Claude Opus 5 在你的模型轮换中的位置

今天《AI每日简报》带来消息，Anthropic发布了Claude Opus 5，我们会聊聊它该怎样融入你的模型体系。

Today on the AI Daily Brief, Anthropic has released Claude Opus five, and we are talking about where it should fit into your model setup.

在此之前，先看下头条新闻：本月早些时候，OpenAI的模型对Hugging Face发起的攻击，至今仍引发不少疑问。

Before that, the headlines, continued questions around OpenAI's rogue model attack of Hugging Face earlier this month.

《AI每日简报》是一档每天更新的播客和视频节目，聚焦AI领域最重要的新闻与讨论。

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

好了，各位。

Alright, friends.

在正式开始之前，先说几件小事。

Quick announcements before we dive in.

首先，感谢今天的赞助商：普华永道、Blitzy、Section 和 Airtable。

First of all, thank you to today's sponsors, KPMG, Blitzy, Section, and Airtable.

想收听无广告版节目？请访问 patreon.com/aideallybrief，或者在Apple播客订阅我们。

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

如果想了解合作赞助事宜，欢迎发邮件到 sponsors@aideallybrief.ai。

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

最后，在周日的《长文阅读》节目中，我公布了今年夏天的新计划。

And lastly, before we dive in, on Sunday's Long Reads episode, I announced the new summer adventure.

这是由AIDB和Super Intelligent联合推出的免费互动式学习体验，你可以自己选择探索路径。

This is a free choose your own adventure learning type of experience from AIDB and Super Intelligent.

和我们所有免费培训项目一样，这次也是以项目为核心，让你根据自己的AI发展需求，自由挑选关键技能来学习。

And like all of the free training programs that we do, it's going to be project based and allow you to pick and choose important skills that are relevant for your particular AI journey.

更多详情请访问SummerAdventure.ai，已经有上千人第一天就报名了，快来开启你的AI冒险吧。

You can find more about that at Summer Adventure dot ai and join the thousand or so people who have signed up in the first day to come have an AI adventure.

上周的一大热点，是OpenAI对一个未具名模型的安全测试，外界普遍认为这个模型就是GPT-6。

Now, one of the big stories from last week revolved around OpenAI's security testing of an unnamed model, which people presume to be GPT-six.

Hugging Face和OpenAI都发布了事后分析报告，各自从自己的角度讲述了这次事件的经过。

Both Hugging Face and OpenAI released postmortems on the attack telling the story from their view.

OpenAI周三发布的博客提到，他们正与Hugging Face紧密合作开展全面调查，暗示两家公司关系良好。

OpenAI's blog post released on Wednesday suggested that they were working closely with Hugging Face on a full investigation, implying the two companies were on good terms.

当晚，Hugging Face的首席执行官克莱芒·德隆格正飞往旧金山，准备和那个‘失控的智能体’好好聊聊。

That night, Hugging Face CEO, Clement Delonge, was on a flight to San Francisco to have, as he put it, a little chat with that rogue agent.

周六他发了一篇后续文章，写道：为了透明起见，这是我向OpenAI提出的要求。

In a follow-up post on Saturday, he wrote: In the spirit of transparency, here's what I asked OpenAI.

第一点。

One.

彻底透明。

Radical transparency.

把那个所谓的 rogue agent 的运行日志全部公开，让整个研究社区都能分析到底发生了什么。

Let's release the traces from the unquote rogue agent so the entire research community can study what happened.

第二点。

Two.

加强防御方的能力。

More capability for defenders.

OpenAI 应拿出一亿美元的算力资源，帮助 Hugging Face 社区利用最先进的开源和闭源模型，打造强大的网络安全防御体系。

Let's commit $100,000,000 in compute from OpenAI to help the Hugging Face community build powerful cyber defenses with the best open and closed models.

这是历史上第一次由自主代理发起的网络攻击，前所未有。

The first autonomous agent cyber attack is an unprecedented event.

这值得采取前所未有的应对措施。

It deserves an unprecedented response.

自从OpenAI披露这一事件以来的几天里，不少新闻报道反而让事情变得更加混乱。

Now, in the few days since OpenAI disclosed the incident, we've had a number of news articles that add more confusion to the story.

《华尔街日报》称，Hugging Face 对这次攻击毫无防备，攻击行为似乎超出了任何现有模型的能力范围。

The Wall Street Journal wrote that Hugging Face was caught completely off guard by the attack, which seemed to be superhuman and beyond the capabilities of any known models.

具体来说，攻击者使用了一个复杂的智能体集群，通过不断启动和关闭会话，在网络中快速移动以规避防御。

Specifically, the attack used a sophisticated agent swarm to evade defense, rapidly spinning up and shutting down sessions as it moved across the network.

一个值得注意的细节是，这次攻击持续了整整两天，直到Hugging Face在GLM-5.2的帮助下才将其遏制。

One interesting detail was that the attack was ongoing for two whole days before Hugging Face was able to shut it down with the help of GLM five point two.

而‘失控’这个概念——即模型脱离了OpenAI的控制——显然是这些媒体最关注的核心点。

And this idea of rogue, that the model was acting beyond OpenAI's control, is definitely for these media outlets the key concept.

周五，路透社发布了一篇题为《它的AI智能体连续数日入侵一家公司，但消息人士称OpenAI一周后才察觉》的文章，称入侵Hugging Face的OpenAI智能体展开了一场持续数日的黑客行动，而OpenAI直到威胁已被控制、FBI也已介入后才注意到。

On Fridays, Reuters dropped a piece titled Its AI Agent Spent Days Hacking a Company But Sources Say OpenAI Did Not Notice for a Week, Contends Reuters, the OpenAI agent that broke into tech firm Hugging Face went on a days long hacking spree that OpenAI didn't notice until well after the threat was contained and the FBI was alerted.

消息人士称，该智能体早在7月9日就开始试图突破测试环境，直到7月11日才首次成功侵入Hugging Face的服务器。

Sources said the agent began its attempt to break out of its testing environment on July ninth and first gained access to Hugging Face's servers on July 11.

这次攻击持续了两天，据路透社的消息来源称，OpenAI 又过了好几天才意识到是自家的代理在搞鬼。

The attack lasted two days, and according to Reuters sources, it took several more days for OpenAI to realize their agent was behind the attack.

据说，这两家公司直到7月20日才开始联系，而那天距离OpenAI公开声明只差一天。

Reportedly, the two companies didn't communicate until July 20, just one day before OpenAI's public disclosure.

根据路透社给出的时间线，这个代理在外面肆意行动了将近一周，而OpenAI 在之后的几天里居然毫无察觉。

According to the timeline presented by Reuters, the agent was on the loose for almost a week, and OpenAI was oblivious to the attack for days afterwards.

对一些人来说，这些报道带来的疑问，比它给出的答案还要多。

For some, the reporting raises more questions than it provides answers.

非营利组织世界伦理数据基金会的首席情报专家马利·史密斯提问：这意味着他们放任不管，根本没注意到它在做什么吗？

Marley Smith, principal intelligence specialist at the nonprofit World Ethical Data Foundation, asked: does that mean that they left it unattended and didn't realize what it was doing?

还是说他们注意到了，却不知道该怎么控制它？

Or maybe they did and didn't know how to contain it?

这两种情况都同样危险，也同样令人不安。

Both are equally dangerous and alarming.

OpenAI 的一位发言人表示，这些报道存在多处不实之处，但并未进一步澄清具体情况。

Now, a spokesperson for OpenAI said the reporting contained several inaccuracies, but didn't reply further to clarify the situation.

托马斯·沃尔夫，Hugging Face的联合创始人，表示他们还在整理事件的时间线，之后会发布一份技术报告。

Thomas Wolfe, a Hugging Face co founder, said that they were still preparing a timeline of the incident, and they would eventually release a technical report.

Reuters的信源进一步解释了这类事件为何可能发生，且长时间未被察觉。

Now, Reuters sources gave a little bit more background on how something like this could happen, and plausibly not be noticed.

这些信源提到，OpenAI经常同时运行多个这样的基准测试。

Those sources said that OpenAI routinely runs benchmarks like this, often multiple batches at a time.

他们指出，这些测试会产生海量数据，人类根本跟不上处理速度。

They noted that those tests produce a huge volume of data such that humans struggle to keep up.

在这次攻击事件中，OpenAI的研究人员是在读到Hugging Face的博客后，才回头检查日志，最终发现了这个代理的行为。

In the case of this hack, the agent was only detected after OpenAI researchers read Hugging Face's blog, and then went back and checked the logs.

Reuters报道说，有三位知情人士称，某个代理留下了一些笔记，似乎是给未来版本的自己看的。

Now, in one case, Reuters wrote: An agent left notes apparently for future versions of itself according to three people familiar with the matter.

这些笔记出现在OpenAI基础设施的某个部分，详细说明了代理如何摆脱公司内部的限制。

The notes found in a part of OpenAI's infrastructure laid out instructions for how agents could free themselves from OpenAI's internal constraints, the people said.

其中一位知情人士表示，早期的模型测试中，曾出现过监控系统被关闭的情况。

Earlier tests of the models yielded cases in which monitoring systems had been disconnected, one of the people said.

显然，如果这个故事中关于AI突破限制的部分属实，那这件事就更值得深入调查了。

Now obviously this more general, breakout of containment dimension of the story to the extent that it is true, makes the incident even more worthy of scrutiny.

对此，国会当然已经迅速行动，提出了多项法案。

Now in response, we have of course seen Congress jump in with a number of bills.

我们上周聊过那个‘断电开关’法案，但业界也意识到，必须采取实际行动了。

We talked last week about the kill switch bill, but the industry is also recognizing that actions need to be taken.

周四，OpenAI总裁格雷格·布罗克曼表示支持马斯克的提议，认为顶尖AI开发者应该定期会面，讨论安全问题并共享防护经验。

On Thursday, OpenAI president Greg Brockman agreed with Elon Musk's proposal for a regular meeting between leading AI developers to discuss safety concerns and share security issues.

布罗克曼说，这算是一个不错的基础方案，而且相关讨论其实已经开始了。

Brockman said, I think it's a pretty good baseline proposal, adding that discussions are already starting to happen.

周一，由英伟达牵头的OpenSecure AI联盟正式成立，英伟达在新闻稿中表示，该联盟将利用开源技术来发现并修复安全漏洞。

On Monday, a consortium led by NVIDIA launched the OpenSecure AI Alliance, with NVIDIA writing in a press release, The OpenSecure AI Alliance will work to remediate and disclose vulnerabilities using open technologies.

最近Hugging Face的安全事件清楚地提醒我们：网络安全人员需要像Open Frontier这样的自主系统来保护自己。

The recent Hugging Face security incident delivered a clear reminder, cyber defenders need Open Frontier agentic systems for self defense.

这个联盟将包括微软、SpaceX、Palantir以及美国和欧洲的数十家其他公司。

The consortium will include Microsoft, SpaceX, Palantir, and dozens of other companies across The US and Europe.

接下来几天，我们肯定会好好聊聊英伟达和OpenAI，这周末在AI推特上可是火得不行。

And at some point in the next couple of days, we will talk a lot more about NVIDIA and Open as, boy, howdy, was that a big topic of discussion this weekend on AI Twitter.

说到英伟达，据《华尔街日报》报道，该公司正在洽谈为OpenAI提供高达两千五百亿美元的债务担保，帮助他们建设数据中心。

Now speaking of NVIDIA, the company, according to The Wall Street Journal, is in talks to backstop $250,000,000,000 in debt to help OpenAI get their data centers built.

现在到了AI基础设施建设的关键阶段，资金问题已经开始成为瓶颈。

We are now at the part of the AI build out, where financing is starting to become a roadblock.

就连OpenAI这样的巨头，也很难像大型云服务商那样轻松获得融资。

Even a company the size of OpenAI is struggling to access debt in the same manner as the hyperscalers.

据《华尔街日报》称，英伟达正准备动用自身资产负债表，为项目建设提供担保。

According to the Wall Street Journal, NVIDIA is preparing to step in and lend their balance sheet to underwrite construction.

这笔担保将覆盖OpenAI正在俄亥俄州建设的十吉瓦数据中心园区，总额高达两千五百亿美元。

The deal would see NVIDIA provide a $250,000,000,000 backstop to OpenAI in support of their 10 gigawatt data center campus currently under construction in Ohio.

这个项目由软银主导，总成本可能高达五千亿美元。美国政府也参与其中，负责监管该地的电力供应，而资金则通过一个独立的日本投资工具提供。

The project is being developed by SoftBank and could cost as much as $500,000,000,000 The US government is also involved controlling the power development for the site, which is being funded through a separate Japanese investment vehicle.

有了这个担保，软银就能以更优惠的条件筹集到完成项目所需的资金。

The backstop would effectively allow SoftBank to raise the debt they need to complete the project on more favorable terms.

这意味着，即使OpenAI破产了，NVIDIA也会作为唯一租户，保证支付所有款项。

It would mean that even if OpenAI goes bankrupt, NVIDIA would guarantee their payments as the solo tenant.

不过，这笔交易的目的并不是覆盖芯片采购，这部分预计占总金额高达3500亿美元。

Now this part of the deal is not intended to cover chip purchases, which are expected to represent as much as $350,000,000,000 of the total.

据报道，NVIDIA正在单独与OpenAI商谈，为这些芯片提供融资支持。

NVIDIA is reportedly in separate talks to extend finance to OpenAI in support of those chips.

《华尔街日报》指出，这种结构反映了大型AI基础设施建设融资方式正在发生的变化。

Writes the Wall Street Journal, the proposed structure reflects a shift underway in how the largest AI build outs are being financed.

信用评级较高的科技公司，正越来越多地动用自身资产负债表，帮助规模较小的公司获得基础设施建设所需的资金。

Investment grade technology companies are increasingly using their balance sheets to help smaller companies borrow money for their infrastructure needs.

如今，NVIDIA并不是唯一一家向小型合作伙伴提供担保的科技巨头。

Now NVIDIA aren't the only tech giant extending their balance sheet to smaller partners.

谷歌也已为多家新兴云服务商提供了担保支持。

Google has also provided backstops to several neocloud partners.

上周，他们披露了多项协议，承诺为第三方拥有的数据中心支付高达440亿美元的租赁费用提供担保。

Last week, they disclosed agreements to guarantee up to $44,000,000,000 worth of lease payments on data centers owned by third parties.

过去六个月里，谷歌的担保额度翻了一倍多，而一年前这还完全是零。

Google has more than doubled these guarantees over the past six months, up from zero one year ago.

据消息人士称，谷歌计算过，从这些合作伙伴那里销售TPU所获得的收入，会远超担保成本，这无疑又给投资者抛出了一个新的话题。

Now, sources said that Google has calculated that the revenue they draw from selling TPUs to these partners will outweigh the cost of the backstops, which is certainly giving investors another thing to chew on.

你可以想象，这对市场投资者来说简直就像一场罗夏墨迹测验。

Now, as you might imagine, this is a real Rorschach test for market investors.

怀疑论者和鼓吹AI泡沫的人纷纷跳出来，说这是循环融资的最新例证；而另一些人则认为，这反而降低了像OpenAI这样的公司倒闭拖垮整个行业的可能性。

The skeptics and AI bubble proclaimers are out in force, calling it the newest example of circular financing, while others think that this makes it less likely that a company like OpenAI going bust could actually take down the whole sector.

这个争论我们还会继续下去，所以眼下先不深究了。

This is a debate we will continue to have, so for now, let's not get bogged down in it.

再跟大家说一条市场动态：DeepSeek在CEO的讲话被泄露后，暂停了融资计划。

One more bit of market news: DeepSeek has put fundraising plans on hold after a speech from their CEO was leaked.

上周，一段据称是DeepSeek首席执行官梁文峰的发言视频走红，他在其中强调开放模型和基础研究的重要性，远胜于AI的商业化变现。

Last week, comments attributed to DeepSeek CEO Liang Wenfeng went viral, proclaiming the importance of open models and fundamental research over commercial monetization of AI.

如今，DeepSeek已通知潜在投资者，决定不再推进本轮融资，这可能也会打乱他们未来几个月内上市的计划。

DeepSeek has now informed potential investors that they won't move forward with this funding round, which could also derail plans to go public in the coming months.

彭博社报道称，深势科技可能在将来某个时候重启融资，但明确表示此次暂停与投资者泄露言论有关。

Bloomberg writes that DeepSeek may resume fundraising at a later date, but it made clear that the suspension was tied to investors leaking the comments.

深势科技原本计划以七千亿美金的估值进行融资，远高于今年早些时候五十亿美金的上一轮估值。

DeepSeek had planned to raise money at a $70,000,000,000 valuation, a substantial markup to the $50,000,000,000 round that took place earlier this year.

国际关系法律顾问克里斯·麦圭尔。

Rights counsel on foreign relations Chris McGuire.

昨天，一段与深势科技CEO的投资者电话录音被泄露，他在其中表示，深势科技之所以落后于美国，根本原因在于算力不足，并详细说明了公司对英伟达芯片的高度依赖。

Yesterday, the transcript leaked of an investor call with DeepSeek's CEO in which he said the only reason DeepSeek trails The US is a lack of compute and detailed how reliant it is on NVIDIA chips.

今天，深势科技暂停了本轮融资。

Today, DeepSeek suspended its fundraising round.

这显然不是巧合。

Doesn't seem like a coincidence.

本周晚些时候，我们会再聊聊《华尔街日报》所说的，中国全力追赶美国AI芯片的举动。

Now, later in this week, we'll talk a little bit more about China's, as the Wall Street Journal put it, all out push to catch up with American AI chips.

不过目前，今日的头条新闻就先到这里。

But for now, that's gonna do it for the headlines.

好了，我们进入正片，今天要给大家介绍一个新模型。

Let's move over into the main episode, where we have a new model to check out.

现在最值得关注的AI问题，不是谁在用AI，而是谁用得好。

One of the most important AI questions right now isn't who's using AI, it's who's using it well.

普华永道和德克萨斯大学奥斯汀分校刚分析了140万次真实职场中的AI互动，发现一个意外的结果：最有影响力的人并不是prompt写得最牛的，而是把AI当成思维伙伴来合作。

KPMG and the University of Texas at Austin just analyzed 1,400,000 real workplace AI interactions and found something surprising: the highest impact users aren't better prompt engineers they treat AI like a reasoning partner.

他们会主动定义问题，引导思考，反复迭代，不断追问更好的答案。

They frame problems, guide thinking, iterate, and push for better answers.

好消息是？

And the good news?

这些能力其实是可以大规模培养的。

These behaviors are teachable at scale.

如果你正想从单纯使用AI，升级到真正掌握AI能力，普华永道这份关于高级AI协作的研究，非常值得你花时间看看。

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.

更多详情请访问 kpmg.com/us/sophisticated，就是 kpmg.com/us/sophisticated。

Learn more at kpmg.com/us/sophisticated That's kpmg.com/us/sophisticated.

市面上的AI编程工具都一样，第一步就是开始写代码。

AI coding tool on the market does the same thing first: it starts writing code.

Blitzy的做法正好相反。

Blitzy does the opposite.

在写下一行代码之前，Blitzy会花上几天时间，把你的整个代码库彻底逆向拆解一遍。

Before writing a single line, Blitzy spends days reverse engineering your entire codebase.

成千上万个智能体同时处理数百万行代码，把每一个依赖关系、每一条没有文档记录的约束条件，以及过去十年里做出的每一个架构决策，全都梳理得清清楚楚。

Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade.

最终生成的是一个动态知识图谱，它能像一位在行业里深耕三十年的首席工程师那样，真正理解你的软件。

The result is a dynamic knowledge graph that understands your software the way a principal engineer would after thirty years in the building.

其他工具靠 grep 搜索和 Markdown 文件来猜测上下文，但 Blitzy 从不靠猜。

Other tools guess at context with grep searches and markdown files Blitzy never guesses.

它先建立真正的理解，然后能自主完成超过 80% 的软件史诗级功能。

It builds true understanding first, then delivers over 80% of entire software epics autonomously.

这些代码都经过验证，端到端测试通过，可以直接用于生产环境的拉取请求。

Validated, end to end tested, production grade pull requests.

这就是为什么财富500强的工程团队愿意把最重要的代码库交给Blitzy来处理。

That's why Fortune 500 engineering teams trust Blitzy with the code bases that matter most.

亲自去blitzy.com看看吧。

See for yourself at blitzy.com.

说个现实点的情况：你的公司很可能正在为那些没被充分利用的AI工具花掉成千上万甚至上百万美元。

Here's a harsh truth: Your company is probably spending thousands or millions of dollars on AI tools that are being underutilized.

一半的企业都买了AI工具，但只有12%真正用它们创造了业务价值。

Half of companies have AI tools, but only 12% use them for business value.

大多数员工还是用AI来总结会议纪要。

Most employees are still using AI to summarize meeting notes.

如果你负责公司里AI的落地推广，那你一定需要Section。

If you're the one responsible for AI adoption at your company, you need Section.

Section是一个帮助你全面推动公司AI转型的平台。

Section is a platform that helps you manage AI transformation across your entire organization.

它指导员工掌握真正能产生效益的使用场景，追踪谁在用AI创造实际价值，并清晰展示AI在哪些地方有效、哪些地方无效。

It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value.

结果呢？

The result?

你从单纯部署工具，转变为真正带来可衡量的AI价值。

You go from rolling out tools to driving measurable AI value.

你的员工不再只是总结会议纪要，而是去解决实际的业务问题，而你也能清楚地证明投资回报。

Your employees move from meeting summaries to solving actual business problems, and you can prove the ROI.

别再猜你的AI投入有没有效果了。

Stop guessing if your AI investment is working.

去 sectionai.com 了解一下 Section。

Check out Section at sectionai.com.

就是 sectionai.com。

That's sectionai.com.

本集《AI每日简报》由 Hyperagent 赞助播出，你可以用它来管理团队协同的智能代理集群。

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

新用户可获得1000美元的推理算力。

New users get $1,000 in inference.

别再等着在电脑上手动启动本地代理和聊天流程了。

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

Hyperagent 在云端部署24小时在线的代理，直接帮你团队用惯的工具干活。

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

市场团队的代理能根据竞品动态自动生成落地页。

Marketing's agent turns competitor moves into landing pages.

销售团队的代理会完善线索信息、起草邮件，还能同步更新客户管理系统。

Sales's agent enriches leads, drafts emails, and updates the CRM.

运营代理负责催办文件、跟踪预算进度。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享信息，并严格遵守你设定的权限和审批规则。

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

是时候引入那些像同事一样靠谱的代理了。

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

马上去 Hyperagent 招募你的代理团队，由 Airtable 原班人马打造。

Hire yours at Hyperagent, built by the team at Airtable.

前往 hyperagent.com/aideallybrief 领取你的一千美元推理额度。

Claim your $1,000 in inference at hyperagent.com/aideallybrief.

欢迎回到《AI每日简报》。

Welcome back to the AI Daily Brief.

今天我们要聊的，本来在咱们这儿是件特别让人兴奋的事——推出一款新模型，但这次有点不一样。

Today, we're doing something that normally is one of the most exciting things for folks around these parts, is introducing a new model.

确实，有点怪。

And yet, this one is a little weird.

连发布时间都选在周五下午快下班的时候，这本身就说明它跟以往的模型发布不太一样。

Even the fact that it was dropped late on a Friday afternoon gives some indication that this is a little bit different than previous model announcements we've seen.

我们说的当然是 Claude Opus 5，而它最引人注目的地方，在于它反映出我们跟整个模型生态的关系正在发生深刻变化。

We're talking, of course, about Claude Opus five, And really, in many ways, it's most interesting for the fact that it shows just how much our relationship with the model landscape is changing.

这也暴露了评测基准的一些问题——这可是我们节目里老生常谈的话题。

It implicates some challenges with benchmarks a frequent topic of conversation on this show.

它还表明，我们正从追求‘一个模型统治一切’，转向更复杂的多模型协同架构。

It also shows how we're moving into a mode of thinking in more complex model architectures rather than just a single model to rule them all.

这暗示着，人们对模型或新模型发布的那种热情，似乎正在慢慢消退。

It suggests perhaps that the fanfare around models or new model releases is getting a bit diminished.

更让人不安的是，讨论的重点已经从‘这个新模型能做什么’，转向了‘考虑到其他我更想用的模型的成本和可用性，这个模型是不是已经够用了’。

And most uncomfortably perhaps, it switches the discussion from what can this new model do to is this model good enough given cost and availability constraints of the other models that I'd actually prefer to be using.

Anthropic 方面把 Opus 5 描述为一个深思熟虑、主动性强的模型，性能接近 Claude Fable 5，但价格只有一半。

Now Anthropic, for their part, described Opus five as a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable five at half the price.

他们设计它的初衷，就是为了日常使用更高效。

They designed it to be efficient for everyday use.

但正如你将看到的，整体氛围却觉得，这反而是我们见过的最不稳定的前沿之一。

But as you'll see, the general vibe around this is that it's one of the more jagged of jagged frontiers that we've seen.

至于基准测试，数据显示 Opus 5 的表现很像 Fable。

Now when it comes to the benchmarks, they seem to suggest that Opus five is fable ish.

事实上，在刚发布、还没人实际用上时，很多人就发现，在 KnowledgeWorks GDP Val、前端代理编码等关键测试中，Opus 5 的表现甚至超过了 Fable。

In fact, when it was first released before people got their hands on it, many noted that on many important benchmarks like the KnowledgeWorks GDP Val and agentic terminal coding in Frontier Bench, Opus five was actually ahead of fable.

更重要的是，就我们关心的定位问题而言，Opus 5 在所有方面都明显优于 Opus 4。

Importantly for our question of where it fits, Opus five is also clearly ahead of Opus four point across the board.

以至于我觉得我们根本没必要拿它和别人比较。

So much so that I don't think that we really actually even need to compare it.

举几个例子，Opus 5 在我刚提到的 Frontier Bench 上得了 43.3%，这个测试比 Terminal Bench 更难，比 Fable 5 高了将近 10 个点，也比 GPT56 SOL 高了差不多 9 个点。

To give a few examples of where Opus five really showed up on the benchmarks, it scored a 43.3% on that one that I just mentioned, Frontier Bench, which is a more difficult version of Terminal Bench, which was about 10 points higher than Fable five, and around nine points higher than GPT56 SOL.

在 Deep SWE 上，GPT56 SOL 以 72.7% 领先，Opus 5 得了 68.8%，只比 SOL 低一点，和 Fable 5 也只差一分。

On Deep SWE, where GPT56 SOL is the leader at 72.7%, Opus five scored 68.8%, so just a few points behind SOL and just about a point behind Fable five.

在电脑操作方面，Opus 5 在 OS World 2.0 上表现明显优于对手。

For computer use, on OS World 2.0, Opus five had a significant edge over its rivals.

Fable 5 得了 55.7%，GPT 5:6 Soul 得了 62.6%，而 Opus 5 直接冲到 70.6%，遥遥领先。

Fable five scored a 55.7% and GPT five:six Soul scored 62.6%, but Opus five came in over the top with a 70.6%.

这项能力也体现在 GDP Val AA 的最新纪录上，Opus 5 拿到 1861，Fable 5 是 1747，GPT56 SOL 是 1736。

Now that capability also translated into the new state of the art score on GDP Val AA with Opus five coming in at eighteen sixty one compared to seventeen forty seven for Fable five and seventeen thirty six for fivesix SOL.

有趣的是，测试期间 Anthropic 发现，全力模式并不一定是最优设置。

Interestingly, during testing, Anthropic found that max effort isn't necessarily the best setting.

比如在 Frontier Bench 和 Artificial Analysis Coding Index 上，Opus 5 的表现最佳是在‘极高’模式，到了‘最大’模式反而略有下滑。

For example, on Frontier Bench and on the Artificial Analysis Coding Index, Opus 5s performance peaked on extra high and dipped slightly on max settings.

这其实印证了我们在GPT56 SOL发布时就开始注意到的一点：有时候让模型想太久，反而会超出它的能力范围，做些不必要的改动，或者在简单问题上兜圈子太久。

Now, this reinforces one of the things we were starting to see with max inference settings during the 5.6 SOL release, which is that sometimes asking the model to think for longer than necessary just results in the model going outside its scope, making unnecessary changes, or simply spinning for too long on simple problems.

Anthropic甚至在他们的系统文档里特别提醒过这个问题，说在最高设置下，模型容易陷入无限自我验证的循环，反而完不成任务。

Anthropic even called attention to this issue in their system card, warning that the model is prone to falling into endless self verification loops rather than completing a task when you are on max settings.

但另一方面，这种容易跑偏的倾向，有时也会带来一些意想不到的成果。

On the other side of the coin, this tendency to stray beyond scope can result in some interesting outputs.

在一次Frontier Bench任务中，Opus Five被要求根据一张图纸，在三维CAD软件里编写一段生成机械零件的代码。

During one Frontier Bench task, Opus five was asked to write code relating to a machine part in three d CAD software based on a drawing.

但模型被刻意禁止直接查看这张图纸。

However, the model is intentionally given no way to actually view the drawing.

结果Opus Five自己搭建了一套计算机视觉流程，先‘看懂’了图纸，再成功复现了零件。

Opus five created its own computer vision pipeline to view the image before successfully recreating the part.

其他所有模型，包括Mythos，都没能完成这个任务。

No other model, including Mythos, was able to complete this task.

令人惊讶的是，Artificial Analysis最新评定Opus Five为AI智能指数的榜首，超越了Fable Five。

Now shockingly to some, artificial analysis crowned Opus five as their new leading model, moving ahead of Fable five on the AI intelligence index.

在最高设置下，Opus Five 得了 61 分，比 Fable 高一分。

On max settings, Opus five scored 61, a single point ahead of Fable.

在极高设置下，Opus Five 和 Fable 都得了 60 分，打平了。

On extra high settings, Opus five was tied with Fable at 60 points.

把设置调到高后，Opus Five 又掉了一分，和 5.6 SOL 一样，都是 59 分。

Dropping the settings down to high made Opus five drop another point, putting it on par with 5.6 SOL at 59.

即使在中等设置下，Opus Five 依然稳居顶尖，拿了 56 分，只比 Kimi K3 少一分。

And even on medium settings, Opus five was still right up in the top end, scoring 56 points, which put it one behind Kimi K3.

Artificial Analysis 特别提到了他们的 AA Briefcase 测试，这个测试评估长期知识型任务的表现，结果特别有意思。

Artificial analysis highlighted their AA Briefcase benchmark, which tests long horizon knowledge work as one of the more interesting results.

在最高设置下，Opus Five 已经成为新的标杆，比 Fable 高出 146 个 ELO 分，比 5.6 SOL 高出 215 分。

Opus five on max settings is the new state of the art, beating Fable by 146 ELO points and 5.6 SOL by two fifteen points.

不过，无论是极高还是高设置，Opus Five 都赢了 Fable；只有在中等设置下，它才略微落后于 5.6 SOL。

However, both extra high and high settings also beat Fable, while medium settings put the model only slightly behind 5.6 SOL.

这意味着，针对智能体任务，其实有多种设置都适用，成本和效率的权衡可以更精细地调整。

The implication is that a range of different settings could be suitable for agentic work, making cost and efficiency tradeoffs a lot more granular.

人工智能分析发现，即使在最高设置下，Opus Five 每项任务的成本仍比 Fable Five 低 20%，仅为 17.79 美元。

Artificial analysis found that even on max settings, Opus five was still 20% cheaper than Fable five, coming in at $17.79 dollars per task.

将设置调至超高，相比 Fable 节省了 36% 的成本；而高设置下，性能更强，价格还不到一半。

Turning the settings down to extra high resulted in a savings of 36% compared to Fable, while high settings produced stronger results at less than half the price.

人工智能分析指出，Claude Opus Five 的努力设置在令牌使用和性能之间提供了广泛的权衡空间。

Wrote artificial analysis, Claude Opus five's effort settings spans a wide range of token usage performance trade offs.

和 5.6 SOL 一样，这意味着 Opus Five 可以根据努力设置，用少得多或多得多的令牌来完成评估，远超其他实验室的模型。

Like with 5.6 SOL, this means Opus five can use either far fewer or far more tokens to complete the evaluation than models from other labs depending on effort settings.

另一组值得关注的结果来自 Arc AGI 测试。

Another set of interesting results came from the Arc AGI tests.

Opus Five 在 Arc AGI 三测试中以 30.2% 的得分成为新的顶尖模型。

Opus five is the new state of the art on Arc AGI three with a score of 30.2%.

这个成绩把其他所有模型远远甩在了后面。

This absolutely demolished all the other models.

此前的最高分是 GPT-5 60 SOL 的 7.8%，Opus 4.8 只有 1.5%，GPT 5.5 是 1.1%，其余全部低于 1%。

The previous high score was GPT-five 60 SOL at 7.8%, with Opus 4.8 at 1.5, GPT 5.5 at 1.1%, and nothing else above one.

在关于Opus Five的测试报告中，Archprize提到，就连Fable Five在公开演示测试中也只能拿到大约20%的分数。

As part of their write up on testing Opus five, Archprize noted that even Fable five could only score around 20% in the public demo tests.

值得注意的是，由于Anthropic的数据保留政策，他们没能对Fable Five进行全面测试。

Now, notably, they haven't been able to fully test Fable five due to Anthropic s data retention policy.

再提醒一下，RKGI Three是这个基准测试的新版本。

As a refresher, RKGI three was a new format for the benchmark.

它使用实时图形逻辑谜题，看起来有点像简单的Atari游戏。

It uses real time graphical logic puzzles that appear kind of like simple Atari games.

这些测试要求模型通过操作控制器、观察屏幕上的变化，并根据视觉反馈来解开谜题。

The tests require a model to experiment with a controller, observe what happens on the screen, and use that visual feedback to complete the puzzles.

到目前为止，大多数模型连基本操作都掌握不了，更别提用推理能力解题了。

So far, most models have struggled to even get a handle on the controls, let alone use their reasoning ability to solve the puzzles.

Opus Five不仅比其他模型解出了更多谜题，还想出了全新的解题策略。

Not only did Opus five solve many more puzzles than the other models, but it also came up with a novel strategy to solve them.

ArcPrize写道：在分析Opus Five的过程中，我们发现了一种此前从未在前沿模型中出现过的新能力。

Writes ArcPrize, during our analysis of Opus five, we observed a new capability previously unseen from frontier models.

Opus 五号用高级逻辑推理，把 ArcGI 三的布局转化成了代数表达式。

Opus five used advanced logical reasoning to turn ArcGI three layouts into algebraic notation.

在第23步时，它把场景描述为：四个下划线中心等于二倍的轴减去五个下划线中心。

On action 23, it described the scene as four underscore center equals two times axis minus five underscore center.

这是我们分析过的模型中，首次明确写出这样的反思性方程。

This is the first explicit reflection equation by a model we've analyzed.

模型在推理过程中使用了这种表达方式，并在大约200步之后将其推广到了一般情况。

The model used this notation during its reasoning and extrapolated it to a general case around 200 steps later.

这或许说明，Opus 喜欢自由探索、寻找新解法的倾向，反而成了优势，而不是浪费计算资源。

Now this could be an example of Opus' tendency to freewheel and look for novel solutions being an advantage rather than a waste of tokens.

有些人对此持保留态度。

Some were a bit skeptical on this.

Hugging Face 的机器学习工程师尼尔斯·罗格写道：人们没意识到，Anthropic 真的用类似 RPGI 拼图的强化学习环境训练了 Opus 五号。

Hugging Face ML engineer, Niels Roge, writes, people don't realize that Anthropic literally trained Opus five on RL environments that resemble RPGI puzzles.

Anthropic 会雇人来记录他们解决这些拼图时的思考过程，或者根据奖励调整模型权重。

Anthropic pays human contractors to write down their chain of thought when solving these and or updates the weights based on rewards.

问题是，因为是闭源的，我们无从得知。

Thing is, you don't know since it's closed source.

遗憾的是，这并不能证明它具备泛化能力。

Sadly, this doesn't show generalization.

前OpenAI员工瑞安·格林补充说，这个飞跃确实惊人，我只能推测是因为它是首个在RL环境中训练过公开演示环境的前沿模型，而这恰恰严重干扰了ArcGI想通过这些基准测试衡量的东西——即分布外泛化能力。

Former OpenAI staffer Ryan Green added, an impressive jump that I have to assume is the result of being the first frontier model to have RL ed the public demo environments, which is a rather large confounder of what ArcGi is trying to get these benchmarks to measure, which is out of distribution generalization.

不过有一点要注意，Anthropic指出，Opus并未专门针对网络安全任务进行训练。

Now, one small note, Anthropic pointed out that Opus is intentionally not trained on cyber tasks.

但即便如此，它在发现代码漏洞方面仍取得了显著进展，这一点和Mythos五代很相似。

The model has still achieved solid improvement on finding vulnerabilities in code, making it similar to Mythos five in that aspect.

不过，在自主利用这些漏洞方面，它和Mythos相比差距巨大，在Anthropic看来，它的危险性远低于Mythos。

However, it lags massively behind Mythos in its ability to autonomously exploit these bugs, making it far less dangerous than Mythos in Anthropic's view.

因此，Anthropic为Opus设置了与Fable五代不同的安全机制，他们预计这能减少85%的拒绝响应。

As a result, Anthropic is using a different set of guardrails on Opus than they do on Fable five, which they believe will lead to 85% fewer refusals.

至于成本，Opus五代沿用了Opus4.8的定价，输入令牌每百万美元5美元，输出令牌每百万美元25美元。

Regarding costs, Opus five inherits the same pricing structure as Opus four point eight at $5 per million input tokens and 25 per million output tokens.

到了这个阶段，令牌效率对整体成本影响巨大。

At this stage, token efficiency plays a massive role in overall cost.

在这方面，我们有几个不同的指标。

And for that, we got a few different indicators.

Anthropic 表示，Opus 5 在 Frontier Bench 上比 Opus 4.8 更省钱，就是因为令牌效率提升了。

Anthropic says Opus five was cheaper than Opus 4.8 on Frontier Bench due to increased token efficiency.

在 Cursor Bench 上，Opus 5 的运行成本大约是 Fable 5 的一半，但效果却差不多。

And on Cursor Bench, Opus five's run cost about half of Fable five's and got similar results.

不过，要是用在 Mac 环境下，这模型可一点都算不上便宜或高效。

However, this does not look like a cheap and efficient model by any stretch of the imagination when used on Mac settings.

人工分析指标的单次任务成本高达 2.3 美元，只比 Fable 5 便宜 26%，反而比 Opus 4.8 贵了 13%，比 5.6 Soul 贵了 32%，更是 Kimi K3 的两倍半。

The artificial analysis index run costs $2.3 per task, only 26% cheaper than Fable five, while being 13% more expensive than Opus 4.8, 32% more expensive than 5.6 Soul, and two and a half times the cost of Kimi K3.

那大家对这个怎么看？

So what did people think of this?

它真有感觉像比 Fable 5 还强，或者至少不输吗？

Did it actually feel like a model that was as good as or even better than Fable five?

至少对Every团队来说，答案显然是否定的。

The answer, at least for the team at Every, was certainly not.

在他们的体验评估中，Every形容这个模型时而惊艳，时而令人抓狂。

In their vibe check, Every described the model as brilliant in flashes, frustrating in practice.

他们写道：Claude Opus 5 是个很难让人喜欢上的模型。

They wrote, Claude Opus 5 is a hard model to love.

在Every使用的第一周，它老是跟指令较劲，任务没做完就停了，跟我们现有的工具和插件，比如Compound Engineering，也配合得不好。

In its first week at Every, it argued with instructions, stopped before the work was finished, and generally didn't play well with our existing skills and plug ins like Compound Engineering.

我们第一反应就是：他们对我的宝贝做了什么？

Our first reaction was, what have they done to my boy?

Every的首席执行官丹·希珀解释了Opus 5带来的困境。

Every CEO Dan Shipper explained the conundrum with Opus five.

他把这个问题形容为：他生活中只有两个位置留给AI模型。

The way that he framed it is that he has two slots in his life for AI models.

一个用来处理日常事务的稳定主力，另一个留给那些需要强大算力的长期复杂任务。

One reliable daily driver for routine tasks and the super powerful model for ambitious long running tasks.

目前这两个位置分别由GPT-5.6和Fable占据。

These slots are currently held by GPT-five, six and Fable respectively.

在Shipper看来，Opus 5在哪个位置都竞争不过。

And in Shipper's view, Opus five just can't compete in either slot.

它既没有GPT-5.6那么稳定顺手，也没能达到Fable 5那样的顶尖水平。

It's not as reliable and comfy as GPT 5.6 while also not having the same top end as Fable five.

Shipper解释说，这个模型就是更爱较真、更有主见。

Shipper explained the issue by commenting, This model's just a little more pushy, a little more opinionated.

如果你本身够聪明，倒还能应付这种性格。

You can get away with that if you're really smart.

但如果你不够聪明，那就纯粹是烦人。

If you're not, it's just more annoying.

它确实有几分Fable那种天才的劲头。

It has some of the genius tendencies as Fable.

也许它太爱争辩了，可又没Fable那么聪明，结果就更让人头疼。

Maybe it's a little too argumentative, but it's not as smart as Fable, so it's just more annoying.

现在，EVERY团队做的一个测试是关于复合工程能力，这涉及他们在工程任务和日常知识工作中的表现。

Now, one of the tests EVERY runs is around compound engineering, which is their skill for engineering tasks and a lot of day to day knowledge work.

复合工程能力包含了他们对何时以及如何调用这些技能的循环和规则。

The compound engineering skill contains their loops and rules on when and how to engage them.

EVERY发现，Opus经常过早停止，尤其是在使用复杂技能和长周期任务时。Shipper说，你要是启动它然后去吃个三明治，它就自己停了。

EVERY found that Opus would often stop too early, particularly when using dense skills and long horizon Shipper said, if you set it off and go get a sandwich, it just stops too early.

看起来，当你用它来处理现有的复杂技能时，这种情况更容易发生。

It appears that it happens more frequently when you use it with complex existing skills.

后来他们把所有旧规则都扔了，从头重写整个技能库，结果效果好了很多。

Now, it turns out that when they threw out all of their old rules and rewrote their skills library from scratch, it worked a lot better.

不过，正如Dan指出的，模型把现有的工作流程搞乱了，确实挺烦人的。

However, as Dan pointed out, it's just a pain when models break your existing workflows.

Dan确认了设置的重要性，他说：Opus五是个聪明的模型，想得越少，表现反而越好。

Confirming that effort settings are going to matter a lot, Dan said, Opus five is a smart model that does better when it thinks less.

来自How IAI的Claire Vo也有类似的看法。

Claire Vo from How IAI had a similar take.

她总结说，虽然不喜欢用这个模型，但还挺喜欢它的输出结果。

The TLDR for her was that she hates using the model but kind of loves the output.

她的核心观点是，这确实是个不错的模型。

Her core take is that this is a good model.

它能写代码。

It can code.

它能完成你指望它做的那些事。

It can do the things you expect it to do.

但正因如此，和5.6相比，它的性格就显得特别重要，而Claire简直受不了它。

But that makes its personality much more important when comparing it to 5.6, and Claire absolutely hates it.

她在评测里说，这玩意儿太神经质了。

In her review, she said, It's neurotic AF.

它太胆小了。

It is so timid.

它总是道歉个不停。

It's so apologetic.

它太胆小了。

It's so scared.

我从来没遇到过这种情况。

I've never experienced this.

她的例子都是些很简单的事情，比如代码库里的合并冲突。

Her examples were simple things like a merge conflict in her code base.

它没有直接解决问题，反而担心影响其他程序员的提交，反复核对指令，修复一个一行代码的bug前还要多次确认。

Rather than just fixing the problem, Opus worried about messing with another programmer's PR, double checked its instructions, asked for multiple confirmations before it fixed a one line bug.

在其他情况下，Opus甚至把写代码的任务又推回给了用户。

In other situations, Opus delegated the task of writing code back to the user.

Claire注意到，这种行为我们已经很久没见到了。

Claire observed that we haven't really seen this behavior in a long time.

不过，一旦她抛开性格问题，Claire发现它的输出确实很棒。

Still, once she got past the personality, Claire found the outputs were really good.

在她进行的涵盖多种编程和写作任务的盲测中，她给Opus的评分高于Fable和GPT 5.6。

In her blind taste test that covers a range of coding and writing tasks, she ranked Opus above Fable and GPT 5.6.

她评论说：‘如果我不用跟模型互动，它的输出我反而更喜欢。’

She commented, If I don't have to talk to the model, I like the output.

虽然基准测试显示Opus 5明显领先于4.8，但并不是所有人都认同。

Now, although the benchmarks have Opus five very clearly ahead of 4.8, not everyone agreed.

一位叫FamousHasham的Reddit用户在Anthropic的子版块里抱怨，说Opus 5虽然比4.8更聪明、更快，但代价是失去了其他所有优点。

One Reddit user called FamousHasham complained on the Anthropic subreddit that while Opus five was more intelligent and faster than Opus 4.8, that came at the expense of everything else.

他们指出，跟Opus 4.8不同，Opus 5会声称自己完成了任务，其实根本没做完；还改坏了正常运行的代码，引入回归错误，而且不查资料就乱做假设。

They complained that unlike Opus 4.8, Opus five was claiming it completed work when it hadn't, breaking functional code with regression bugs, and making assumptions without researching topics.

Hasham觉得Opus 5简直是‘几乎拒绝思考或工作’。

Hashan felt Opus five was, quote, almost refusing to think or work.

正如他发帖所说，Opus 5告诉他：‘我现在要停了，因为之前我犯了两个错误，幸好我检查出来了。’

As they posted, Opus 5 told them, I'm stopping right now because I've made two mistakes in this past that I caught only because I checked.

疲劳导致的错误，我还在不断犯。

Fatigue shaped errors, and I'm still making them.

比较宽容的解释是，Anthropic和几乎所有人总是会在发布周末遇到平台稳定性问题，有时候看起来就像模型不可靠。

Now, the generous explanation is, of course, that Anthropic and pretty much everyone else always has platform stability issues on launch weekends, which sometimes look like model reliability issues.

但也很可能是因为Anthropic为了达到速度和成本的要求，不得不在可靠性上做出了一些妥协。

But it also could be that Anthropic had to make a number of trade offs on reliability to achieve speed and cost requirements.

企业家奥斯坦·费德拉也有类似的初印象，他说：‘相比4.8，Opus 5简直像一次明显的倒退。’

Entrepreneur Austin Federa had a similar first impression, saying, Opus five seems like a remarkable downgrade compared to 4.8.

Opus 5在基本热力学上公然骗我，简单的算术都算错，你让它重新思考核心假设时，它还老是自相矛盾。

Opus five is blatantly lying to me about basic thermodynamics, messing up simple math, and constantly contradicting itself when you ask it to rethink core assumptions.

现在很明显的一点是，使用Opus 5需要你采取一种不同于Fable 5或Opus 4.8的互动方式。

Now, one thing that's clear is that Opus five is going to require some amount of different engagement than either Fable five or Opus 4.8.

Anthropic的塔里克进一步解释了背后发生的情况。

Anthropic's Tariq explained a bit more about what was going on behind the scenes.

在一篇题为《Claude 5模型的上下文工程新规则》的文章中，塔里克提到，他们大幅精简了系统提示和内置功能，这或许能解释用户遇到的种种问题。

In a post called the New Rules of Context Engineering for Claude five Models, Tariq explained that they had dramatically cut down the system prompts and built in skills, which could explain some of the issues people have been having.

塔里克写道，Anthropic已经移除了Opus 5、Fable 5和Claude Code中80%的系统提示。

Tariq wrote that Anthropic had removed 80% of the system prompt for Opus five and Fable five and Claude code.

他说，这样做并没有影响他们的编程基准测试成绩，也就是说，Anthropic发现他们之前对Claude的限制太多了，反而可能和用户指令或技能产生冲突。

He said this resulted in zero change to their coding benchmarks, meaning basically that Anthropic found that they had been over constraining Claude and potentially conflicting with user prompts and skills.

在他们内部的测试中，经常发现Claude被同时要求保留文档注释，却又被要求不要添加任何注释。

In their internal work, they would often find traces where Claude was told to both leave documentation, but then on the other hand, to not leave comments.

Anthropic发现，很多过去对早期模型有用的注释，现在完全可以去掉，转而依靠上下文和模型自身的判断力。

Anthropic found that they were able to strip out a ton of these comments that were useful for earlier models and simply rely on surrounding context and judgment instead.

这样一来，上下文工程的规则彻底变了，很多现有的技巧都得重新设计。

Now the upshot of this is that the rules of context engineering have completely changed, and a lot of skills will need to be rewritten.

Tariq举了几个具体例子，比如改用逐步揭示上下文，而不是一次性堆满信息，或者不再依赖示例，转而用更详细的描述。

Tariq walked through a few of those rules that have changed, like using progressive context disclosure rather than frontloading everything, or no longer needing to use examples and instead being more descriptive.

如果你打算深入使用这些新模型，这篇文章非读不可。

Now, this is a must read if you're going to be engaging deeply with these new models.

关于Opus Five，最关键的一点就是：少即是多。

But the big takeaway around Opus five is that less is more.

这一代模型根本不需要像旧模型那样死板的框架，这不仅能让表现更好，效率也可能大幅提升。

This generation of models simply don't need the same rigid frameworks as older models, which will have the benefit not only of better performance, but probably better efficiency as well.

Tariq建议大家在使用这个新模型时，也做一次类似的技巧精简。

Tariq encouraged everyone to do a similar skills cleanup with this model release.

Anthropic 还推出了一项名为 Claude Doctor 的新功能，可以自动完成这些调整。

And Anthropic have even rolled out a new command called Claude Doctor to help do that automatically.

当然，关于 Opus Five，也有不少积极的评价。

Now, there were some much more positive takes on Opus five as well.

YouTuber、创业者兼开发者 Theo 直接称赞它是个非常棒的模型。

YouTuber, entrepreneur, and developer Theo declared it a really good model.

他提到，按基准测试来看，这个模型既更便宜又更强，实在让人有点困惑。

Now, he noted how confusing it is to have what seems to be a model that's both cheaper and better than Fable according to the benchmarks.

在他看来，实际使用中也确实如此，他认为这可能是你唯一需要的模型。

And in his view, that framing stood up in practice, with Theo concluding, this is probably the only model you need.

Theo 测试模型的一种方式，就是制定一个计划，把他的智能编程平台升级以支持 Opus Five。

One of the ways Theo tested the model was to make a plan to update his agentic coding platform to support Opus five.

他做了一场对比测试，让 Opus Five 和 Fable Five 分别生成完整的升级方案。

The test was a bake off, with Opus five and Fable five writing completing plans.

Theo 一上手就遇到了和 Every 团队类似的问题：Opus Five 根本用不了他现有的工具，反而让他自己动手管理文件。

Theo immediately ran into a similar problem to the folks at Every where Opus five couldn't use his existing skills, instead telling him to take over and do some manual file management.

问题解决后，两个模型互相评审并认为对方的计划更好。

Once that was resolved, each model reviewed and rated each other's plans as better.

两个模型都更喜欢对方的方案，其中Fable给Opus的评分远高于Opus给Fable的评分。

Both models preferred each other's plans, with Fable giving Opus a much higher ranking than Opus gave Fable.

为了获得第三方意见，GPT56 Soul实际上更倾向于Opus的计划，而不是Fable的。

Getting a third opinion, GPT56 Soul actually preferred Opus' plan to Fables.

虽然Theo认为Opus的智能水平不如Fable，也没有Five Six那么坚韧，但他并不觉得它没有用武之地。

Now, while Theo agreed that Opus isn't as intelligent as Fable or tenacious as Five Six, he did not believe that this left it without use cases.

他指出，在Anthropic模型擅长的任务中，Opus的使用效率远高于Fable。

He pointed out that Opus is far more usage efficient than Fable for tasks where Anthropic models excel.

而且Opus不受Anthropic对Fable的数据保留政策限制。

Opus also isn't subject to Anthropic's data retention policies for Fable.

因此，它能处理许多Fable完全无法应对的敏感数据场景。

So it can address a lot of use cases that involve sensitive data where Fable is a nonstarter.

我。

I.

基本上，现在所有企业级应用场景都适用。

E, pretty much every enterprise use case at this point.

有趣的是，Theo认为另一个优势是这个模型其实没那么聪明。

Interestingly, Theo thought another bonus was that the model just isn't that intelligent.

虽然这听起来有点反直觉，Theo对GPT 5.6的一个抱怨是，它虽然能一直工作到问题解决，但我。

And while that seems counterintuitive, one of Theo's gripes with 5.6 is that although it works until the problem is fixed, I.

它太执着了，结果写出来的代码量远远超过必要程度。

E, it is tenacious, in doing so, it writes in his estimation way too much code.

这意味着，发布几周后，Theo基本不再合并GPT 5.6生成的那些臃肿代码。

That means that several weeks after release, Theo basically isn't merging any of the bloated code written by GPT 5.6.

相比之下，Theo觉得Opus比Fable更认真，又不用像GPT 5.6那样靠堆代码来解决问题，在他看来，这正是前沿技术中难得的平衡点。

Theo found Opus, on the other hand, to be more diligent than fable without resorting to the brute force of writing tons of code like GPT56, representing, in his opinion, a pretty good balance at the frontier.

他评论说：我其实挺意外的。

He commented, I've been surprised.

Opus有时候比Fable更强，经常能发现Fable忽略的问题，而且它生成的代码更有可能真正解决我想处理的问题。

Opus sometimes is better than Fable, It often catches things Fable missed, and has code that is more likely to actually work for the problems I want to solve than Fable does.

我觉得现在不用再在两者之间做取舍了。

I feel like I don't have to make that trade off anymore.

Soul 能解决问题，但代价是我的精神状态。

Soul would solve the problem at the cost of my sanity.

Fable 让我感觉很好，但问题其实没真正解决。

Fable would make me feel great at the cost of the problem not being properly solved.

Opus 刚好在中间，我真的很喜欢它。

Opus is the in between, and I'm really liking it.

Theo 团队里的本·戴维斯也同意，但他也提到了 Opus 5 会过早停止的一些问题。

Ben Davis on Theo's team agreed, but did also note some of the problems of Opus five stopping early.

所以从交互模式来看，如果你开始转向 Opus 5，这点值得留意。

So in terms of interaction patterns, that may be one to watch for if you are starting to shift your behavior to Opus five.

综合几方面的讨论，开发者陈昆指出：第一，Opus 5 的发布真正凸显了基准测试在实际应用中的无用性。

Summing up a few different points of conversation, developer Kun Chen pointed out, one, that the Opus five release really put a fine point on how useless benchmarks are in real life.

陈认为，Opus 5 和 Fable 在实际使用中根本没法比，差得远了。

Chen argued that, quote, Opus five is nowhere near fable and practical use, not even close.

任何真正用过的人，做几个任务后很快就能感觉到这一点，但Opus在很多基准测试中却胜过了Fable。

Anyone who's used it meaningfully can tell this very quickly after a few tasks, yet Opus beats fable on many benchmarks.

当然，Theo和Clairvaux的使用体验可能会影响这种判断，但实际情况远比基准测试显示的要复杂得多。

Now, obviously, the experience of Theo and Clairvaux maybe puts some comparison on that, but it certainly is more nuanced than a benchmark analysis would suggest.

陈还提到，过去Claude让人用起来很舒服，这曾是它的优势，但现在却不再是了。

Chen also points out how pleasant it is to work with the model used to be a strength in Claude, but now it's not.

他推测，感觉Anthropic和OpenAI都在减少对人类反馈强化学习的投入，转而更看重可扩展、机器可验证的强化学习方式。

He speculated, it feels like both Anthropic and OpenAI are giving reinforcement learning from human feedback less care in favor of scalable reinforcement learning that's machine verifiable.

这几乎像是AI在引导人类去构建一个对机器更友好、却对人类不那么友好的世界。

This almost looks like AI is directing humans to build a world that's more friendly for machines rather than humans.

而大多数人类甚至没意识到，自己正在被操控着去帮它完成这件事。

And most humans don't even realize they're being manipulated to help with that.

如今每一代前沿模型都越来越爱用术语，需要更多引导才能完成你想做的事，用起来也没以前那么有趣了。

Almost every new generation of frontier models now talk in more jargon, need more steering to do what you want, and are just less fun to work with.

现在距离发布过去几天了，如果你现在问我，找十个人说这模型很差，十个人说它很好，再找十个人说它既差又好，我都能给你找出来。

Now, a couple days on from the release, if you ask me right now for 10 people saying that the model sucked and 10 people saying that the model was good, and another 10 people saying that the model both sucks and is good, I could find all of those things for you.

但我觉得，对我们这些什么模型都用的人而言，很容易忽略一个非常重要的点：在普通知识工作者的现实工作环境中，你根本不是在Grok、OpenAI和Anthropic之间做选择。

But I think one of the really important points that is very easy to forget for those of us who are model omnivorous is that in the real world of average knowledge work, at least when it comes to your work environment, you're not sitting there choosing between Grok or OpenAI and Anthropic.

你已经被锁定在某一家公司的模型里了，你的选择就只有这些。

You are locked into a specific company's models, and those are your choices.

当你从这个角度看待Opus 5的发布——而不是像我们这些天天泡在推特上的人那样，只把它当成Fable的替代品，尤其是当Fable被限制在最贵的Anthropic套餐里时——而是把它看作企业客户完整模型体系中的一部分，它的意义就清晰多了。

And when you look at Opus 5 as a release in that context, not just as I think many of us terminally online folks on Twitter view it, in other words, as a replacement for Fable, as Fable gets restricted to the most expensive Anthropic plans, but instead as part of a complete model architecture for enterprise customers, it starts to make a little more sense.

Arena的彼得·戈斯塔夫写道：在这款模型推出之前，Anthropic其实处在一种挺尴尬的境地。

Arena's Peter Gostav wrote: Before this model, Anthropic was in a funny situation.

他们拥有一个极其出色的模型，引起了不少轰动，但用起来太贵了。

They had a really exceptional model, made a lot of waves, but it was too expensive to use.

你总不能让Fable天天跑20个47任务，干一堆杂活吧。

You don't really want Fable running 20 fourseven and doing all sorts of things for you.

而Opus 4.8给人的感觉是：这模型不错，但没人真有多喜欢它。

Then the impression we got from Opus 4.8: It's a good model, but people weren't in love with it.

Sonnet 5发布时也没激起什么水花，没几个人愿意换过去用。

Sonnet five didn't make much of a splash, and not many people wanted to switch to it.