要求。

This is not just an SEO evolution, but a visibility mandate.

而毕马威的GEO要求其实很简单。

And indeed, the GEO mandate from KPMG is simple.

如果AI在影响决策，你的专业见解就必须出现在答案里。

If AI is shaping decisions, your expertise needs to show up inside the answer.

详情请访问kpmg.com/us/geo。

Read all about it at kpmg.com/us/geo.

再说一遍，网址是kpmg.com/us/geo。

Again, that is kpmg.com/us/geo.

这就是为什么大多数传统系统现代化项目会失败。

Here's why most legacy modernization projects fail.

因为做这项工作的AI无法理解大规模的代码库。

The AI doing the work can't understand codebases at scale.

它只能看到一小部分上下文，检查语法，却错过了分布在全球应用生态系统中的多年决策。

It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem.

Blitzy解决这个问题的方式和它解决其他问题的方式一样。

Blitzy solves this the way it solves everything.

在任何迁移开始之前，Blitzy的代理会基于你的代码，将整个遗留系统逆向工程成一个持久的知识图谱。

Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph.

每一个依赖、每一个约束、每一条曾经只存在于某个工程师脑子里的隐性知识。

Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head.

基于这种理解，Blitzy自主执行语言迁移、框架升级以及从单体架构到微服务的转型，并且全程进行端到端验证。

From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith to microservices transformations, all validated end to end.

Blitzy的一个客户使用编码代理，在十六周内完成了价值一千万美元的单体保险系统的现代化改造，而基准时间是137周。

One Blitzy customer modernized a $10,000,000 monolithic insurance stack in sixteen weeks against a one hundred and thirty seven week baseline with coding agents.

这就是9倍的压缩效率。

That's 9x compression.

在加速你的路线图的同时，清除技术债务。

Retire technical debt while accelerating your roadmap.

去blitzy.com看看吧。

See how at blitzy.com.

就是blitzy.com。

That's blitzy.com.

到了这个阶段，问题已经不再是公司是否在积极使用AI了。

At this point, it's no longer a question of whether companies are actively using AI.

但能不能用好，就完全是另一回事了。

Using it well, on the other hand, is a whole different story.

不过，Robots and Pencils这家公司，我觉得就是为这个时代而生的。

Robots and Pencils, though, is a company that I can point to that is actually built for this time.

他们是一家应用AI公司，直接和客户合作，解决真正对业务重要的问题，而不是那些只存在于幻灯片里的实验。

They're an applied AI firm working directly with clients on problems that matter to the business, not experiments that live in a slide deck.

每个项目一开始都是从客户真正需要的结果倒推着来做的。

Every engagement starts by working backwards from the outcome a client actually needs.

如果你想在这个领域里分辨出真正的AI和那些虚头巴脑的东西，这就是关键的区别。

If you're trying to tell real AI apart from noise in this space, that's the difference maker.

去 robotsandpencils.com 看看吧。

Head to robotsandpencils.com.

本期《AI每日简报》由Hyperagent赞助播出，你可以通过它来管理你的智能体团队，让它们协同工作。

This episode of the AI Daily Brief is brought to you by Hyperagent, where you run fleets of agents your team can manage together.

别再想着本地智能体和那些等着你在笔记本电脑上操作的聊天流程了。

Forget local agents and chat workflows waiting on your laptop to be prompted.

Hyperagent 在云端部署了全天候运行的智能体，在你团队已经使用的工具中完成实际工作。

Hyperagent deploys always on agents in the cloud, doing real work across the tools your team already uses.

营销智能体能把竞争对手的动作变成落地页。

Marketing agents turn competitor moves into landing pages.

销售智能体能丰富潜在客户信息、起草邮件，还能更新CRM系统。

Sales agents enrich leads, draft emails, and updates the CRM.

运营代理负责处理文书工作和跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享上下文，并遵循你关于范围和审批的规则。

Every agent has access to shared context and follows your rules about scope and approvals.

是时候让代理像队友一样工作了。

It's time you add agents that feel like teammates.

去Hyperagent雇佣你的代理吧。

Hire yours at Hyperagent.

在hyperagent.com/aidailybrief获得100美元积分。

Get $100 in credits at hyperagent.com/aidailybrief.

欢迎回到AI每日简报。

Welcome back to the AI Daily Brief.

今天是AI每日简报中我最喜欢的日子之一，那就是新模型发布日。

Today is one of my favorite types of days around these parts of the AI Daily Brief, and that is a new model day.

周二，Anthropic发布了Claude Fable 5.1和Mythos 5.1。

On Tuesday, Anthropic released Claude Fable 5.1 and Mythos 5.1.

有趣的不只是能力提升了多少，还有Anthropic在这次发布中选择关注的其他方面。

And what's interesting is not just how the capabilities have improved, but the other aspects that Anthropic chose to focus on with this launch.

不过，我们还是先从能力说起吧。

Still, let's start with the capabilities.

但从现在开始，围绕每一项最先进进展的问题都会是：鉴于我们现有模型已经很强大了，这些能力上的飞跃或者其他新功能，是否值得让我们切换过去？

From here on out though, the question around every single state of the art advance will be given how powerful our existing models are, are the capabilities jumps or some other new feature worth making the switch to?

不过，我们还是先谈谈能力，因为如果能力没有大幅提升，那后面的讨论就没什么意义了。

Still, let's talk about capabilities first because if they aren't a big upgrade, the rest of the conversation is kind of pointless.

简而言之，Fable 5.1毫无疑问是新的最先进水平。

In short, Fable 5.1 is the new state of the art, unambiguously.

5.1在测试智能体编码的Terminal Bench 4.0上得了55.8分，而Mythos 5.1更是达到了60.9%。

5.1 scored 55.8 on Terminal Bench 4.0 which tests agentic coding, and that goes all the way to 60.9% for Mythos 5.1.

这比Fable 552.3%的42%和Opus five的52.3%都要高，也远高于GPT five:six Soul的37.3%。

That's up from 42% for Fable 552.3% for Opus five and way above GPT five:six Soul at 37.3%.

在Cursor Bench 3.2.0上也有类似的提升，Fable 5.1得了73.4%，而Fable five是70.5%。

There was a similar jump on Cursor Bench 3.2.0 with Fable 5.1 scoring 73.4% against Fable five's score of 70.5%.

GPT-five:six:Soul 得了 67.2%，所以差距还是挺大的。

GPT-five:six:Soul scored 67.2%, so again a pretty significant gap.

Fable 5.1 在 GDP Val AA 上也创下了新的最高分，比 Fable five 高出 130 个 Elo 点，比之前最强的 Opus five 高出 29 个点。

Fable 5.1 has also got a new state of the art score on GDP Val AA beating Fable five by 130 Elo points and Opus five which was the previous state of the art by 29 points.

GPT-five:six Soul 本来就已经落后 Fable 12 分了，现在比 Fable 5.1 落后了 140 多分。

GPT-five:six Soul was already 12 points behind Fable and is now over 140 points behind Fable 5.1.

在商业任务方面，Fable 5.1 在 Automation Bench 上得了 31.4%，比 Fable five 的 17.1% 和 GPT-five-six SOL 的 19.6% 有了巨大的提升。

For business tasks, Fable 5.1 scored 31.4% on Automation Bench, which is a huge jump from the 17.1% score that Fable five achieved and 19.6% for GPT-five-six SOL.

计算机使用能力，这显然是 Agenda 能力的关键部分，也提升了，Fable 5.1 明显超过了之前的模型。

Computer use, which is obviously a key part of Agenda capabilities, is also up with Fable 5.1 coming in meaningfully above previous models as well.

不过，从发布来看，Anthropic 显然不仅关注能力的提升，还关注成本的降低。

Still, it's very clear from the announcement that Anthropic was concerned not just with an improvement in capability, but also an improvement in cost.

团队最想在社交媒体上分享的图表，不仅仅是分数对比，还有一张展示分数与每任务平均成本关系的图。

The charts that the team was most keen to share on social media were the charts that not just showed the score comparison, but a graph of score matched against mean cost per task.

在代理科学研究、代理终端编码、多学科推理和更广泛的代理编码方面，Fable 5.1 不仅在从低到高的每个努力级别上得分更高，而且在每个可比级别上，每任务的平均成本也更低。

Across agentic scientific research, agentic terminal coding, multidisciplinary reasoning, and broader agentic coding, not only did Fable 5.1 score higher at each effort level from low to max, but each of their mean cost per task were lower at each comparable level.

换句话说，不管是在低、中还是高强度的设置下使用Fable 5.1，你都能以更低的成本获得比Fable five同等级设置更好的成绩。

In other words, at a low, medium, or high effort setting with Fable 5.1, you are going to get a better score and at a lower cost than the low, medium, or high effort setting on Fable five.

而且博客文章开头就明确强调，价格是重点。

And right up top in the blog post, it is clear that price is a major focus.

Anthropic写道，对于按token计费的典型工作负载，Fable 5.1的预计成本将比Fable five低约25%。

Anthropic writes, Fable 5.1 will cost an estimated 25% less than Fable five for typical workloads wherever usage is built by token.

这是因为我们降低了缓存读取的定价，也就是模型读取已经处理并存储的输入内容。

This is because we're reducing our pricing on cache reads, where the model reads inputs that have already been processed and stored.

对于高度智能体化的工作，节省的成本通常会更多，最高可达约45%。

For highly agentic work, the savings will often be much larger, up to approximately 45%.

换句话说，正如我们一直在讨论的，新模型发布的问题不再仅仅是能力提升，还包括效率提升，你可以看到，即使是像Anthropic这样纯粹的公司也无法回避这个新现实。

In other words, as we have been discussing, the question around new model releases is no longer just about capabilities jumps but also about efficiency increases, and you can see that even a purist company like Anthropic is not immune to that new reality.

不过，公司对自己模型的基准分数和成本做出声明是一回事，当这些模型在实际应用中得到检验时又是另一回事。

Then again, it's one thing for a company to make claims about its benchmark scores and costs for its own models, it's another thing when they get tested in the wild.

所以，就人工智能分析智能指数而言，Fable 5.1无疑在基准测试中名列前茅。

So when it comes to the Artificial Analysis Intelligence Index, Fable 5.1 is undisputedly at the top of the benchmarks.

它的总分从Fable 5的62分跳到了Fable 5.1的66分。

It jumped from an overall score of 62 with Fable five to 66 for Fable 5.1.

这也让它超过了之前的领先者Opus 5，后者是63分。

That puts it ahead of Opus five as well which was the previous leader at 63.

Anthropic现在在指数上占据了前三名，都略微领先于GPT-5 0.6。

Anthropic now has the top three models on the index, all slightly ahead of GPT-five 0.6 sold.

然而，AA发现每个任务的成本高得离谱。

However, AA find that the cost per task was brutal.

事实上，Artificial Analysis发现，即使降低了缓存读取价格，Fable 5.1实际上还是比Fable 5贵了一点。

In fact, artificial analysis found that Fable 5.1 was actually a little more expensive than Fable five even with the cut to cash read pricing.

这个模型每个任务花费3.76美元，而Fable 5是3.14美元。

The model cost $3.76 per task compared to $3.14 for Fable five.

Artificial Analysis认为这主要是由于Fable 5.1的token消耗量大幅增加，在基准测试中多用了70%的token。

Artificial analysis blamed much higher token consumption with Fable 5.1 using 70% more tokens across the benchmark run.

他们承认，缓存成本的降低确实平均每个任务节省了1.40美元，主要集中在代理型基准测试上，但这不足以抵消模型对token的更大胃口。

They acknowledged that the reduction in cash costs did save an average of $1.40 per task largely concentrated in the agentic benchmarks, but that wasn't enough to offset a much more token hungry model.

值得注意的是，用额外高而不是最高设置来测试Fable 5.1，成本降低了28%，而整体性能只下降了一分，总分65。

Notably testing Fable 5.1 on extra high rather than max produced a 28% reduction in cost with only a one point overall drop in performance, scoring a 65 overall.

公共LLM评估平台Vals AI也发现Fable 5.1在其排行榜上领先，并问道：Anthropic是否已经巩固了自己作为前沿领导者的地位？

Public LLM evaluation platform Vals AI also found Fable 5.1 in the lead of its charts and asked: Has Anthropic solidified itself as the frontier leader?

Fable 5.1在VALS指数上首次亮相就排名第一。

Fable 5.1 debuts at number one on the VALS Index.

而Opus five和Fable five紧随其后，Anthropic现在占据了前三名。

And with Opus five and Fable five right behind it, Anthropic now holds the top three spots.

现在ARC Prize发现的结果与Anthropic的承诺更一致一些。

Now the ARC Prize found results a little bit more in line with Anthropic's promises.

Five One在RKGI two上得了90%，在RKGI one上得了97.5%，他们报告说其平均每任务成本比Fable 5低了约32%，这得益于更好的token效率。

Five One scored a 90% on RKGI two and a 97.5% on RKGI one, and they reported that its average cost per task was about 32% lower than Fable 5s driven by better token efficiency.

不幸的是，他们没能得到清晰的RKGI three结果，因为他们写道：请求经常被Anthropic误认为是逆向工程尝试，导致我们无法在发布前完成测试。

Unfortunately, they couldn't really get clear RKGI three results as they write: The requests were frequently misclassified by Anthropic as reverse engineering attempts, preventing us from completing testing before release.

Anthropic非常想强调的另一个基准是Terminal Bench Science Benchmark上代理式科学研究的巨大飞跃，Fable 5.1得分52.6%，几乎是之前Opus five最高29%的两倍。

One other benchmark that Anthropic was very keen to highlight was the big jump on agentic scientific research on the Terminal Bench Science Benchmark, where Fable 5.1 at 52.6% scored almost double the previous high of Opus five at 29%.

最近有很多迹象表明，Anthropic 想在医学、生物学以及更广泛的科学研究领域投入更多时间和精力。考虑到这次把“智能体科学研究基准”放在这么显眼的位置，这似乎又是一个佐证。

There have been a lot of indications recently that Anthropic wants to spend more time and more focus in the areas of medicine and biology and scientific research more broadly, and given the top billing of the Agenic Scientific Research Benchmark, this seems to be more evidence of that.

Anthropic 在发布公告的博客里还重点提到了一个事情，就是 Fable 5 到 5.1 之间，安全护栏有了很大改进。这对生物学和医学问题这类场景有特别的意义，他们声称已经把回退率降低了，我。

Another big thing that Anthropic was pitching in the announcement blog was the fact that the guardrails were much improved between Fable five and five point one, which has specific implications for something like biology and medical question, where they claim that they've reduced the fallback rate, I.

就是说，

E.

模型从 Fable 5.1 切换到 Opus 模型的次数，大概减少了 85%。

The times when the model switches from Fable 5.1 to instead an Opus model, by about 85%.

确实，这次发布整体上很有意思的一点是，重点并不完全放在能力上。

Indeed, what's interesting about the announcement overall is how much the focus is not strictly on the capabilities.

Axios 的资深 AI 记者 Madison Mills 写道：所以现在光发布一个新模型已经不够了。

Axios' senior AI reporter Madison Mills writes: So it's not enough to release a new model anymore.

现在我们看到的是新模型、新安全措施、网络安全和干预方面的变化、成本削减，还有新的企业知识产权保护，全都打包在一次发布里。

Now we're getting new models, new safeguards, cybersecurity and intervention changes, cost cuts, and new enterprise IP protections all in one release.

她指的是，在那篇博客文章的最上面，除了价格之外，Anthropic 还宣传他们有一个新的“企业前沿安全系统”，也就是 EFS，这个系统能让他们为企业提供零数据保留的服务。

And what she's referring to is that right up at the top of that blog post, in addition to price, Anthropic is also selling that they have a new Enterprise Frontier Safeguard System or EFS, which allows them to offer enterprises zero data retention.

使用Fable 5的最大障碍之一就是企业根本不愿意或者没法应对三十天的数据保留政策，这个政策是在Fable 5被美国政府叫停后重新上线时引入的。所以这次升级非常重大，不过EFS系统不会一次性全面推出，Anthropic表示会在今年秋天晚些时候开始分阶段向企业客户提供。

One of the absolute biggest blockers to Fable five usage has been that enterprises simply weren't willing or weren't able to deal with the thirty day retention policies that came when Fable five came back online after being shut down by the US government, and so this is a major, major upgrade, although the EFS system is not going to be rolled out all at once, with Anthropic saying that it will be made available to enterprise customers in phases beginning later this fall.

不过他们表示，在EFS推出之前，符合条件的企业客户可以使用Fable 5.1，并且数据保留为零。

Still, they say until EFS is available, eligible customers will be able to use Fable 5.1 with zero data retention.

最后，除了价格和数据保留之外，他们还强调了安全防护方面的改进，特别是减少了误报的情况，也就是系统把正常内容误判为有害内容的情况。

Finally, in addition to price and data retention, they also pitched this improvement in their safeguards, specifically improvements to reduce false positives where the system flags benign content.

除了我刚才提到的在生物和基础医学问题上误报率降低了85%之外，他们还表示，新的安全防护在网络安全方面也比以前减少了60%的误报。

In addition to the 85% reduction that I just mentioned for biology and basic medicine questions, They also said that the new safeguards block 60% fewer false positives than before in cybersecurity as well.

在网络安全方面，听起来他们在概念上做了重新调整，让Fable 5.1能够用于发现漏洞，但无法利用这些漏洞来开发攻击工具。

Now on the cybersecurity front, it sounds like they've conceptually refocused things, making it so that Fable 5.1 can be used to discover vulnerabilities without being able to develop exploits for them.

那么大家的第一印象如何呢？

So what were people's first impressions?

Anthropic团队很多人还提到的一点是，Fable 5.1听起来不那么像Claude了，我

The one other thing that a lot of folks from the Anthropic team were pitching was Fable 5.1 sounding less Claude like, I.

就是

E.

少了那些AI写作的痕迹，也少了那种让人烦的居高临下的语气，就是最近几版Claude让人不爽的地方。

Having less of the hallmarks of an AI writer and less of some of the patronizing tone that people have been annoyed with with recent iterations of Claude.

Claude Code的创始人Boris Churney发帖说：我们听到了大家的反馈，正在积极减少那种Claude腔。

Claude Code creator Boris Churney posted: We heard your feedback and are actively working on reducing Claude speak.

5.1版本已经取得了不错的进展，后面还有更多改进。

Solid progress with 5.1, more to come.

那Anthropic之外的用户发现了什么呢？

But what did users outside Anthropic find?

Ethan Moloch教授确实觉得整体上这个模型是个显著的进步，不过可能没有我们期望的那么大幅改善那种Claude腔。

Professor Ethan Moloch did find that overall the model is a meaningful advance, although perhaps not as much of an advance on that Claude speak as we might like.

他写道：我提前用上了Claude Fable 5.1。

He wrote: Had early access to Claude Fable 5.1.

在需要判断力和品味的长期工作方面，它确实是个真正的进步，但在减少Claude腔方面进步没那么大。

It's a real advance in long run work that requires judgment and taste, but less of an advance in the Claude ish.

他测试的一部分方式是做了一个复古游戏。

Part of the way he tested it was creating a retro game.

有意思的是，很多人似乎都在用游戏来测试这些东西。

And interestingly, a lot of people seem to be looking for games as the way to test things.

BridgeMind AI 分享了一个马里奥赛车克隆版的视频，说 Fable 5.1 一下子就做出了这个马里奥赛车游戏，这是我目前得到的最好的结果之一，我对它的游戏开发能力印象超级深刻。

BridgeMind AI shared a video of a Mario Kart clone saying Fable 5.11 shotted this Mario Kart game, one of the best results I've had so far and I am super impressed with the game development capabilities.

来自 Claude 的 Alex Albert 展示了他是如何用 Fable 5.1 通过代码生成视频的。

Alex Albert from Claude showed how he used Fable 5.1 to generate videos through code.

对于没在看视频的人来说，这个视频就像你在房产广告里看到的那种虚拟看房演示。

For those of you not watching, the video is a walkthrough of the type that you might see in a real estate listing.

Alex 说：这次我给它一张地块的照片。

Alex says: For this one, I gave it a picture of a property lot.

它根据地块设计了一栋房子，渲染出来，还生成了一个电影般的虚拟看房视频。

It designed a house for the lot, rendered it, and produced a cinematic walkthrough.

Meng2 发现 Fable 5 在处理高级 JavaScript 方面非常出色，特别适合制作更视觉化、更互动的网站。

Meng2 found that Fable five was really good at advanced JavaScript for more visual and interactive sites.

他写道：它更快，能更好地理解复杂的设计指令，并且能精准地还原参考效果。

He wrote, It's faster, understands complex design instructions better, and recreates references with surgical precision.

有了这么强大的功能，就很难满足于静态网站了，尤其是现在很多AI网站看起来都千篇一律。

With this much power, it's hard to settle for static sites especially when so many AI sites look generic.

不过，他也指出，这并不意味着它一下子就完美了。如果你不指定图片，它还是可能生成那种千篇一律的AI插图；在处理像人物和狗这样的三维物体时，它仍然有些困难；你还是需要花心思去调整重叠元素、留白和滚动行为；而且因为它运行得更快，他消耗token的速度也非常非常快。

That said, he did point out that it's not all of a sudden perfect, that it can still create generic AI illustrations if you don't specify the images, that it still has some difficulty with three d subjects like people and dogs, that you still had to deploy taste, fixing overlapping elements, negative space, and scroll behavior, and that because it works faster, he went through tokens very, very quickly.

这种烧token的现象，我们稍后还会再提到。

That token burning effect is something that we'll come back to in just a minute.

在前端设计方面，Mia写道：我让Claude Fable 5.1创建了100个HTML文件。

On front end design, Mia writes: I've asked claud fable five point one to create a 100 HTML files.

规则很简单：看起来惊艳，设计零重复，完全开启创意模式。

The rules were simple: look stunning, zero repeat designs, go full creative mode.

所有100个文件都是通过一个提示词生成的。

All 100 files created in one single prompt.

这是我在这类实验中得到的最好结果，超越了其他任何模型。

These are the best results I've had with this type of experiment beating any other model.

它在前端方面真的很出色，而且几乎没有损坏的文件。

It's really good on the front end and there's almost no broken files.

这真的很厉害。

It's truly impressive.

企业家马修·米勒写道：Fable 5.1 是我用过的最好的模型。

Entrepreneur Matthew Miller wrote: Fable 5.1 is the best model I have ever used.

自从它发布以来，我什么都试过了。

I have thrown everything at it since it dropped.

每一项任务都完成得完美无缺。

Every single task done to perfection.

它的单次生成能力是我从未见过的。

The one shot capabilities are unlike anything I have seen.

你只需问一次，它就能直接搞定。

You ask once and it just delivers.

但真正让我震惊的是它的安全性。

But the thing that actually blew me away is security.

我可以把安全任务交给 Fable 5.1，它不会退缩或拒绝。

I can hand Fable 5.1 secondurity tasks and it does not fall back or refuse.

它发现并修复了我代码库里的漏洞，而Fable 5却拒绝这么做。

It found and patched vulnerabilities in my codebase that Fable five refused to.

这是我见过AI发展最快的一次，而且GPT Astra和Grok 4.7都即将发布。

This is the fastest I have ever felt AI Advance and GPT Astra and Grok 4.7 are both about to release.

世界即将发生巨变。

The world is about to change.

现在每次有新模型出来，你总能指望Every给出最全面的评测之一。

Now every time there's a new model, you can always count on every to have one of the most comprehensive reviews.

这当然就是他们的“Vibe Check”系列，他们对Fable 5.1的总结是：Anthropic又回来了。

This is, of course, their vibe check series, and their summation of Fable 5.1 is Anthropic is so back again.

CEO Dan Shipper写道：这是我们用过的最强的编程模型，但现在它速度快、token效率高，而且关键的是，它说话真的像个正常人。

CEO Dan Shipper wrote: It's the strongest coding model we've used, but now it's fast, token efficient, and crucially actually speaks like a normal person.

Every团队发现它在编程方面是个怪物。

The team at Every found that it was a monster at coding.

Dan说，Kieran Claassen仅凭一个提示词就重建了他们一个产品的可用版本，而且Fable 5.1还添加了他没有要求的有用细节。

Dan said that Kieran Claassen rebuilt a working version of one of their products from one prompt, and Fable 5.1 added useful details that he hadn't requested.

在写作方面，丹说它的文字更清晰，AI痕迹更少，而且接受修改时不会争辩。

On writing, Dan said it had clearer prose, fewer AI tells, and it takes an edit without arguing.

他们发现，在代理任务上，它使用的令牌大约是Opus五的一半，交付时间大约是原来的60%。

They found that on agentic tasks, used about half the tokens as Opus five and delivered things in about 60% of the time.

丹之前说，对Anthropic的主要批评是，他们在数据中心里造了一个几乎没法用的超级天才。

Previously, Dan said, the big knock on Anthropic was that they built a super genius in a data center that was almost unusable.

它太慢了，会顶嘴，还满口技术术语。

It was too slow, argued back, and talked to technical gibberish.

他们用Fable 5.1解决了这些问题，甚至更多。

They managed to solve those problems and more with Fable 5.1.

而比这更重要的是，我觉得丹找到了很多Empower用户可能会采用的使用模式。

And what's even more important than that is that I think that Dan landed on the usage pattern that many Empower users might.

他写道：我日常工作中还是更多用ChatGPT，但在Fable 5.1上消耗的令牌多得多。

He wrote: I still use ChatGPT for work more day to day, but I use way more tokens in Fable 5.1.

我一天开始时就把它派出去做大型编程项目，比如端到端的MVP构建，然后时不时检查一下。

I send it off at the beginning of the day to do big programming projects like end to end MVP builds and check-in every once in a while.

这其实已经有一段时间是高级用户的分工方式了。

This has sort of been a power users division of labor for some time at this point.

GPT-5.0.6模型和编解码器用于交互式任务，就是你和AI一起协作的那种；而Fable模型则用于那些不需要太多交互的长时间运行任务。

The GPT-five 0.6 models in codecs for interactive tasks where you are co working with the AI, and the Fable models for long running tasks that don't require as much interaction.

Prime Intellect的Will Brown也表示同意，说：“天哪，这个模型太厉害了。”

Will Brown from Prime Intellect agreed saying, God, this model is nuts.

他们真的只是让它变得更聪明，编程能力也更强了。

They really just made it smarter and better at coding.

它真的能直接干活。

It can just do things.

他们让它变得合理，不再是垃圾输出。

They made it reasonable and not slop.

这太酷了。

This is so cool.

你可以给它更多的工作，它就直接完成了。

You can give it way more work and it just does it.

代码写得相当不错。

The code is pretty good.

它把重要的东西解释得很清楚，能遵循指令，还能自己发现错误。

It explains the important stuff well, follows instructions, catches its own mistakes.

至少这几周以来，对我来说它是最像AGI的模型。

The most AGI pilling model for me in several weeks at least.

到目前为止，要说有什么批评的话，那绝对就是模型太耗token了，而且很快就撞上订阅的使用上限。

Now to the extent that there are critiques so far, it is absolutely about how token hungry the model can be and how quickly that runs up against subscription usage limits.

Steve Jabs写道：Fable 5.1加上大约12个子代理，在20倍Clawed Max套餐上就用掉了一个小时的使用量。

Steve Jabs writes: Fable 5.1 and about 12 sub agents equals one hour of usage on the 20x clawed max plan.

Chubby写道：简直没法用。

Chubby writes, Literally unusable.

速率限制太离谱了。

The rate limits are absurd.

而且顺便说一句，Fable的自动续写功能有bug，根本用不了。

And oh, by the way, Fable's automatic continuation is bugged and doesn't even work.

伊苏·阿加瓦尔写道，Fable 5.1 根本没法用。

Issu Agarwal writes, Fable 5.1 is unusable.

它太贵了，你几乎用不到三十分钟就没了，而且两周后每周的限额还会更低。

It's so expensive that you can barely get more than thirty minutes of usage out of it, and weekly limits will also be lower in two weeks.

这不是一个适合长时间工作的模型。

This is not a model for extended work.

就连平时不爱夸张的人，比如前投资人杰弗里·伊曼纽尔也写道：Fable 5.1 的使用情况肯定有点不对劲。

Even people not prone to hyperbole like former investor Jeffrey Emanuel wrote: Something definitely seems screwy with the Fable 5.1 usage.

我猜可能是新的 Claude 代码里有个缓存 bug？

Probably a caching bug in the new Claude code if I had to guess?

我今天把我所有 28 个 MAX 20X 账户的额度（至少是五小时的使用限制）全用光了，就只是审计了我的一堆项目。

I managed to blow through all of my 28 MAX 20X accounts today (at least the five hour usage limit) just doing audits of a bunch of my projects.

这可是头一回！

First time ever!

企业家亚当·B。

Entrepreneur Adam B.

莱文深入调查后表示，他可能找到了问题所在。

Levine dug in and suggested that he might have found the problem.

他写道，小提示：Fable 5.1 消耗了大量积分，结果发现它把每个子代理都默认设为 Fable 5.1，完全忽略了我们一贯的相反规则。

Pro tip, he writes, Fable five point one was burning a lot of credits and, turns out, it decided every sub agent should be a Fable five point one ignoring our long standing rule to the contrary.

在另一条推文中，他说，看起来 5.1 在大工作流中特别容易触发，比如用大约 10 个 Fable 5.1 子代理，如果你同时运行多个代理或项目较大，即使 20 倍的限制也会被迅速耗尽。

In another tweet, he said, seems like five point one is super trigger happy with big workflows that use like 10 Fable five point one sub agents that then eat even a 20x limit if you're running more than one agent or it's a bigger project.

基本上，如果你只是用默认设置让它运行，它会用 5.1 来启动子代理去干活，这样很快就会把积分烧光。

Basically, if you just let it go in the default settings, it's going to use 5.1 to spin up the sub agents that it uses to do work, and that could burn through things very quickly.

已经有人开始分享自己的成本优化方法，但我有点觉得 Jan Vailic 说得对，他说订阅制会结束，API 定价在等着我们。

Already people started jumping in with their own cost optimization approaches, but I sort of think that Jan Vailic has it right when he says subscriptions will end, API pricing is awaiting us.

我觉得到这一步，这几乎是不可避免的，但我也认为人们总是这样，在模型刚出来最初几个小时就评判成本，那时候大家还没真正搞清楚怎么用，各种规则也还没定下来，所以如果你实际使用情况比那些反馈要好一些，我也不会感到意外。

I think at this point that is pretty inevitable, but I also think that people always do this thing when they judge costs in the very first hours of even having a model before people have really figured out how to use it and before all the norms settle, so I wouldn't be surprised if your mileage actually goes a bit farther than some of the responses that you're seeing.

但问题是，尤其是如果有严格的使用限制，你很快就会发现自己在用 API 定价，而那相当昂贵。

The question though is, especially if there are strict usage limits you're going to find yourself on API pricing which is pretty expensive soon.

Matt Fee 写道：严肃的问题。

Matt Fee writes: Serious question.

Fable 5.1 能做什么是 Opus、Soul、Kimi、Composer 或 Grok 做不到的？

What can you do with Fable 5.1 that you can't do with Opus, Soul, Kimi, Composer or Grok?

给我讲讲你的实际用例。

Give me your actual use cases.

告诉我我错过了什么。

Tell me what I'm missing.

一方面，我觉得这是人们应该问的正确问题。

On the one hand, I think this is the right type of question for people to be asking.

就像现在几乎每家企业都在努力构建多模型架构，以便把合适的任务和合适的能力水平匹配起来一样，大多数个人也会面临类似的情况——也许他们没有什么自动路由机制，但他们只是理解并设计好了系统，知道针对不同类型的请求该用哪个模型和设置，这样就不会把所有东西都浪费在最先进、最贵的模型和高配置上。

In the same way that pretty much every enterprise right now is trying to figure out a multi model architecture that allows them to connect the right task with the right level of capability, most individuals are going to have something similar, where perhaps they don't have any sort of automated router, but they just understand and have designed systems so that they know which model and setting to use for different types of requests, so that they're not just burning everything on the most state of the art, most expensive model on the highest settings.

同时，有一个想法已经存在一段时间了：现在的模型已经足够好，以至于在很多用例中，没人能真正分辨出它们之间的差别，而如果你还考虑用最贵、最先进的模型，那你就是在自欺欺人，以为它们之间真的有区别。

At the same time, there's this idea that's been around for a while: that the models are so good now that for many use cases no one can really tell the difference between them, and to even consider using the most expensive state of the art models, you must be deluding yourself into thinking that there's actually a difference.

我完全不认同这种说法。

I reject that pretty wholesale.

仅仅因为多个模型都能成功完成一项任务，就认为它们可以互换，这就像说如果两个人能完成同样的工作任务，那么谁来做都无所谓，因为任务完成了。

The idea that just because multiple models can successfully complete a task means that they are all interchangeable with one another is akin to saying that if two people can complete the same work task, it doesn't matter which one does because the task got done.

当然，有些任务确实是这样，而这些任务恰恰是你应该用更便宜的模型来优化的。

Now, certainly there are going to be tasks for which that is the case, and those are precisely the tasks that you should be optimizing using cheaper models for.

但在很多高端重要的工作上，我仍然发现，尽管这些模型都很强大，它们之间还是有巨大的差异。

But when it comes to a lot of high end important work, I still find that as capable as all of these models are, there are still massive differences between them.

我非常提倡的一件事是，为测试新模型准备一套固定的个人基准测试。

One thing I strongly advocate for is to have a standing slate of personal benchmarks for new model testing.

这些基准不必是别人的任务，可以只是对你重要的事情。

They don't have to be anyone else's tasks, they can just be the things that matter to you.

你可能会发现，对于你自己的特定任务，别人抱怨的模型可能表现很好，而别人喜欢的模型可能表现不佳。

And you might find that for your particular tasks, models that other people are complaining about work great and models that other people love don't work so well.

对我来说，我的个人基准测试列表包括几项内容。

For me, that personal benchmark list includes a few things.

基本上就是研究、写作、战略和批判性思维，以及构建，其中构建包括界面设计和架构设计。

It's basically some combination of research, writing, strategic and critical thinking, and building, which includes both an interface design and an architecture component.

你会发现，尤其是在写作或战略思维方面，很多偏好都是主观的。

And what you'll notice is that especially when it comes to something like writing or strategic thinking, a lot of preference is going to be subjective.

换句话说，Anthropic 没法给我展示一个基准测试，来衡量我如何把 NLW 那些疯狂的新商业想法一步步落实。

In other words, Anthropic can't show me some benchmark for iterating on NLW's mad ideas for new businesses.

这得看 Fable 5.0.1 Opus 和 Soul 在实际操作中表现如何，我得亲自试试。

That's something that I have to see how Fable 5.0.1 Opus versus Soul handle in practice.

而且即使在这个模型普遍都很强的时代，我仍然发现这类事情上差异巨大。

And even in this era of generally capable models, I still find massive differences in things like that.

这里要提醒大家的是，对我们所有人来说，当新模型出来时，问题不再是“我该不该换到这个模型？”

The reminder here is that for all of us, the question when a new model comes out is no longer Should I switch to that model?

而是“这个模型如何融入我个人的模型架构？”

Instead, it's How does that model fit into my personal model architecture?

用在哪些场景？

For what uses?

这个模型是不是更好，值不值得它带来的金钱或其他方面的成本？

Is that model better and worth whatever financial or other types of costs that come with it?

换句话说，最优秀的用户会想办法从新模型中获得最大价值，而不是抱着“只能选一个”的旧观念，从一个模型随便跳到另一个。

The best users in other words are going to figure out how to get the most out of new models rather than just clunking around from one to the next with some old idea that you have to pick just one.

不过最后再补充一点，如果你经济上没办法轻松地在不同模型之间切换，那很多考虑就不一样了。

Now for one last qualification on that, I will note that if you are not in a financial position where you can be blithely shifting between models, a lot of these considerations get different.

在这种情况下，说它们都挺全能的是没错的。

And for that, the advice that they are all pretty generally capable is accurate.

确实，现在锁定一个生态系统的时机比以往任何时候都好，因为它们各自都很强，虽然各有取舍。

It certainly is the case that it has never been a better time to be locked into just one ecosystem because they are all so individually capable even if they do have different trade offs.

不过，现在有趣的部分来了，你可以去亲自测试和体验这些东西。

Still, now the fun part begins where you get to go test and try these things.

我很期待这个劳动节周末花点时间试试看，当然，下周我会汇报结果。

I'm excited to spend some time this Labor Day weekend testing things out, and I will, of course, report back next week.

那今天就到这里，这就是今天的AI每日简报。

For now, though, that is gonna do it for today's AI daily brief.

一如既往，感谢你的收听或观看。

Appreciate you listening or watching as always.

下次见。

Until next time.

再见。

Peace.