ChatGPT 刚刚成为工作代理

今天《AI每日简报》带来最新模型动态，还有OpenAI的重大工具更新。

Today on the AI Daily Brief, more new models plus a big harness update from OpenAI.

在那之前，先看下头条新闻：Cursor似乎也在开发一款工具，瞄准更大的知识型工作市场。

And before that in the headlines, Cursor also appears to be developing a harness to go after the larger knowledge work sector.

《AI每日简报》是一档每日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。

All right, friends.

在正式开始之前，先说几件事。

Quick announcements before we dive in.

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

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

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

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

想了解如何赞助本节目？请访问a idailybrief.ai/sponsors，或发送邮件至sponsors@aidailybrief.ai。

And, of course, to learn more about sponsoring the show, head on over to a idailybrief.ai/sponsors or send us a note at sponsors@aidailybrief.ai.

既然我们本期的重点是关于一个重要的辅助工具更新，那我们先从这条新闻开始：Cursor正在开发一款通用型智能代理，打算和Claude Cowork一较高下。

Quite appropriately, given that our main episode is about a big harness update, we kick off our headlines with news that Cursor is planning a general purpose agent to compete with Claude Cowork.

报道称，该项目早在四月就已启动，就在Cursor与SpaceX达成合作后不久。

The information reports that work began on the project in April, shortly after Cursor signed their deal with SpaceX.

这款代理预计将使用Grok 4.5，也是Cursor首次面向非专业程序员群体推出的产品。

The agent is expected to use Grok 4.5 and will be Cursor's first project aimed at anyone other than professional coders.

它名叫SAND，目标是充当个人助手，处理日常办公任务，比如回复邮件、操作电子表格等。

Called SAND, it's designed to function as a personal assistant performing standard office tasks like dealing with email or working with spreadsheets.

据透露，它未来还可能演变成一个统一平台，因为有消息说，这款代理同样能胜任AI编程任务。

It sounds like it could also eventually become a unified platform, with the information's reporting suggesting that the agent will also be functional at AI coding.

消息人士称，该平台已在六月内部上线，但目前还不清楚它是否会获得公开发布的批准，以及具体时间会是什么时候。

Sources said the platform was rolled out internally in June, however it's still unclear whether it will get the green light for a public release, or when that would happen.

毫无疑问，这款产品表明Cursor正有意突破其传统的软件工程师用户群体。

Certainly, the product suggests that Cursor is looking to grow beyond their traditional user base of software engineers.

这很合理，接下来我们会看到，今天OpenAI所有发布的核心主题，都是把在编程领域验证成功的东西，推广到更广泛的知识型工作中去。

This makes sense as we'll see, a huge theme of all of the announcements today with OpenAI are all about taking what has been working in coding bringing it to a broader set of knowledge work.

说到编程方面有效果的东西，OpenAI 抓住了当下热点，直接宣布主流编程评测基准根本不靠谱。

Now speaking of what's working with coding OpenAI has captured the zeitgeist and declared that the leading coding benchmark is bunk.

在一份新报告中，OpenAI 对 SuiteBench Pro 进行了审查，发现这个基准问题一大堆。

In a new report, OpenAI audited SuiteBench Pro and found the benchmark to be sorely lacking.

他们的测试显示，基准里有三成任务根本没法用，因此正式撤回对它的支持。

In their testing, they found that 30% of the tasks on the benchmark were broken, and are now formally retracting their support of the benchmark.

很多问题出在部分任务是公开的，导致这些题目被提前放进训练数据，严重扭曲了结果。

Many of the issues stem from some of the tasks being public, which can distort results, by having those specific problems be included in training data.

还有些任务要求藏得太深、指令自相矛盾、测试标准太死板，或者评分标准不完整。

Others had hidden requirements, contradictory instructions, overly strict tests, or incomplete grading criteria.

他们的结论是：SuiteBench Pro 已经没法准确衡量前沿编程能力了。

Their conclusion was that SuiteBench Pro no longer reliably measures frontier coding capability.

要说明的是，其实大家早就开始远离 SuiteBench 了。

And to be clear, the shift away from SuiteBench was already well underway.

Cursor 已经用了好几个月自己的私有评测标准，而 Cognition 和 Databricks 这周也推出了各自的评测体系。

Cursor has been using their own proprietary benchmark for months, while Cognition and Databricks also launched their own benchmarks this week.

我觉得我们现在正式进入了一个阶段，各家公司将纷纷推出自己的新基准，包括我敢打赌他们还是会拿SuiteBench Pro出来展示，但主要就是等大家凭感觉去判断这些基准到底靠不靠谱。

I think we're officially at the point where there's going to be lots of introduction of new benchmarks companies are going to present all of them, including I guarantee they will still present SuiteBench Pro and mostly just wait for people to have the vibes that confirm whether the benchmarks are bunk or not.

我们再稍微聊一下OpenAI，不过换个完全不同的方向——这家公司刚刚发布了一份新声明，阐述了他们对政府和军方合作的态度。

Now staying on OpenAI for a minute but going to a very different area, the company has published a new statement that explains their approach to government and military partnerships.

在公布新的国家安全原则时，OpenAI写道：我们相信，民主社会应当能够利用人工智能来保护民众、保卫关键基础设施、提供公共服务，并应对新兴威胁。

Announcing their new national security principles, OpenAI writes: We believe democratic societies should be able to use AI to protect people, defend critical infrastructure, deliver public services, and respond to emerging threats.

但随着AI系统能力不断增强，部署时必须强化民主问责制、真正的人类判断和法治，增强民主制度，而不是集中权力。

But increasingly capable AI systems must be deployed in ways that reinforce democratic accountability, meaningful human judgment and the rule of law, and strengthen democratic institutions rather than concentrate power.

把这些原则提炼成几个关键点：OpenAI明确表示，他们不会支持将技术用于大规模国内监控、在缺乏适当人类判断与问责的情况下做出高风险决策，或用于规避法律义务、监管和问责的用途。

Distilling the principles into a few key points: OpenAI explicitly stated that they will not support the use of their technology for mass domestic surveillance, high stakes decisions including decisions over the use of force without appropriate human judgment and accountability, or uses that evade legal obligations, oversight, and accountability.

这些内容你可能觉得眼熟，因为和Anthropic的红线几乎一模一样，而之前Anthropic那套说法已经让政府乱了阵脚，所以现在这份文件到底想达到什么目的，还真不好说。

Now, these you might recognize are essentially the same as Anthropic's red lines, and given that that created such chaos with the government, it's not clear what the goal of this document is supposed to be.

不过至少，我们总算清楚地知道了他们坚持什么立场，也为未来和政府的对话打下了基础。

But I guess at least we have a clear articulation of what they stand for, and something to build from in their conversations with the government.

说到政府和AI的交集，Anthropic刚刚任命前美联储主席本·伯南克加入他们的长期公益信托董事会。

Speaking of the overlap between the government and AI, Anthropic has appointed former Fed Chair Ben Bernanke to the board of their long term benefit trust.

伯南克当然是由小布什政府在2006年任命为美联储主席的。

Bernanke, of course, was appointed Fed Chair by the W.

他任职到2014年，全程主导了全球金融危机爆发前后的各项应对措施。

Bush administration in 2006 and served until 2014, presiding over the events that led up to and followed the global financial crisis.

他是近年来最具争议的美联储主席之一，人们对他的评价两极分化：有人觉得2008年的银行救助拯救了世界经济，也有人认为那是加剧美国贫富差距的原罪。

He is one of the more controversial Fed shares in recent history, with views on him differing depending on whether you think the bank bailouts in 2008 saved the world from a depression, or you view them as the original sin that accelerated the wealth divide in The United States.

当克劳德被要求客观评价伯南克时，它回应说，他总体上备受尊重，但左右两派都有批评者，认为他代表了一种缺乏问责的技术精英体制，只为权贵利益服务。

When Claude was asked to provide an objective viewpoint on Bernanke, it responded that he's generally well regarded but has critics on both the left and the right who view him as emblematic of an unaccountable technocracy protecting elite interests.

那么，Anthropic的长期公益信托到底是什么？

So what then is Anthropic's long term benefit trust?

他们自己说：Anthropic是一家公益企业，成立的初衷就是在商业成功与创造社会公共利益之间取得平衡。

In their own words, they write: Anthropic is a public benefit corporation, meaning the company was created to balance commercial success with generating social and public good.

这个信托基金的作用，是帮助公司长期稳健地维持这种平衡，对Anthropic开发和部署人工智能形成监督制约。

The benefit trusts exist to help the company responsibly maintain that balance over the long term, providing a check on how Anthropic develops and deploys AI.

说白了，这就是一个独立的监督与咨询机构，地位高于Anthropic的常规董事会。

Essentially, this is an independent oversight and advisory board that sits on top of Anthropic's normal board of directors.

重要的是，这个受益信托有权选举或罢免公司董事会的一名成员，随着时间和资金达到特定里程碑，这一权力将逐步扩大到两名，再增至三名。

Importantly, the benefit trust has the power to elect or remove one member of the corporate board, which escalates to two and then three according to time and funding based milestones.

到明年，他们将对董事会拥有多数控制权，尽管股东仍可通过超级多数投票否决信托的决定。

By next year, they will have majority control over the corporate board, albeit with a shareholder override that requires a supermajority vote against the actions of the trust.

与普通公司董事会不同，信托董事会的任何成员都不允许持有公司股份。

Unlike normal corporate boards, no one on the board of the trust is allowed to be a shareholder.

关于这一点，尤其是FinTwit上讨论很多，但我认为汤姆·布鲁尼点出了我们很多人关心的核心问题：他会像压低利率那样，把代币价格也压到零吗？

There is a lot of discourse on this one, especially from folks on FinTwit, but I think Tom Bruni got at the key question that many of us are asking: Will he bring token prices to zero like he did with rates?

再来看芯片领域，Meta计划在加拿大投资100亿美元，建设下一个大型数据中心。

Moving over to the chips world, Meta is heading north for their next big data center project with a $10,000,000,000 facility in Canada.

Meta于周三宣布了在阿尔伯塔省的破土动工，计划在未来几年内建成总容量达1吉瓦的设施。

Meta announced the groundbreaking for the Alberta site on Wednesday, with plans to build a total of one gigawatt of capacity over the coming years.

有趣的是，社区承诺在公告中占据了核心位置，Meta计划投入6000万加元用于当地基础设施升级，包括道路和供水系统。

Interestingly, community pledges were front and center in the announcement, with Meta planning to contribute $60,000,000 in Canadian dollars to local infrastructure upgrades, including roads and water.

在电力方面，Meta承诺承担全部基础设施成本，当地官员预计，这一项目反而会促使电价下降。

On electricity, Meta pledged to pay full infrastructure costs, and officials expect local rates to actually decrease as a result of the project.

他们还会直接向当地非营利组织提供资金支持。

They will also provide direct funding to local nonprofits.

此外，他们在施工高峰期计划雇佣3000名建筑工人，并长期保留300个运营岗位。

In addition, they plan to hire 3,000 construction workers at the peak of the build and support 300 ongoing operational jobs.

我一直在大声疾呼：这类社区投入，对当地来说价值巨大，而对这些数据中心的预算而言却只是九牛一毛，简直是科技巨头与所在社区达成共赢最简单的方式——只要他们愿意花点时间真正去做。

Now, as I continue to scream from my bully pulpit, these sort of community contributions can be on the one hand incredibly valuable to those communities, and are a rounding error in these data center budgets, making it one of the easiest potential alignments between hyperscalers and the communities they operate in than they could possibly imagine if they only take the time to actually do it.

看到Meta真的走上了这条路，真是让人欣慰。

So it's great to see Meta actually going down that path.

不过，这里另一个更重要的信号是，Meta的AI基础设施建设根本没打算放缓。

Now, the other big implication here is, of course, that Meta's big AI build out is not slowing down.

上个月，当Meta计划出售多余算力的消息传出时，很多人以为他们会削减资本支出。

When news broke last month that Meta was planning to sell extra capacity, many drew the implication that they'd wind back on CapEx.

现在又追加十亿瓦的建设规模，说明他们的步伐依然飞快。我想等你听完正片，就会慢慢明白背后的原因了。

Adding another gigawatt suggests Meta's plans are still ticking along at a rapid pace, and I think when you get to the main episode, you'll kinda start to understand a bit more why.

与此同时，Meta自研芯片项目已经重获生机，预计在未来几个月内投入量产。

Meanwhile, Meta's in house chip program is back from the dead and set to enter production in the coming months.

据路透社援引的一份内部备忘录显示，Meta预计将在九月开始生产首批芯片。

According to an internal memo cited by Reuters, Meta is on track to begin producing the first chips in September.

报道称，至少有一款芯片设计仅用六周就顺利通过了测试阶段。

The report states that at least one chip design sailed through the testing phase in just six weeks.

Meta正与博通合作设计芯片，并将委托台积电生产处理器、三星生产内存。

Meta is working with Broadcom on design and will contract with TSMC for processors and Samsung for memory.

大多数半导体分析师原本认为，这一代芯片也会被搁置，就像之前的那些一样。

Most semiconductor analysts assume this generation of chips would be shelved, just like all the others.

自2022年至少以来，他们就在自主研发先进的AI芯片，但每一代都遇到障碍，从未投入大规模生产。

Has been working on their own advanced AI chips since at least 2022, but each iteration hit roadblocks was never put into full scale production.

直到现在，这一代芯片看起来也和之前没什么两样。

This generation had a similar vibe up until now.

它最早在三月公布，距离Meta取消上一代芯片仅一个月，当时宣布了该芯片家族包含四种配置，分别针对不同的推理和训练任务优化。

It was first announced in March, just one month after Meta had canceled the previous generation, with that announcement describing four configurations across the chip family tuned for different inference and training workloads.

上个月有消息说，Meta曾要求三星暂时暂停定制芯片的开发，很多人因此认为这标志着项目即将终结。

Then last month, it was reported that Meta had asked Samsung to temporarily pause development on their custom chipset, and many assumed that was the beginning of the end.

相反，现在看来Meta已经定下了最终设计，只等与芯片合作伙伴安排好生产排期了。

To the contrary, it now seems that Meta has a finalized design, and just needs to wait for some foundry time with their chipmaking partners.

这些芯片预计会部署在Meta的数据中心，帮助他们减少在英伟达和AMD上的开支。

Chips The are expected to be installed in Meta's data centers and allow them to cut down on their spending with NVIDIA and AMD.

Meta还计划从明年开始每六个月推出一款新芯片，这比行业常规速度要快得多。

Meta also plans to design a new chip every six months, beginning next year, which would be significantly faster than the standard industry cadence.

路透社看到的这份备忘录也再次确认了Meta扩大算力的计划。

The memo viewed by Reuters also reaffirms Meta's plan to scale compute.

公司表示，今年计划部署7吉瓦的算力容量，并在2027年将这一速度翻倍。

The company said that they plan to deploy seven gigawatts of capacity this year and double that pace in 2027.

是啊。

So yeah.

看来Meta短期内不会削减投入了。

Doesn't seem like we're getting a cutback from Meta anytime soon.

不过话说回来，为什么会有这样的变化？我觉得咱们还是先放下标题，进入正片吧。

But again, for why, I think we need to end the headlines and move on to the main episode.

现在最重要的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更擅长写提示词，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.

他们会明确问题、引导思考、反复打磨，不断追求更好的答案。

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

好消息是？

And the good news?

这些行为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，提升到真正掌握AI能力，普华永道关于高级AI协作的研究非常值得一看。

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

了解更多，请访问 kpmg.com/us/sophisticated。

Learn more at kpmg.com/us/sophisticated.

网址是 kpmg.com/us/sophisticated。

That's kpmg.com/us/sophisticated.

现在企业级AI领域一个挺有意思的变化是，讨论焦点正迅速转向基础设施和运维。

One of the more interesting shifts in enterprise AI right now is how quickly the conversation is moving towards infrastructure and operations.

随着AI进入核心工作流程、受监管的数据环境和智能代理系统，企业迫切需要一套有管控的基础设施和推理能力，能够日复一日稳定运行，并从一开始就内置明确的运维责任机制。

As AI moves into core workflows, regulated data environments, and agentic systems, enterprises need governed infrastructure and inference that can operate reliably day to day, with clear operational accountability built in from the start.

当这些系统规模扩大后，运维模式本身也越来越成为AI战略的核心部分。

As those systems scale, the operating model increasingly becomes part of the AI strategy itself.

Rackspace Technology 负责管理整个企业级AI栈，覆盖私有云、混合云和边缘环境中的智能代理与底层设施。

Rackspace Technology is the operator of the full enterprise AI stack, from agents to infrastructure across private cloud, hybrid cloud, and edge environments.

Rackspace 为那些对数据主权、合规性和系统可用性有严格要求的组织，构建并运维受管控的AI基础设施、推理服务和生产级AI系统。

Rackspace builds and operates governed AI infrastructure, inference, and production AI systems for organizations where sovereignty, compliance, and uptime are non negotiable.

因此，他们的工程师在部署后仍长期驻场，协助将AI真正融入日常运营并保障稳定运行。

Therefore, deployed engineers stay embedded beyond deployment to help operationalize and run AI in live environments.

想了解更多企业级AI的落地场景和规模化成果，请访问 rackspace.com。

To learn more about where Enterprise AI runs and outcomes scale, go to rackspace.com.

Blitzy 正帮助大型企业将工程效率提升五倍以上。

Blitzy is driving over 5x engineering velocity for large scale enterprises.

一家上市公司保险公司借助Blitzy开发了一个定制的支付处理应用，原本预计需要十三个月的项目，最终只用了六周就上线运行了。

A publicly traded insurance provider leveraged Blitzy to build a bespoke payments processing application, an estimated thirteen month project, and with Blitzy, the application was completed and live in production in six weeks.

一家上市的垂直领域SaaS公司用Blitzy从一个五十万行代码的单体系统中拆分服务，速度比原先预估快了21倍，而且全程没有影响生产环境。

A publicly traded vertical SaaS provider used Blitzy to extract services from a 500,000 line monolith without disrupting production 21 times faster than their pre Blitzy estimates.

这些都不是实验性的尝试。

These aren't experiments.

这就是2026年全球最具创新力的企业交付软件的方式。

This is how the world's most innovative enterprises are shipping software in 2026.

你可以在《现代CTO》或《CIO分类播客》里直接听到其他财富500强CTO们对Blitzy的评价。

You can hear directly about Blitzy from other Fortune 500 CTOs on the Modern CTO or CIO classified podcasts.

想了解Blitzy如何改变你的软件开发生命周期？去blitzy.com预约一位AI解决方案顾问吧。

To learn more about how Blitzy can impact your SDLC, book a meeting with an AI Solutions Consultant at blitzy.com.

网址是blitzy.com。

That's blitzy.com.

本集《AI每日简报》由Hyperagent赞助播出，你可以在这里管理由团队共同协作的智能代理集群。

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.

别再等着本地代理或聊天流程在你的笔记本上被唤醒了。

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

Hyperagent 将代理部署在云端，7x24小时运行，直接使用你们团队已有的工具完成实际工作。

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

市场部门的代理会把竞争对手的动态自动生成落地页。

Marketing's agent turns competitor moves into landing pages.

销售部门的代理负责完善线索、起草邮件，并更新客户关系系统。

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

运维代理自动跟进文书工作，追踪预算使用情况。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享信息，并严格遵守你们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工作的代理了。

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

立即在 Hyperagent 上雇佣你的AI助手，由 Airtable 原班团队打造。

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

前往 hyperagent.com/aideallybrief 领取你的1000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.

这周，显而易见，所有焦点都集中在新模型上。

This week, obviously, the entire story has been new models.

我的意思是，一波接一波，根本停不下来。

I mean, it has just been a barrage.

GPT-5.6 全系列终于正式发布，所有人都能用了，相关公开信息也一并公布了。

We've got GPT-5.6, the entire family, which are finally officially out, available for everyone with all the related public release information.

还有意外登场的 Grok 4.5，以及突然冒出来的 News Spark 1.1，连 Meta 都被重新拉回了人们的讨论中。

But then also the fairly unexpected Grok 4.5 as well as a new one, News Spark 1.1, out of left field that actually has people talking about Meta again.

但在大家已经越来越清楚——模型本身固然重要，但支撑它的系统更关键的这一年，本周压轴的重磅消息，恰恰是 OpenAI 对使用模型的底层系统做出的重大调整。

And yet, in a year where people have come to understand just how significant not only the model is, but the harness in which it operates, it is so appropriate that we are also closing out the week with, in some ways, the biggest news being big changes to the OpenAI harness through which you are going to use the model.

OpenAI 推出了 ChatGPT Work，我们待会儿在初体验环节聊聊这意味什么，不过首先，咱们先来补一补刚刚正式发布的 GPT-5.6 系列模型。

OpenAI has released ChatGPT Work and we are going to talk about what that means in First Impressions, but first let's finally catch up on the now fully released GPT-5.6 model family.

这是 OpenAI 第一次把他们的模型家族分成不同的类别。

Now, this is the first time that OpenAI has split their model family into a set of classes.

除了旗舰模型 Sol，GPT-5.6 还推出了中等规模的 Terra 和更小巧、更省钱的 Luna 版本。

In addition to the flagship model Sol, GPT-5.6 also has a mid sized version in Terra and a small cost efficient version in Luna.

这周我们其实已经陆续聊过这些模型了，因为 OpenAI 之前允许一些早期用户分享他们的使用感受，但直到昨天官方正式发布，我们才拿到完整的性能数据。

Now, we have started to talk about these models throughout the week, as OpenAI gave clearance to people to start talking about their early adopter impressions of it, but we just now with the official announcements yesterday got the benchmarks.

有意思的是，这次公布的性能数据方式，跟 OpenAI 过去的做法很不一样。

And interestingly, the benchmark story as it's presented is very different than the way that OpenAI has done this in the past.

萨姆·阿尔特曼发推说：‘这不仅是迄今为止我们最好的模型，也是我们写过的最棒的博客之一。’

Sam Altman tweeted, Obviously the best model we've ever produced, but also one of the best blog posts we've ever produced.

原因之一是，他们不再只用一张简单的数字表格来展示性能，而是特别强调用图表来表现性能与成本的关系。

And part of that is that instead of just a simple easy table of a bunch of numbers for their benchmarks, OpenAI is now highly focused on charts that show performance per cost.

比如在代理能力测试中，图表用三种方式呈现：Y 轴是得分，X 轴是某个变量，而最关键的图表是以 API 成本为基准的。

For example, on the agent's last exam test the chart is actually presented in three ways: the score on the Y axis and some other variable on the X axis, with the Premier chart being the API cost.

他们不仅用分数对比了GPT-5.0.5、Claude Opus 4.8和Claude Fable 5，还对比了成本，重点强调GPT-5.6系列不仅表现更好，成本还低得多。

They show the comparison not only to GPT-five 0.5 (CLAUD Opus four point and Claude Fable 5) in terms of the score, but also in terms of the cost with the strong emphasis being that not only do the GPT-five 0.6 series of models perform better, but they do so at a much reduced cost.

你还可以把图表按延迟时间或输出的token数量来排序，这同样是衡量效率的另一个指标。

Now you can also vary that chart to be organized by latency or by output tokens used, which is of course another measure of efficiency.

在人工智能分析编码代理指数上，他们也用了类似的图表，这现在已经成为他们展示基准测试的标准方式。

You see a similar chart presented this way for the Artificial Analysis Coding Agent Index and this is pretty much now the standard for how they present the benchmarks.

萨姆·阿尔特曼在一条公告推文中进一步强化了这一点：我们听到了企业对AI成本的担忧，而5.6 SOL在每项任务的花费上是一个巨大进步，Terra和Luna也是如此。

Sam Altman reinforced this focus with one of his announcement tweets where he said, we have heard enterprises on their concerns about AI costs, and 5.6 SOL is a huge step forward for dollars per task, as are Terra and Luna.

考虑到很多听众还是通过音频收听播客，我们就不一一解释这些图表了，简单来说，GPT-5.6在最高设置下，性能超越了Opus 4.8，接近甚至超过Fable 5，最重要的是，它的成本低得多。

Now rather than trying to explain all of these different charts on a podcast that many, if not most of you, still consume via audio, the big overview is basically that 5.6 SOL on max settings, benchmarks ahead of Opus 4.8, near or above Fable five, and most importantly at a significantly lower cost to each.

在人工智能分析测试中，GPT-5.6只比Fable 5少1分，屈居第二，但它的成本只有Fable的三分之一，甚至比Opus 4.8还便宜40%。

For the Artificial Analysis Index run, GPT-5.6 was a close second to Fable five, falling short by a single point, but it completed the run at a third of the cost of Fable and was even 40% cheaper than Opus 4.8.

而在编码代理测试中，GPT-5.6已经登顶，比Fable 5高出3分。

Meanwhile, on the Coding Agent Index, GPT-5.6 is the new state of the art, beating Fable five by three points.

事实上，至少在人工智能分析这个测试里，中等规模的Tera版本表现和Fable一样好，如果这些结果能反映到实际使用中，它作为日常编程模型会便宜得多。

In fact, at least on that artificial analysis benchmark, the mid sized Tera variant performed at the same level as Fable and would of course be significantly cheaper as a daily driver coding model if those benchmark results translate to real world experience.

现在回到之前备受关注的模型，也就是Fable 5发布期间的那款，西蒙·史密斯指出，5.6 Luna在人工智能分析智能指数上与GLM 5.2表现相当，但成本低了43%。

Now bringing it back to the model that had so much attention in the Fable five shutout period, Simon Smith pointed out that 5.6 Luna actually matches GLM 5.2 on the Artificial Analysis Intelligence Index, and comes in at 43% cheaper.

谈到我们几天前那期节目讨论过的、高效西方模型带来的新市场机会，西蒙继续说：我很高兴开源模型存在，它们推动了前沿实验室不断创新和发布产品，但我认为企业不会仅仅因为成本原因就全面转向开源模型。

Getting at the idea that we were talking about in our episode a couple of days ago of the new market opportunity for more efficient Western models, Simon continued: I'm glad open weights models exist because they push Frontier Labs to innovate and release, but I don't think enterprises will wholesale shift to open weight models strictly for cost reasons.

我觉得前沿实验室会同时兼顾智能和效率，推出不同性能和价格层级的模型，并训练他们最好的模型，懂得在什么时候调用更便宜的模型，以实现成本效益最大化。

I think Frontier Labs will optimize for both intelligence and efficiency, offer models at multiple performance and price points, and train their best models to know how and when to use their cheaper models to maximize impact for cost.

这样一来，就能形成极具吸引力的价值主张，根本不需要为了省钱而转向开源模型。

And that will make for a compelling value proposition that negates the need to shift to open weight models strictly to save money.

正如我之前提到的，自从本周初首批测试者被允许公开分享体验后，我们在过去几期节目中已经讨论了不少第一手反馈。

Now as I mentioned, since early testers were allowed to break their silence from the beginning of the week, we've been able to cover a number of first impressions over the last couple of shows.

简单回顾一下，核心结论是：我们现在有了两款表现截然不同的前沿模型。

And to recap, the big takeaway is that we now have two frontier models that perform very differently.

总体来看，早期共识是：Fable 5适合处理大规模、长时间的任务，尤其是那些需要高度自主性的场景；而GPT-5.6 SOL不仅在日常使用中是巨大升级，在需要深度参与中间决策的大型任务中也同样出色。

Broadly speaking, the early consensus is that Fable five is the big model for massive, long running tasks especially ones where there can be a lot of autonomy while GPT-5.6 SOL is a huge upgrade not only as a daily driver, but for large tasks where you want to be more involved in the intermediate decisions.

进一步印证了这些早期评价，每位CEO丹·希珀写道：5.6强大、快速，价格只有Fable的一半，现在几乎成了我所有任务的默认选择。

Reinforcing some of the early takes, every CEO Dan Shipper wrote: 5.6 is powerful, fast, half the price of Fable, and my default for almost everything.

他提到，5.6在编程方面不如Fable，但也指出Every的编程评测是针对大规模、长时间的代码重构任务，而这正是Fable更擅长的领域。

He noted that 5.6 is not as good as Fable for coding, but also pointed out that Every's benchmark for coding is a massive long running code refactoring task, which Fable is naturally more suited to than 5.6.

有趣的是，在写作方面，他说5.6是Every用过最喜欢的一个模型，认为它比Anthropic的模型更清晰、更简洁。

Fascinatingly, on writing, he said that 5.6 is the model that every has liked more than anything else, arguing that it's clearer and more concise than anthropic models.

在整个评测过程中，Shipper不断提到它的运行速度。

And throughout his review, Shipper continuously comments on the speed.

Fable是个庞大、缓慢但极其强大的模型，而GPT-5.6则快得多，更像是一个能实时互动的合作伙伴，而不是让你启动后就不管、过会儿再回来查看的工具。

Fable is a big, slow, incredibly powerful model, while GPT-5.6 is very fast, making it more of an active collaborator rather than a model you leave running and come back to.

对很多听众来说，最让人振奋的一点是，Shipper说真正的突破发生在知识型工作领域。

In one of the biggest green flags for a lot of you listening, Shipper said that the real leap is around knowledge work.

他说，这是第一个他敢完全托付、让它独立运行整个知识工作流程的模型，而不仅仅是帮着处理单个任务。

Soul, he wrote, is the first model I've trusted to run whole loops of knowledge work, not just help with individual tasks.

它让我从亲自做事，转变成管理那个替我做事的系统。

It has shifted my job from doing the work to tending the system that does it.

过去几个月，你们可能听了很多新术语，比如‘/goals’、‘loops’之类的，但一直不太清楚这些概念在编程之外的知识工作里到底意味着什么。现在Dan和其他人开始在5.6 SOUL上实际体验到这些，确实很值得关注。

Now probably many of you have spent the last couple of months listening to all of these new conventions like slash goals and discussions of loops, not exactly sure how they relate when you move outside of the coding domain and into knowledge work and so the fact that Dan and others are starting to experience some of that with 5.6 SOUL is pretty notable.

他总结说：如果非要我一句话说清楚，我觉得Fable身上那种大模型的味道更浓。

Summing up he wrote: If I really had to put my finger on it, I'd say Fable has way more big model smell.

但要从它身上榨出价值，本身就需要一门本事，而99%的人现在还没达到那个水平。

But that means it's a skill in itself to get value out of it and 99% of people are still not there yet.

GPT-5.6几乎一样强大，但用起来简单、响应快，还相对便宜。

GPT-5.6 is almost as powerful, but it's easy to use, fast, and relatively cheap.

它能让你提前感受到模型工作未来的发展方向。

It should give you an early sense of where model work is going.

Theo用一份作品集的形式做了他的评测，写道：过去一个月，我用GPT-5.6 SOL消耗了超过二十万的token。

Theo presented his review in the form of a portfolio, posting: Over the last month, I burned over 200,000 in tokens with GPT-5.6 SOL.

我做了很多东西。

I built a lot.

如果你听Theo描述他怎么用它来创作，就会再次明白：Fable是那种你放它自己去跑的模型，而5.6则是你需要不断互动的伙伴。

And if you listen to the way that Theo describes his building process, it once again reinforces the idea that whereas Fable is a model that you let it go off on its own to do things, 5.6 is one you interact with.

总的来说，大家对5.6 SOL的评价非常积极。

Now, overall, the sentiment towards 5.6 SOL is extremely positive.

在我看来，某种程度上，它也受益于这样一个事实：在Fable 5被禁之前，OpenAI在最活跃的AI用户群体中一直占据着主导地位。

And it feels to me like, to some extent, it's also being the beneficiary of the fact that heading into the Fable five ban, OpenAI kinda had all the momentum when it came to the most enfranchised AI users.

Codecs 已经逐渐成为大家默认的选择。

Codecs had increasingly become the default harness that people were looking for.

在Fable出现之前，很多人已经从Opus 4.8转向了GPT-5.5。

And before Fable at least, a lot of folks has shifted over from Opus 4.8 to GPT-5.5.

现在我们因为盼着Fable 5回归，最近没怎么提它刚发布时那些限制，但对一些用户来说，这些问题依然存在。

Now given that we've been so excited to have Fable five back, we haven't talked for a while about some of the constraints that were placed on the model when it was announced, but certainly for some buyers, those concerns haven't gone away.

Gurgler Rose发推说：一位在大型AI乐观公司、大量投入AI预算的开发者，对Fable和GPT-5.6 SOL的对比很有意思。

Gurgler Rose tweeted, Interesting take on Fable versus GPT-5.6 sole from a dev at a large and AI bullish company spending lots of money on AI.

他们告诉我：Anthropic对Fable的数据保留政策没变，也就是说，他们的系统会保存我们的数据。

They told me: Anthropic has not changed their data retention policy on Fable, meaning they would store our data.

所以我们没法用。

So we cannot use it.

因此，我们全力转向了GPT-5.6 SOL。

We're going hard on GPT-5.6 SOL as a result.

但正如我开头说的，2026年的故事远不止是模型的故事。

But as I said at the top of this, the story of 2026 is not just a model story.

事实上，在很多方面，这更是一个工具生态的故事。

In fact, in many ways, it is equally if not more a harness story.

随着我们进入智能代理时代，围绕模型构建的系统——让模型能调用工具、与其他模型互动、启动子代理、获取上下文等等——

As we move into the agentic era, the systems that we put around our models to allow them to access tools, interact with other models, spin up sub agents, access context, etc.

其重要性丝毫不亚于模型本身。

Are every bit as important as the underlying model themselves.

很明显，OpenAI之所以早早允许人们讨论GPT-5.6，就是想让官方发布当天的焦点放在新界面‘ChatGPT Work’带来的工具升级上。

It's pretty clear that part of the reason that OpenAI started to let people talk about GPT-5.6 early is that they wanted the emphasis on the official announcement day to be on the harness updates that we got in a new user interface that they call ChatGPT Work.

OpenAI写道：‘ChatGPT Work是一个能跨应用和文件执行任务、为项目持续工作数小时，并将目标转化为完整成果的智能代理。’

OpenAI wrote: ChatGPT Work is an agent that can take action across your apps and files, stay with the project for hours if needed, and turn a goal into finished work.

表面上看，这是OpenAI对Claude Cowork的回应。

Now, at a surface level, this is OpenAI's answer to Claude Cowork.

但更深层看，这显然是他们超级应用战略的下一步。

But it's also clearly the next step of their super app strategy.

这是一个全新的智能代理框架，把编解码器的理念扩展到了所有知识型工作场景。

This is a new agentic harness that takes the functionality of codecs and extends that approach out to all knowledge work.

OpenAI提到：学习使用ChatGPT Work最好的方式，就是交给它你已经很熟悉的任务。

OpenAI wrote: The best way to learn how to use ChatGPT Work is to give it tasks you already know well.

比如分析月度预算差异，把原始材料整理成营销方案简报，或者为销售会议做准备。

Analyze a month end budget variance, turn source materials into a marketing campaign brief, or prepare for a sales meeting.

你可以随时跟进它的进展，回答它的问题，调整方向，优化关键操作。

You can follow its progress, answer questions, change direction, and improve important actions.

这个界面支持你期待的所有功能，集成了Notion、Google Drive和Microsoft 365等主流工具，能轻松访问完整的工作环境。

Now, the interface supports everything you would expect it has connectors for all of the typical tools including Notion, Google Drive, and Microsoft three sixty five allowing easy access to full work context.

它还支持定时任务，运行在云端服务器上，即使你关上笔记本，代理依然能继续工作。

It also supports scheduled tasks and functions on a cloud instance so your agents can work even when your laptop is closed.

它支持企业级的安全配置和权限管理，这正是企业对前沿AI工具的普遍期待。

It's configurable for enterprise grade security and access controls, as companies are coming to expect from Leading Edge AI tools.

早期用户的反馈非常有力：Zapier企业业务负责人Angela Ferrante说：我用ChatGPT Work搭建了一个每月自动处理数千条潜在客户线索的标准化流程。

The early testimonials sound pretty compelling: Angela Ferrante, the head of enterprise at Zapier, wrote: Used ChatGPT Work to build a repeatable system for reviewing thousands of leads each month.

它追踪了Zapier客户在CRM、邮件和其他工具中的所有接触点，找出跟进中断的地方，并生成了一份每周高管仪表盘，揭示了流失的销售线索，还发现了高达七位数的潜在销售额。

It traced customer touchpoints across Zapier's CRM, email, and other tools, found where follow ups broke down, and generated a weekly executive dashboard that highlighted missed pipeline and revealed 7 figures in potential sales.

实际上，从所有发布材料来看，核心卖点就是把编码工作中的模式推广到更广泛的领域。

And really, across the entire set of announcement materials, the biggest part of the pitch is about taking work patterns in codecs and applying them more generally.

OpenAI不再鼓励人们去细抠任务的每个小环节，而是建议知识工作者先明确目标、加载上下文，然后用ChatGPT Work来完成复杂的多步骤任务。

Rather than micromanaging small components of tasks, OpenAI is encouraging knowledge workers to define a goal, load context, and use ChatGPT Work to complete large multi step tasks.

更重要的是，他们表示团队已经迅速适应了这种新方式。

What's more, they said that their teams have already adapted to this new approach.

现在所有团队都在用ChatGPT Work处理编码工作，OpenAI写道：在销售中，ChatGPT Work把一场初步沟通转化成了针对关键业务问题的定制化方案演示，全程只用了24小时，而过去通常要几周。

All of their teams are now using ChatGPT Work in Codecs with OpenAI writing: In sales, ChatGPT Work turned a discovery conversation into a tailored proof of concept for a mission critical problem within twenty four hours, a process that normally takes weeks.

ChatGPT整理了会议记录，把需求转给解决方案架构师，并与技术团队协作，让销售负责人能专心服务客户，真正扮演高价值顾问的角色。

ChatGPT structured the notes, routed the request to a solutions architect, and collaborated with the technical team, freeing the lead to focus on the customer and serve as a high value consultative partner.

在财务部门，ChatGPT Work把月末结账和预测的时间从几天缩短到几小时，它帮团队找数据、导入Excel、对账、做PPT，还核对结果。

In Finance, ChatGPT Work reduced month end close and forecasting from days to hours, by helping teams find source data, move it into Excel sheets, reconcile it, create slides, and verify the results.

这样一来，财务团队就能腾出更多时间去分析预测变化的原因，解释变动背后的逻辑，并为管理层提供下一步行动建议。

This lets the finance team spend more time understanding what's changed in the forecast, explaining why it changed, and advising leaders on what the company should do next.

不过说到消费者的反应，比OpenAI可能期待的要冷淡一些。

Now when it came to consumer response, it was a little bit more muted than I think OpenAI might have hoped.

彼得·杨总结了我所感受到的一种情绪，他写道：我觉得ChatGPT Work和Codex之间的区别让人困惑。

Peter Yang summed up one set of feelings I think when he wrote: I think the ChatGPT Work versus Codex thing is confusing.

这让人不禁想问：那开发者是不是不用干活了？

It raises questions like: So developers aren't working?

那如果我用它来规划假期呢？

And what if I'm using it to plan my vacation?

这管用吗？

Does that work?

我觉得干脆就叫ChatGPT好了，要么就叫ChatGPT Codex，别搞什么标签或切换按钮了。

I think the whole thing should just be called ChatGPT, Codex or ChatGPT Codex and there should be no tabs or toggles.

在我看来，Codex对普通用户来说并不难懂，也不难用。

In my opinion, codex is not that complicated for a normie to understand and use.

你还是在聊天，只不过现在它能帮你做事了。

You're still chatting, except now it can do stuff.

这件事越早统一越好。

The sooner this is all unified, the better.

伊桑·莫洛赫说：这让人困惑。

Ethan Moloch wrote: is confusing.

Claude Cowork 是特意设计得更安全的版本，防止非程序员不小心搞出大乱子。

Claude Cowork was a purposefully more secure alternative to code designed so non coders couldn't cause too much trouble.

我明白了。

I got that.

但我搞不懂什么是 ChatGPT Work，用它和用 Codex on Work 相比，我到底得到了什么，又失去了什么。

But I don't understand what ChatGPT Work is, or what I am gaining or giving up using it versus codecs on Work.

有些人甚至开始担心这两个应用会混在一起。

Now some were even actively worried about the blending of the apps.

丹·施珀再次写道：我特别担心这个，因为我很喜欢 Codex 这个应用。

Dan Schipper again wrote, I was extremely worried about this because I love the Codex app.

OpenAI 当时陷入了一个两难境地：怎么在一个应用里同时满足普通用户、程序员和企业的不同需求。

OpenAI was caught in an interesting position: how to make an agent orchestration app for regular ChatGPT consumers, coders, and businesses all in one app.

现在他们把界面分成了 ChatGPT Work 和 ChatGPT Codecs 两个部分。

They now split the interface between ChatGPT Work and ChatGPT Codecs.

其实两者基本一样，只是 Work 会隐藏代码。

They're basically the same except Work hides code.

在这两个版本里，聊天功能都被降级为仅用于快速提问的次要功能。

And chat has been demoted to second tier status for quick questions in either one.

这算不上多大的跨越，但也不算什么重大倒退。

It's not a big leap, but it's not a huge setback either.

不过它依然是我最喜欢的桌面智能代理管理应用。

And it remains my favorite of the desktop agent orchestration apps.

但当你发布一个重磅新功能时，大家的反应却是‘怎么合并了’，就像丹的标题那样，这显然不是你希望看到的。

Still, when you're coming with a big new announcement, having the response be the merged app is fine which is what Dan's headline was is not necessarily exactly what you want.

我觉得命名上确实容易让人困惑，要真正理解这套设计，还得靠大家实际用一段时间，慢慢体会它的好处和难点在哪里。

I think that there are confusions in the naming conventions, and to get a real feel for this, it's just gonna take some time for people to actually experience it, and figure out what the benefits or the particular challenges are.

另外，伴随 ChatGPT Work 一起推出的，还有他们更新过的 Sites 功能。

Now one additional feature that came alongside ChatGPT Work was their updated Sites feature.

这让你能把任何知识型工作直接变成一个网站或网页应用，分享给公司里其他用户，哪怕他们根本不用ChatGPT——我之前专门做过一期节目，讲过我觉得哪些知识型工作场景更适合用网站来呈现，而不是传统的文档形式。

This basically allows you to turn any knowledge work into a website or web app that you can share with other users inside your company, even if they're not using ChatGPT, which I already did a whole show about how many different knowledge work use cases I think you should be building websites for instead of the traditional artifacts.

所以我觉得，让这个过程更简单，再配上更好的托管功能，会真正改变人们用ChatGPT输出成果的方式。

And so I think that making that easier and integrating better hosting is going to actually be a pretty significant change in how people output work with ChatGPT.

但归根结底，这些都还是第一印象。

Ultimately, all of this is still first impressions.

很明显，OpenAI团队对这些新功能非常兴奋，也相当有信心：等大家用多了5.6和Work这个新框架，第二次、第三次使用时的感受，肯定会比第一次好得多——不过到底是不是这样，咱们还得等时间来验证。

What's clear is that the OpenAI team is really excited about all of these things, and seemingly fairly confident that as users get more reps on with both 5.6 and with the Work harness, our second and third impressions are going to be even better than our first, but for that, we'll just have to wait and see.

我们早就知道5.6要正式上线，至少在OpenAI开始暗示新框架要来之前，就已经提前24小时有所预感。

Now, we knew the official release of 5.6 was coming, and at least twenty four hours before OpenAI started teasing that a new harness was coming as well.

但谁也没想到，扎克伯格会时隔三年，第一次发推宣布：Meta今天也发布了一个新模型，叫Muse Spark 1.1。

What I don't think anyone expected was for Mark Zuckerberg from Meta to tweet for the first time in three years, announcing that Meta was also releasing a new model today, this time called Muse Spark 1.1.

Muse Spark的第一个版本今年四月就推出了，当时没掀起什么浪花。

Now, the first edition of Muse Spark landed in April, without much fanfare.

但这次更新后的版本，简直像换了个东西。

But this updated version seems like an entirely different beast.

在基准测试中，这个模型的表现与Opus 4.8和GPT-5.5相当。

On the benchmarks, the model looks competitive with Opus 4.8 and GPT-5.5.

在人类最后一场考试中，它在网页搜索、工具使用和知识测试方面大幅领先这两者。

It beat both by a significant margin on Humanity's last exam, which tests web search, tool execution, and knowledge.

在编程方面，它跟两者差距不大，Terminal Bench 2.1上比Opus和5.5低几个百分点，在SuiteBench Pro上则介于两者之间。

On coding, it's in the ballpark, a couple of points behind Opus and 5.5 on Terminal Bench 2.1, and between the two rival models on SuiteBench Pro.

在日益成为标准的Deep Suite测试中，MuSparq 1.1得分53.3%，落后于Opus的59%和GPT-5.0.5的67%。

On Deep Suite, which is increasingly important as a standard, MuSparq 1.1 was a little further behind with a score of 53.3% compared to 59 for Opus and 67 for GPT-five 0.5.

但对Meta来说，自从LAMA 3之后一直没推出过像样的大模型，这次已经是重大突破了。

Still, for Meta, who have struggled to produce a viable LLM since LAMA three, this is a big improvement.

而这款模型真正脱颖而出的地方，是在个人代理任务上。

And where the model shows signs of actually leading the pack is on personal agentic tasks.

在Jobbench测试中，MuSpark击败了Opus 4.8和GPT-5.0.5，这个测试评估的是完成真实职场任务的能力。

MuSpark beats both Opus four point and GPT-five 0.5 on Jobbench, which tests the ability to complete real life professional work.

在MCP Atlas上，它的表现达到了顶尖水平，这个测试模拟了通过MCP服务器收集信息的场景。

It's state of the art on MCP Atlas, which is a toolathon style benchmark that tests the ability to gather information through MCP servers.

在DeepSearch QA这个代理搜索基准上，它的表现略低于GPT-5.5，但稍微领先于Opus 4.8。

And on DeepSearch QA, which is an agentic search benchmark, it's slightly behind GPT-five point five and slightly ahead of OPUS 4.8.

总的来说，这个模型已经展现出一些前沿水平的迹象，尤其是在Meta专注于更贴近消费者的个人助手型AI的背景下。

Basically, this is a model that has glimmers of frontier performance, especially in the context of Meta focusing on more consumer friendly personal assistant style AI.

目前我们还没见过太多人和这个模型互动，但它显然也不是那种刚发布、完全没人用过的原型。

Now, we haven't seen a ton of interaction with the model yet, but it also doesn't fully appear to be one that just released with no one ever having had a chance to build with it.

比如Julius团队就用MuSpark在Julius里搭建了一个Minecraft的克隆版，只花了五分钟，消耗了0.073美元的token。

The team from Julius, for example, used MuSpark to build a Minecraft clone inside Julius, which it was able to do in about five minutes for zero seven three dollars in tokens.

在Valve的AI公开大模型评估中，New Spark 1.1排在第四名，超过了GPT-5.5和Grok 4.5；Valve的AI团队还提到，它的运行速度是前三名模型的三倍。

On the Valve's AI Public LLM evaluation, new Spark 1.1 landed at number four, ahead of both GPT-five 0.5 and Grok 4.5, and as Valve's AI pointed out, running three times faster than the top three models.

Valve的Rayon写道：这模型也太快了吧。

Rayon from Vals wrote: Wow this model is fast.

在我们的所有测试中，它的延迟只有Opus 4.8的四分之一，是GPT-5.5的一半。

Across our benchmarks we found it to be one fourth the latency of Opus four point eight and one half the latency of GPT-five 0.5.

我原以为Meta的网络基础设施肯定很强，但真没想到他们是怎么做到在高并发下还能跑得这么快的。

I would expect Meta to have incredible web infra, but really don't know what witchcraft they're pulling to host the model for such fast inference at high rate limits.

不过，雷昂提到的、也是这次发布最关键的一点，是成本问题。

Still, the biggest thing that Rayon talked about and that was really key to this announcement was the cost.

Chubby写道：MetaMuse Spark 1.1是个非常出色的智能代理模型，但最重要的是，它便宜得惊人，成本效率极高。

Chubby writes: MetaMuse Spark 1.1 is a very good agentic model, but above all it is incredibly affordable and cost efficient.

比如在Vibe Code Bench 1.1上，它排第五，但测试成本只有5.9美元，而Opus 4.8要5.9美元，Fable 5更是高达12.51美元。

For example, on Vibe Code Bench 1.1, it came in fifth overall, but the cost to test it was compared to, for example, dollars 5.9 for Opus 4.8 and $12.51 for Fable five.

雷昂再次从Valve那边留言：这模型太便宜了，我简直不敢相信。

Rayon again from Valves wrote: The model is so cheap, I almost don't believe it.

实际使用中，它的成本只有Fable和GPT-5.0.5的十分之一。

In practice, we see it's one tenth the cost of both Fable and GPT-five 0.5.

如果你以为开源模型会把利润空间挤掉，那等你看到这个，才真叫大开眼界。

If you thought open source models would compete away margins, just wait till you see this.

用MuSpec 1.1的成本，居然比自己部署一个开源模型还低。

It's somehow cheaper to use MuSpec 1.1 than host your own OS model.

利奥精准说出了很多人的心声：这周的第二个高光时刻，Meta发布了MuSpec 1.1，它同样是能和Opus 4.8、GPT-5.0.5一较高下的前沿模型。

Leo really summed up the feeling of many when he wrote: For my second moment of this week, Meta just announced MuSpec 1.1 and it's also a frontier level model competing with Opus 4.8 and GPT-five 0.5.

扎克伯格和马斯克又回来了。

Zuck and Elon are back.

如果我们把视野放得更广一些，这周发布的每一个模型——Grok 4.5、Cognition 的 SWE 1.7、MuSpark 1.1，甚至 GPT-5.6——在发布时都极其强调成本和效率。

And I think as we zoom out to broader implications, every single model this week Grok 4.5, SWE 1.7 from Cognition, MuSpark 1.1, and even GPT-5.6 had an incredibly strong emphasis on cost and efficiency in its announcement.

就连 GPT-5.6 这样主打顶尖性能的模型，重点也放在了实际表现上，以及它对整体 Token 预算带来的改变。

Even 5.6, which was also competing for the state of the art, so much of the emphasis was about its performance in practice, and the difference that that was going to make to overall token budgets.

这一切意味着，AI 竞赛的格局确实已经变了，而且我们现在正式进入了一个新阶段：各大实验室都意识到，它们的竞争维度已经不再只是前沿性能了。

What this all means is not only that the AI race has shuffled which it clearly has but we are now officially at the point where the labs themselves realize that they are competing on an entirely different vector than just Frontier performance alone.

本周初，xAI 和 Meta 在企业选型时根本没人提起。

At the beginning of this week, xAI and Meta were not functionally in the conversation at all when it came to model selection, especially for enterprises.

但到了周末，它们又重新回到了这场对话中。

And they are now ending the week back in that conversation.

事实上，Semi Analysis 甚至写道：从最基础的角度看，打造一个真正的前沿模型需要三样东西：数据、人才和算力。

In fact, Semi Analysis even wrote: At the simplest level, there are three things you need to build a true frontier model: data, talent, and compute.

我们认为，Meta 是唯一一家在数据、人才和算力这三方面都走在正轨上的超大规模云服务商兼实验室，因此最有可能追上 Anthropic 和 OpenAI。

We believe Meta is the only hyperscalerneolab on track to be world class at all three, and therefore has the best chance at catching up with Anthropic and OpenAI.

AI领域的格局又发生了变化，而这次，我们所有人都是受益者。

The AI tectonic plates have shifted once again, and in this case we are all the beneficiaries.

这个周末和接下来几周，我很期待深入研究这些新模型。

I am excited to dig more into these models over the weekend and in the coming weeks.

目前，今天的AI每日简报就先到这里。

For now, that is gonna do it for today's AI Daily Brief.

一如既往，感谢你的收听或观看。

Appreciate you listening or watching as always.

我们下期再见，保重。

And until next time, peace.