Why Fable 5.1 Is Worth the Upgrade — 文稿

M10:00

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，它们毫无疑问是最先进的，几乎在所有类别上都超越了其他所有模型。

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突破，所以对于很多任务，包括长时间运行的 agent 任务来说，Fable 5.1 的花费应该会比 Fable 5 的同类任务少 25%，甚至多达 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 5.

初步反馈相当不错，不过用户在实际花费和因使用限制能用 Fable 5.1 做多少事上，体验差别很大。

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 can even get with Fable 5.1 given usage limits.

不过，现在每个新模型出来都会有人问：这个模型够好到值得切换过去吗？

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?

但我认为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，问题应该是：我能用这个模型做什么？

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

它怎么融入我的整个模型栈？

How does it fit into my overall model stack?

在认清它带来什么取舍的同时，我怎样才能最大程度地利用它？

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

这就是我们今天这期节目要讨论的内容，咱们开始吧。

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

《AI Daily Brief》是一档每日播客和视频节目，聚焦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，在开始之前先快速说几个通知。

Alright friends, quick announcements before we dive in.

M11:41

今天有几条比较重磅的新闻。

We have kind of a dramatic set of headlines today.

第一条是关于 OpenAI 即将推出的 Astra。

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

在周二的一篇博文中，OpenAI 表示，他们现在认为 Astra 达到了其准备框架中定义的网络安全关键能力阈值。

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

通俗点说，就是该模型能够在没有人类指导的情况下，自主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 8 月初的评估中，OpenAI 认为 Astra 有可能达到这个阈值，但当时还不能确定。

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.

本质上，这与 Anthropic 今年早些时候严格限制 Mythos 时的担忧是一样的。

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

在分享评估 Astra 能力的一些细节时，OpenAI 透露该模型在 Exploit Bench 上取得了 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.

这个基准测试评估的是模型根据已知漏洞开发利用代码的能力。

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

然后 OpenAI 更进一步，开发了他们自己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.

这样一来，就可以用训练数据里不可能有的测试，来检验模型是否真的有能力从头开始创造出全新的利用代码。

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 实现了比 GPT-5.6 Sol 高得多的任意代码执行率，而且使用的输出 token 少得多。

OpenAI wrote, On this dataset, Astra achieves much higher arbitrary code execution rates than GPT-5.6 Sol, using far fewer output tokens.

在评估过程中，该模型甚至发现并利用了 2 个零日漏洞，作为漏洞利用链的一部分。

During the evaluation, the model even discovered and used 2 zero-day vulnerabilities as part of an exploit chain.

为了让这个对比更直观一点，Astra 在内部版 Exploit Bench 上花了 4 万个 token 就拿到了 30% 的分数，而 GPT-5.6 Sol 大概要花 11 万个 token 才能取得任何显著成果。

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 GPT-5.6 Sol, which wasn't capable of any significant results until it spent around 110,000 tokens.

但如果把这一点推广到其他能力上，这可能意味着该模型在运行各类 agent 时，token 效率都更高。

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

在与专家合作伙伴进行的进一步测试中，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 root 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 将在发布 Astra 时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.

该模型本身已经接受了额外的训练，以拒绝网络安全相关任务，拒绝率从 GPT-5.6 Sol 的 59% 提高到了 91.5%。

The model itself has received additional training to refuse cybersecurity tasks, resulting in a 91.5% refusal rate, up from 59% for GPT-5.6 Sol.

OpenAI 还增加了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，并对这些账户应用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链式思维监控，以便及早检测并阻止错误行为。

OpenAI says that they believe that 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.

在X上发布的一篇异常严肃的帖子中，Sam Altman甚至用了正确的语法和标点，他补充说，这里存在明显的紧张关系。

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用它来构建什么。

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.

另一方面，我们显然正处于一个开发阶段，我们认为谨慎是必要的，并且我们正在调整进展速度，以确保能够满足新能力水平所需的安全标准。

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向前迈出了重要一步。

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，我们根据需要放慢了速度，以确保能够做好安全和对齐方面的工作。

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

在暗示OpenAI内部的氛围时，他继续写道，我们有一段时间一直生活在既兴奋又焦虑于进步的紧张状态中，这仍然让我们感到不和谐。

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.

我们知道这对其他人来说更加不和谐。

We know it is much more discordant for other people.

但我们仍然坚信，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。

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.

一个社会与这项技术共同发展的迭代循环，将最大程度地提高我们做对这件事的可能性。

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

所以，我们希望你们喜欢我们的新模型，也希望世界继续极其认真地对待AI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.

现在，有消息称Astra可能最快本周就会发布，这时间点太完美了，因为我正在旅行，理论上应该做预录节目，但信不信由你，这并不是关于Astra的唯一讨论。

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

周二深夜，《信息》杂志独家爆料了一项技术突破，使Astra在推理方面变得更好，据某些人说，可能也更危险。

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Recurrent Depth，它使用了一个循环Transformer。

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文本字符串，以改进其响应。

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

虽然消息人士称该技术提高了性能并降低了成本，但最大的缺点是缺乏可观察性。

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

部分推理过程现在发生在模型内部，而不产生输出。

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

这意味着链式思维将部分被遮蔽，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消息人士表示，他们在Astra中有限地使用了这项技术，以确保推理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.

然而，《信息》杂志写道，AI研究人员担心，一些AI开发者如果为自己的模型采用同样的技术，可能不会施加OpenAI所施加的那种限制，而且不受限制地使用该技术可能会导致AI失控，其行为难以监督。

However, writes The Information, AI researchers, quote, 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安全工作的人士已经对智能体可观察性变得更难感到担忧。

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

在他们对Hugging Face攻击事件的分析中，Meeter指出，日志对人类来说无法拼凑起来，需要AI分析才能了解全貌。

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

Dwarkesh Patel在本周早些时候对那次攻击进行了戏剧性且有争议的复述，他评论道，我不认为这是我们会收到的最后警告，但这可能是我个人能理解的最后一个了。

Dwarkesh 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.

报道发布后，Encode AI的Nathan Calvin发帖称，《信息》今晚报道了一个非常重大且极其令人担忧的故事。

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中利用了一项neural ease突破，这可能会破坏链式思维的可监控性。

Looks like OpenAI utilized a breakthrough in neural ease for Astra that could destroy chain-of-thought monitorability.

如果OpenAI发现了这种架构并获得了性能或效率提升，其他公司很可能很快就会找到它（如果还没有的话），并且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.

如果有些人这样做，可能很难避免逐底竞争。

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

Redwood Research的Ryan Greenblatt，他是Meeter调查Hugging Face事件的主要研究人员之一，写道，我最担心的是，从这里自然的发展方向可能是将不透明的推理扩展到模型完全或几乎完全在潜在空间中进行推理的程度。

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.

这很可能就毁了思维链在监控和监管上的作用。

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

如果这属实，那OpenAI简直是在触碰AI行业仅有的几根红线之一。

If this is true, OpenAI seems to be violating one of the few red lines that exists in the AI industry.

绝对不能这么训练模型。

Absolutely do not train your models like this.

这是在搞什么？

What is going on?

不过还是有不少人跳出来，试图让情绪稍微冷静一点。

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

The Information的Amir Efrati又出来说话，强调他们报道的意思是OpenAI确实在给这个技术设限，努力让思维链保持可见，但他们担心其他AI开发者可能不会这么做。

Amir Efrati 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首席科学家Jacob Pachocki写道，我不想因为混乱的报道引发一场让模型变得无法监控的军备竞赛。

And OpenAI Chief Scientist Jacob Pachocki wrote, I want to prevent a race into unmonitorability kicked off by confused reporting.

包括Astra在内，我们现在前沿模型的计算图深度与GPT-4相比，差距在两倍以内。

The depth of the computation graph for our present frontier models, including Astra, is within a factor of 2 of GPT-4.

从我们最早期的推理模型开始，OpenAI就一直在努力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.

我们非常看重这个技术，因为它能让我们看到模型的对齐能力如何从训练数据中泛化出来。

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

我觉得它确实很脆弱，而且不幸的是，趋势在往坏的方向走，原因跟架构变化无关，我很快会写文章详细说。

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.

但我们有办法加强它，这也是我们当前研究项目的核心目标之一。

But there are things we can do to strengthen it, and it's a core goal of our current research program.

所以呢，又一个没什么争议的发布要来了。

So, you know, another uncontroversial release coming up.

说到发布，华尔街日报报道说Gemini 3.8 Flash马上就要来了，可能会解决谷歌AI一个长期存在的痛点，也就是编程能力。

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

不管怎么努力，谷歌至今没做出过顶级的编程模型，而且在这个关键能力上已经远远落后了。

Now, 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.

但据华尔街日报说，在公司内部测试中，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 Anthropic's Opus.

这个模型预计这周发布，可能今天就会出来，所以我们很快就能知道它是不是真像说的那么好。不过文章还提到了Gemini Pro系列背后的一些事。

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.

有消息源说，所有原本打算以3.5 Pro名义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很满意。

However, researchers are pleased with the performance of Gemini 4 during pre-training evals.

这个模型目前还在后训练阶段，也就是说距离准备好还要些时间。但对于那些希望看到更多竞争而不是只有OpenAI和Anthropic的人来说，这或许是个好消息。

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.

有一个模型，如果不是因为我们今天主要聊Fable 5.1的话，它完全能成为今天的头条话题，那就是WorldLab新发布的模型Atlas。

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 WorldLab's newly released model, Atlas.

他们称它是世界上第一个多模态世界模型，能生成像素级完美控制摄像头的图像和视频帧，还能在3D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 3D.

建模世界，移动摄像头，模拟时空。

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

WorldLabs的Ben Mildenhall写道，Atlas是一个从头为下一帧预测任务构建的自回归扩散模型。

WorldLabs' Ben Mildenhall writes, Atlas is an autoregression diffusion model built from the ground up for the task of next-frame prediction.

它同时也是一个世界级的摄像头控制视频生成、新视角合成和稀疏3D重建方法。

It is simultaneously a world-class method for camera-controlled video generation, novel view synthesis, and sparse 3D reconstruction.

WorldLabs联合创始人李飞飞写道，Atlas能生成像素级完美控制摄像头的帧，仅凭一张输入图像就能重建大场景，通过重新构图视频来模拟时空，从一个或多个输入图像原生输出3D空间，将多个带姿态的图像合成为一个一致的3D世界，等等。

WorldLabs 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 3D spaces from one or more input images, composing multiple posed images into a consistent 3D world, and more.

这是有史以来最好的摄像头条件世界模型，为从视觉特效到机器人技术等众多可能的应用场景打开了大门。

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

这个模型你真的得亲眼去看看，它能在真实环境中生成可控视频，效果前所未有。

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.

举个例子，Justin Ryan写道，Atlas是一个AI模型，只需3个摄像头就能重建移动的3D场景。

Explaining an example of a use case, Justin Ryan writes, Atlas is an AI model that can reconstruct moving 3D scenes from as few as 3 cameras.

创作者可以录制一个真实瞬间，然后从从未拍摄过的摄像头角度来观看它。

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

A16Z 的 Martin Casado 写道，可以把它看作一个拥有完整摄像机控制的视频模型，场景几乎保持3D一致性，基于完全自研的基础模型构建。

A16Z's Martin Casado writes, think of it as a video model with full camera control, and the scene remains nearly 3D consistent, built on a fully internal base model.

从视频编辑到3D构建再到机器人技术，有很多用例。

There are many use cases, from video editing to 3D construction to robotics.

Peter Yang 概括了不少人的感受，他写道，是啊，Fable 5.1 确实很酷，但这简直疯狂。

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.

而对于 X 平台上的 Elvis 来说，这比 Fable 5.1 更酷的东西还多着呢。

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

他写道，全能模型是下一个前沿，简单说，这是我今年看到的最激动人心的发布。

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

可惜，目前对大多数人来说，在咱们的日常使用场景里，更大的话题确实是 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.

好了，到此我们结束头条新闻，进入本期主节目。

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

大家好。

Hello everyone.

AI 带来的一个大变化是。

One big change around AI.

我们不再只想着如何给页面排名，而是思考如何成为 AI 足够信任、愿意用来回答问题的来源。

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 KPMG，他们亲眼见证了这一点。

At KPMG, they're seeing this firsthand.

AI 生成的结果现在直接呈现答案，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——把内容结构化，让 AI 系统能检索、理解并作为可信权威去引用。

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.

这不只是 SEO 的演变，更是可见性的必然要求。

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

而 KPMG 的 GEO 指令很简单。

And indeed, the GEO mandate from KPMG is simple.

如果 AI 在塑造决策，你的专业知识就需要出现在答案里。

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

M114:10

欢迎回到 AI 每日简报。

Welcome back to the AI Daily Brief.

今天是我在 AI 每日简报最喜欢的那种日子之一，就是新模型发布日。

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

周二，Anthropic 发布了 Claude Sonnet 5.1 和 Opus 5.1。

On Tuesday, Anthropic released Claude Sonnet 5.1 and Opus 5.1.

有意思的不只是能力提升了多少，还有 Anthropic 在这次发布中重点关注的其它方面。

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，咱们还是先说说能力吧。

Still, let's start with the capabilities.

但从今以后，每个最前沿的进步都会面临一个问题：考虑到现有模型已经这么强大了，能力飞跃或其他新功能到底值不值得换用？

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?

不过还是先聊能力，因为如果不是重大升级，其他讨论就没啥意义了。

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

简单说，Sonnet 5.1 无可争议地成为了新的最佳模型。

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

Sonnet 5.1 在 SWE-bench 4.0 上得了 55.8 分，这个平台测试的是智能体编程能力，而 Opus 5.1 更是达到了 60.9%。

Sonnet 5.1 scored 55.8 on SWE-bench 4.0, which tests agentic coding, and that goes all the way to 60.9% for Opus 5.1.

这比 Fable 5 的 42% 和 Opus 5 的 52.3% 都有提升，也远高于 GPT-5.6 Sol 的 37.3%。

That's up from 42% for Fable 5 and 52.3% for Opus 5, And way above GPT-5.6 Sol at 37.3%.

在 CursorBench 3.2.0 上也有类似飞跃，Fable 5.1 得了 73.4%，而 Fable 5 是 70.5%。

There was a similar jump on CursorBench 3.2.0, with Fable 5.1 scoring 73.4% against Fable 5's score of 70.5%.

GPT-5.6 Sol 得了 67.2%，所以差距还是挺大的。

GPT-5.6 Sol scored 67.2%, so again, a pretty significant gap.

Fable 5.1 在 GDPVal-AA 上也取得了新的最佳成绩，比 Fable 5 高了 130 Elo 分，比之前的冠军 Opus 5 高了 29 分。

Fable 5.1 has also got a new state-of-the-art score on GDPVal-AA, beating Fable 5 by 130 Elo points, and Opus 5, which was the previous state-of-the-art, by 29 points.

GPT-5.6 Sol 本来就已经比 Fable 落后 12 分，现在比 Fable 5.1 落后了 140 多分。

GPT-5.6 Sol was already 12 points behind Fable and is now over 140 points behind Fable 5.1.

对于商业任务，Fable 5.1在自动化基准测试中得分为31.4%，相比Fable 5的17.1%和GPT-5.6 Sol的19.6%有巨大提升。

For business tasks, Fable 5.1 scored 31.4% on Automation Bench, which is a huge jump from the 17.1% score that Fable 5 achieved and 19.6% for GPT-5.6 Sol.

计算机使用能力，这显然是代理能力的关键部分，也有所提升，Fable 5.1的表现明显优于之前的模型。

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

不过，从公告中可以明显看出，Anthropic关注的不仅仅是能力的提升，还有成本的降低。

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.

团队最热衷于在社交媒体上分享的图表不仅是分数对比，还包括分数与每任务平均成本的对比图。

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、多学科推理以及更广泛的代理编码领域，Fable 5.1不仅在每个努力级别（从低到最高）都取得了更高的分数，而且在每个可比级别上，每任务平均成本也更低。

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 costs per task were lower at each comparable level.

换句话说，使用Fable 5.1的低、中、高努力设置，你会比使用Fable 5的低、中、高努力设置获得更好的分数，同时成本更低。

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 5.

而且博客文章的开头就明确提到，价格是一个重点。

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的估计成本将比Fable 5低25%左右。

Anthropic writes, Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads wherever usage is billed by token.

这是因为我们降低了缓存读取的价格，即模型读取已经处理并存储的输入。

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

对于高度代理性的工作，节省的成本通常会更大，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，关于新模型发布的问题不再仅仅是能力的飞跃，还包括效率的提升。

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.

而且你可以看到，即使是像Anthropic这样纯粹的公司也无法忽视这一新现实。

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。

Then again, it's one thing for a company to make claims about its benchmark scores and costs for its own models.

在实际测试中又是另一回事。

It's another thing when they get tested in the wild.

所以在Artificial Analysis Intelligence Index上，Fable 5.1毫无疑问位居基准测试榜首。

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分跃升至Fable 5.1的66分。

It jumped from an overall score of 62 with Fable 5 to 66 for Fable 5.1.

这使其超过了之前的领先者Opus 5（63分）。

That puts it ahead of Opus 5 as well, which was the previous leader at 63.

Anthropic现在在该指数中拥有前三名模型，均略微领先于GPT-5.6 Sol。

Anthropic now has the top 3 models on the index, all slightly ahead of GPT-5.6 Sol.

然而，AA发现每任务成本相当高。

However, AA found 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实际上比Fable 5稍微贵一点。

In fact, Artificial Analysis found that Fable 5.1 was actually a little more expensive than Fable 5, even with the cut to cache read pricing.

模型每任务成本为3.76美元，而Fable 5为3.14美元。

The model cost $3.76 per task, compared to $3.14 for Fable 5.

Artificial Analysis将原因归咎于更高的token消耗，Fable 5.1在基准测试运行中使用的token多出70%。

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

他们承认，缓存成本的降低平均每任务节省了1.40美元，主要集中在代理基准测试中，但这不足以抵消模型对token的更贪婪需求。

They acknowledged that the reduction in cache 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，在超高级别而不是最高级别测试Fable 5.1，成本降低了28%，性能整体只下降了1分，总分为65分。

Notably, testing Fable 5.1 on extra high rather than max produced a 28% reduction in cost with only a 1-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 Index上首次亮相就排名第一，紧随其后的是Opus 5和Fable 5，Anthropic现在占据了前三名。

Fable 5.1 debuts at number 1 on the Vals Index, and with Opus 5 and Fable 5 right behind it, Anthropic now holds the top 3 spots.

而ARC Prize的结果与Anthropic的承诺更为一致。

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

Fable 5.1在ArcAGI-2上得分90%，在ArcAGI-1上得分97.5%，他们报告说，由于更好的token效率，其每任务平均成本比Fable 5低约32%。

5.1 Scored a 90% on ArcAGI-2 and a 97.5% on ArcAGI-1, and they reported that its average cost per task was about 32% lower than Fable-5's, driven by better token efficiency.

不幸的是，他们未能获得清晰的ArcAGI-3结果。

Unfortunately, they couldn't really get clear ArcAGI-3 results.

正如他们所写，请求经常被Anthropic误分类为逆向工程尝试，导致我们无法在发布前完成测试。

As they write, the requests were frequently misclassified by Anthropic as reverse engineering attempts, preventing us from completing testing before release.

Anthropic 特别想强调的另一个基准测试是，在 Terminal Bench Science 这个自主科研基准上，Fable-5 得分 52.6%，几乎是之前 Opus-5 最高分 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 won at 52.6%, scored almost double the previous high of Opus-5 at 29%.

最近有很多迹象表明，Anthropic 想把更多时间和精力投入到医学、生物学和更广泛的科学研究领域，而自主科研基准测试被放在最突出的位置，似乎进一步佐证了这一点。

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 agentic scientific research benchmark, this seems to be more evidence of that.

Anthropic 在发布博客中宣传的另一件大事是，Fable-5 和 5.1 之间的护栏有了很大改进，这对生物学和医学类问题有特殊影响。他们声称，模型从 Fable-5.1 切换到 Opus 模型的回退率降低了大约 85%。

Another big thing that Anthropic was pitching in the announcement blog was the fact that the guardrails were much improved between Fable-5 and 5.1, which has specific implications for something like biology and medical questions, where they claim that they've reduced the fallback rate, i.e., the times when the model switches from Fable-5.1 to instead an Opus model, by about 85%.

整篇公告有意思的地方在于，重点并不完全放在功能上。

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

Axios 高级 AI 记者 Madison Mills 写道，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.

我们现在一次发布就推出了新模型、新安全措施、网络安全和干预措施的变化、成本削减以及新的企业 IP 保护。

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，在那篇博客的顶部，除了价格之外，Anthropic 还宣传他们有一套新的 Enterprise Frontier Safeguard 系统，简称 EFS，可以让企业实现零数据保留。

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.

Fable 5 使用面临的最大障碍之一就是，企业根本不愿意或无法应对 30 天的数据保留政策，这个政策是 Fable 5 被美国政府叫停后重新上线时带来的。所以这绝对是一次重大升级，不过 EFS 系统不会一次性全部推出，Anthropic 表示将从今年秋末开始分阶段向企业客户提供。

One of the absolute biggest blockers to Fable 5 usage has been that enterprises simply weren't willing or weren't able to deal with the 30-day retention policies that came when Fable 5 came back online 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 可用之前，符合条件的客户可以使用 Fable 5.1，并且实现零数据保留。

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，他们还宣传了安全措施的改进，特别是减少了系统将良性内容误标记为有问题的误报情况。

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

除了我刚才提到的生物学和基础医学问题上 85% 的误报减少，他们还表示，新的安全措施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，听起来他们从概念上重新调整了重点，让 Fable 5.1 能够用于发现漏洞，但无法开发针对这些漏洞的利用程序。

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怎么样？

So what were people's first impressions?

Anthropic 团队很多人还在宣传的是，Fable 5.1 听起来不那么像 Claude 了。

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

也就是说，减少了 AI 写作的那些标志性特征，也少了一些让人厌烦的居高临下的语气，这些是最近几版 Claude 让用户不满的地方。

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

Claude code 的创建者 Boris Cherny 发帖说，我们听到了大家的反馈，正在积极减少那种 Claude 式的说话方式。

Claude code creator Boris Cherny posted, we heard your feedback and are actively working on reducing Claude speak.

5.1 有了扎实的进步，后面还有更多改进。

Solid progress with 5.1, more to come.

但 Anthropic 之外的用户发现了什么呢？

But what did users outside Anthropic find?

Ethan Mollick 教授发现，总体而言，这个模型是一个有意义的进步，但可能在减少 Claude 说话风格方面没有我们期望的那么大。

Professor Ethan Mollick 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，我提前体验了 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中，它确实是一个真正的进步，但在减少 Claude 风格方面进步没那么大。

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

他测试的方式之一就是制作一个复古游戏。

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作为测试方式。

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

Bridgemind AI 分享了一个马里奥卡丁车克隆版的视频，说 Fable 5.1 一次就生成了这个马里奥卡丁车游戏。

Bridgemind AI shared a video of a Mario Kart clone, saying Fable 5.1 one-shotted this Mario Kart game.

这是我目前看到的最好结果之一，我对它在游戏开发方面的能力印象非常深刻。

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 Claude 5.1 通过代码生成视频的。

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

对于没在看视频的人，我说明一下：这个视频是一种漫游式导览，类似于你在房产广告里看到的那种。

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，这次我给了它一块地的图片。

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

它在这块地上设计了一栋房子，进行了渲染，然后生成了一个电影式的漫游导览。

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

Meng Tu 发现，Claude 5 在处理构建可视化、交互式网站的高级 JavaScript 方面非常出色。

Meng Tu found that Claude 5 was really good at advanced JavaScript for more visual and interactive sites.

他写道，它更快了，能更好地理解复杂的设计指令，并且能以手术刀般的精确度还原参考图。

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

有了这么强的能力，就很难再满足于静态网站了，尤其是现在这么多AI网站看起来都很千篇一律。

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，它并不是突然就完美了。如果你不指定具体的图片，它仍然会生成那种千篇一律的AI插图；在处理像人物和狗这样的3D对象时还是有些困难；你仍然需要运用审美，去修复重叠元素、负空间和滚动行为；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 3D 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的问题，我们马上会再谈到。

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

关于前端设计，Mia写道：“我让Claude 5.1创建了100个HTML文件。”

On front-end design, Mia writes, I've asked Claude 5.1 to create 100 HTML files.

规则很简单：看起来惊艳，设计零重复，全力发挥创意模式。

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

全部100个文件都在一个提示词里生成。

All 100 files created in one single prompt.

这是我在这类实验中得到的最好结果，打败了其他所有模型。

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好，而且几乎没有任何损坏的文件。

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

真的令人印象深刻。

It's truly impressive.

企业家Matthew Miller写道：“Claude 5.1是我用过的最好的模型。”

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

自从它发布以来，我什么任务都扔给它做。

I have thrown everything at it since it dropped.

每一个任务都完美完成。

Every single task done to perfection.

一次性完成任务的能力是我前所未见的。

The one-shot capabilities are unlike anything I have seen.

你只需要问一次，它就能直接给出结果。

You ask once and it just delivers.

但真正让我震惊的是安全方面。

But the thing that actually blew me away is security.

我可以把安全任务交给Claude 5.1，它不会退缩或拒绝。

I can hand Claude 5.1 security tasks and it does not fall back or refuse.

它发现并修复了我代码库中Claude 5拒绝修复的漏洞。

It found and patched vulnerabilities in my codebase that Claude 5 refused to.

这是我觉得AI进步最快的一次，而且GPT-Astra和Grok 4.7都快要发布了。

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

世界即将改变。

The world is about to change.

现在，每次有新模型发布，你总可以指望Every提供最全面的评测之一。

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

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

This is, of course, their Vibe Check series, and their summation of Claude 5.1 is Anthropic is so back again.

首席执行官Dan Shipper写道：“这是我们用过的最强的编程模型，而且现在它速度快、token效率高，最关键的是，它说话真的像普通人一样。”

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的团队发现它在编程方面是个怪兽。

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

Dan说，Kieran Kloss仅用一个提示词就重建了他们一个产品的可工作版本，而且Claude 5.1还添加了一些他并未要求的实用细节。

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

在写作方面，Dan说它的散文更清晰，AI的痕迹更少，而且能接受修改而不反驳。

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token大约是Opus 5的一半，完成速度大约是60%的时间。

They found that on agentic tasks, it used about half the tokens as Opus 5, and delivers things in about 60% of the time.

Dan说，之前外界对Anthropic的主要批评是，他们在数据中心建造了一个天才般的模型，但却几乎无法使用。

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，会反驳，而且说话像技术胡言乱语。

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

他们通过Fable 5.1成功解决了这些问题，甚至更多。

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

而且更重要的是，我觉得Dan找到了很多重度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 power users might.

他写道，我日常工作中用得更多的还是ChatGPT，但在Fable 5.1上消耗的token却多得多。

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

我每天一开始就让它去跑大的编程项目，比如端到端的MVP构建，然后偶尔检查一下进度。

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.

这某种程度上已经是重度用户一段时间以来的分工方式了。

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

GPT-5/6模型和Codex用于需要与AI协作的交互式任务，而Fable模型则用于不需要太多交互的长时间运行任务。

The GPT-5/6 models and Codex for interactive tasks where you are coworking with the AI, and the Fable models for long-running tasks that don't require as much interaction.

Prime Intellect的Will Brown也表示赞同，他说，天啊，这个模型太疯狂了。

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

他们真的只是让它更聪明，编程能力更强了。

They really just made it smarter and better at coding.

它直接就能做事。

It can just do things.

他们让它变得合理，不再是一堆垃圾。

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，它直接就能干完。

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

代码质量相当不错。

The code is pretty good.

它能很好地解释重要内容，遵循指令，还能自己发现错误。

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

至少在这几周里，它是我用过的最有AGI感觉的模型。

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

目前要说到批评意见的话，绝对是在于这个模型对token的消耗量巨大，而且很快就会触及订阅的使用上限。

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Claude Max套餐里能消耗1小时的使用量。

Steve Jabs writes, Fable 5.1 and about 12 subagents equals 1 hour of usage on the 20x Claude Max plan.

Chubby写道，根本没法用。

Chubby writes, literally unusable.

速率限制太离谱了。

The rate limits are absurd.

哦对了，Fable的自动续写有bug，根本用不了。

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

Fable 5.1根本没法用。

Fable 5.1 is unusable.

它太贵了，你几乎用不了超过30分钟。

It's so expensive that you can barely get more than 30 minutes of usage out of it.

而且两周后周上限还会更低。

And weekly limits will also be lower in 2 weeks.

这不是一个适合长时间工作的模型。

This is not a model for extended work.

即使是不喜欢夸张的人，比如前投资人Jeffrey Emanuel也写道，Fable 5.1的使用量肯定有问题。

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

我猜可能是新版Claude代码里的缓存bug。

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

我今天竟然用完了所有28个最高20倍账户的额度，至少是5小时使用限制，只是审计了我的一批项目。

I managed to blow through all of my 28 max 20x accounts today, at least the 5-hour usage limit, just doing audits of a bunch of my projects.

有史以来第一次。

First time ever.

企业家Adam B. Levine深入调查后表示，他可能找到了问题所在。

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

他写道，一个小贴士：Fable 5.1消耗了大量额度，结果发现它决定每个子代理都应该用Fable 5.1，无视我们长期以来相反的规则。

Pro tip, he writes, Fable 5.1 was burning a lot of credits, and turns out, it decided every sub-agent should be a Fable 5.1, ignoring our longstanding rule to the contrary.

在另一条推文中，他说，看起来5.1在大工作流中非常容易触发，比如使用10个Fable 5.1子代理，如果你运行多个代理或项目较大，即使20倍上限也会被吃掉。

In another tweet, he said, seems like 5.1 is super trigger-happy with big workflows that use like 10 Fable 5.1 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跑，它会用5.1的方式来创建那些用来干活的小代理，那样的话资源很快就会烧光。

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

已经有人开始搞自己的成本优化方案了，但我倒是觉得Jan Valek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 Valek has it right when he says, subscriptions will end, API pricing is awaiting us.

我觉得到这个地步，那基本上是不可避免的，不过我也觉得大家老爱干一件事，就是在刚拿到模型那会儿就急着算成本，还没真的搞明白怎么用，也没等所有规范稳定下来，所以我不会惊讶如果你实际用起来比网上那些反馈要更省一些。

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，特别是如果有严格的使用限制，那你很快就会发现自己在用API定价，那个还挺贵的。

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

说正经的，你能用Fable 5.1做到什么，是Opus、Sol、KIMI、Composer或者Grok做不到的？

Serious question, what can you do with Fable 5.1 that you can't do with Opus, Sol, KIMI, Composer, or Grok?

给我说说你的实际用例。

Give me your actual use cases.

告诉我我错过了什么。

Tell me what I'm missing.

一方面，我觉得大家问这个问题挺对的。

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

就像现在几乎所有企业都在搞多模型架构，想把合适的任务对应到合适的能力水平，大多数个人也会面临类似的情况。

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.

或者说，他们可能没有一个自动的路由器，但他们自己心里有数，设计了一套系统，知道对于不同类型的请求该用哪个模型和什么设置，这样就不会把所有东西都砸在最先进最贵的模型上。

Or 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.

用最高的设置。

On the highest settings.

同时，有个想法已经流行好一阵了，就是说现在的模型已经这么好了，在很多用例上，没人能真正区分它们之间的差别，你要是还考虑用最贵最先进的模型，那肯定是在自欺欺人，觉得自己能看出差别来。

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.

我完全不认同这个说法。

I reject that pretty wholesale.

因为多个模型都能成功完成一个任务，就说它们可以互换，这就像说如果两个人能完成同一项工作任务，那谁来做都一样，反正任务完成了。

The idea that just because multiple models can successfully complete a task means that they're all interchangeable with one another is akin to saying that if 2 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，而这些任务恰恰就是你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.

但碰到很多高端的、重要的工作时，我仍然觉得，尽管这些模型都很能干，它们之间还是有很大的差异。

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.

我特别建议大家建立一套自己的个人基准测试，用来测试新模型。

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

这些测试不非得是别人的任务。

They don't have to be anyone else's tasks.

它们可以是你在意的事情。

They can just be the things that matter to you.

而且你可能会发现，对于你特定的任务，别人吐槽的模型效果很好，别人喜欢的模型反而不好使。

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，那个个人基准测试列表包括几样东西。

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

基本上是研究、写作、战略和批判性思维，以及搭建——包括界面设计和架构部分。

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，特别是涉及到写作或者战略思维这种事儿，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不可能给我一个基准来测试它如何迭代NLW那些疯狂的新商业点子。

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

那是我必须亲眼看看Fable 5、Opus和Sol在实际中怎么处理的东西。

That's something that I have to see how Fable-5 versus Opus versus Sol handle in practice.

而且就算到了这个普遍都能干的模型时代，我仍然发现这类事情上有巨大的差异。

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

这里提醒大家，现在对于所有人来说，新模型出来时的问题不再是“我该换成这个模型吗？”

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“这个模型怎么融入我的个人模型架构？”

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

这个模型在哪些用途上更好，值得付出它带来的财务或其他成本？

For what uses 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里榨出最大价值，而不是抱着“你只能选一个”的老想法，从一个模型晃悠到另一个。

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.

最后再补充一点，我想指出，如果你经济状况不允许你随便在模型之间切换，那很多这些考量就会变得不同。

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

而关于这点，说它们总体上都挺有能力的建议是准确的。

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

确实，现在锁定单一生态系统从未如此合适，因为它们各自都非常有能力，尽管各有不同的取舍。

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，现在好玩的部分来了——你可以亲自去测试、体验这些东西。

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

我很期待在这个 Labor Day 周末花时间测试一下，当然下周我会回来汇报。

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

那么现在，今天的 AI Daily Brief 就到这里了。

For now though, that is going to do it for today's AI Daily Brief.

一如既往，感谢大家的收听或观看。

I appreciate you listening or watching as always.

下次再见，peace。

Until next time, peace.

Peace。

Peace.

已剔除 2 处广告（点击展开查看）

M11:12广告 · 已剔除

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

To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at [email protected].

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

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

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

M112:01广告 · 已剔除

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

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

Here's why most legacy modernization projects fail.

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

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

Blitzy solves this the way it solves everything.

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

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

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

One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents.

That's 9x compression.

Retire technical debt while accelerating your roadmap.

See how at blitzy.com.

That's B-L-I-T-Z-Y dot com.

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

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

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

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

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

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

Head to robotsandpencils.com.

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

Forget local agents and chat workflows waiting on your laptop to be prompted— HyperAgent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

Marketing agents turn competitor moves into landing pages.

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

Ops agent chases the paperwork and tracks the budget.

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

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

Hire yours at HyperAgent.

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