为什么谷歌没有追赶Claude Code

今天在《AI每日简报》中，我们来看看谷歌在I/O大会上发布了哪些内容，为什么他们的AI战略看起来越来越混乱，但也许这根本无关紧要，因为他们带来的某些优势实在太关键了。

Today on the AI Daily Brief, a look at everything Google announced at IO and why it seems like their AI strategy is getting messier and messier, but it also just might not matter because of some of the significant advantages that they are bringing to the table.

《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，先说几件事。

Quick announcements before we dive in.

首先，感谢今天的支持方：普华永道、Scrunch、Assembly 和 Section。

First of all, thank you to today's sponsors, KPMG, Scrunch, Assembly, and Section.

想收听无广告版本的节目，请访问 patreon.com/aidailybrief，或者在苹果播客订阅我们。

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

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

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

当然，所有关于这个生态系统的其他信息——包括职位机会、通讯简报、免费课程、付费企业培训项目——都可以在 aidailybrief.ai 上找到。

A I daily brief dot a I is, of course, where you can find out about everything else going on in this ecosystem as well, Job opportunities, newsletters, free education programs, paid enterprise education programs, all of that, again, is at aidailybrief.ai.

最后再提一点，大家都知道Google I/O的回顾肯定会是一期完整的节目，但谁也没想到，前OpenAI联合创始人安德烈·卡帕西居然宣布加入了Anthropic。

Now one final note, you always had to know that the Google IO recap was going to be a full end to end episode, but that certainly didn't account for former OpenAI cofounder Andrei Karpathy announcing that he had joined Anthropic.

对很多乃至大多数关注AI的人来说，这个消息比I/O舞台上发布的任何内容都更重磅。

To many, if not most, enfranchised AI watchers, this was a bigger announcement than anything that happened on the stage at IO.

所以，我们明天一定会再回来聊这个事。

And so with that in mind, we will certainly be coming back to it tomorrow.

不过现在，我们先来好好聊聊Google的动态。

But for now, we have a lot of Google to talk about.

今天我们要谈的是Google I/O，这次活动不只是发布一堆新品，更是让我们看到这家顶级实验室如何看待AI的优先级，以及他们在AI竞赛中的位置。

Today, we are talking about Google IO, which in this particular case is more than just a set of announcements, but a chance to see how one of the biggest labs thinks about AI priorities and where they sit in the AI race.

简而言之，这场活动有点乱。

In short, the event was a little confused.

Google现在做的事情真的太多了。

Google is doing a ton.

这一点毫无疑问。

There's absolutely no doubt.

但最终的结果到底是什么，反而不太清楚。在我看来，谷歌的管理层对‘赢得AI竞赛’究竟意味着什么，可能和Anthropic或OpenAI的管理层有着完全不同的想法。

What it adds up to is a little less clear, and in fact, it seems to me like Google's leadership may have a very different idea than either Anthropic or OpenAI's leadership about what winning the AI race will actually look like.

不过在深入讨论今年的活动之前，我们得先铺垫一些背景信息。

But we need a little bit of background and context before we get into this year's event.

谷歌在生成式AI领域的历程，一直挺有意思的。

Google's history with generative AI in general has been interesting.

在那个还叫‘史前时代’的阶段，也就是ChatGPT出现之前，几乎没人关注这些事，而谷歌早就因为一直在关注而领先了。

In the prehistoric times, I e, the pre chat GBT times, when very few were paying attention to all of this, Google was ahead simply by virtue of paying attention.

早在2014年，他们就以5亿美元的高价收购了DeepMind，但问题在于，这并不是他们唯一的AI项目。

Back in 2014, they acquired DeepMind for a then massive $500,000,000, but problematically, it was not their only AI effort.

事实上，2022年11月ChatGPT一发布，谷歌就措手不及，原因就在于他们的AI战略太分散，没有统一管理。

In fact, part of the reason that they were caught flat footed when ChatGPT first launched in November 2022 was that AI strategy wasn't consolidated in a single place.

直到后来，他们才开始整合，说实话，2023年初那一年过得特别艰难，才让他们意识到必须经历这种痛苦的重组。

That wouldn't come till later, in fact, and it honestly took a very rough year out of the gate in 2023 for them to figure out that they needed to go through that sort of painful restructuring.

2023年的IO大会，明显就是一场追赶战。

Twenty twenty three's IO event was certainly a catch up event.

你可能还记得，早在Gemini之前，我们有个叫Bard的产品。

You might or might not remember that before Gemini, we had something called Bard.

Bard是谷歌推出的第一个ChatGPT竞争对手，说实话，表现并不怎么样。

Bard was Google's first ChatGPT competitor, and frankly, it wasn't great.

对很多人来说，它更像微软会推出的那种产品。

Honestly, to many, it felt more like the type of product that Microsoft would put out.

但令人惊讶的是，由于和OpenAI的合作，微软当时普遍被认为领先一步。

But Microsoft, shockingly, by virtue of their deal with OpenAI, was at the time broadly seen as ahead.

整个2024年，谷歌承受着来自多方的压力，尤其是资本市场，要求他们推出超越Bard的产品，最终我们在12月等来了Gemini的发布。

Throughout 2024, there was growing pressure on Google from lots of sources, not least of which was the public markets, to do something beyond and better than Bard, which eventually we got in December with the announcement of Gemini.

Gemini标志着谷歌开始围绕DeepMind的CEO德米斯·哈萨比斯整合战略，但即便看起来有希望，首次发布还是让人挺失望的。

Gemini was Google starting to consolidate their strategy around DeepMind CEO Demis Hassabis, but even if it appeared promising, that first announcement was still pretty disappointing.

最强大的Gemini版本要好几个月后才上线，感觉像是仓促推出的。

The most powerful version of Gemini wouldn't be available for a number of months, and it had the feeling of a rushed move.

2024年初的情况也没好到哪儿去。

Early twenty twenty four didn't go much better.

年初的时候，网上到处都在讨论他们所谓的‘觉醒’图像生成模型，比如你让它生成一张1943年德国士兵的照片，它却画出穿着纳粹军装的日本女性或非裔美国男性。

At the beginning of the year, there was endless digital links build about their quote unquote woke image generation model, which when asked to do things like generating an image of a 1943 German soldier, would put Japanese women or African American men in Nazi regalia.

到了2024年的I/O大会，我们终于看到了功能完整的Gemini模型，但即便如此，谷歌还没真正走出困境。

By the IO event in 2024, we were starting to get fully capable Gemini models, but even then Google wasn't out of the woods yet.

他们在那次大会上大力推广的AI概览，是谷歌第一次尝试把生成式AI能力整合进最重要的核心业务——搜索。

One of the big things they pushed at that event was AI overviews, which was Google's first attempt to integrate this new set of generative AI capabilities with their most important existing business line, which was, of course, search.

问题是，这些早期的AI概览没因为好用而出名，反而因为建议人们往披萨上抹胶水、吃石头而成了笑柄。

The problem was that that first set of AI Overviews became known not for how helpful they were, but for suggesting that people put glue on pizza and eat rocks.

但到了2024年底，谷歌总算找回了状态。

And yet by the end of twenty twenty four, Google had started to get their groove back.

而这种势头的转变，来自一个意想不到的源头。

And part of the shift in momentum came from an interesting source.

内部上，他们已经把所有战略整合到一起，统一管理，所以有些成果只是这一调整带来的回报。

Now internally, they had consolidated strategy and put everything under one roof, and so some of it was just the dividends from that.

更重要的是，他们时隔多年，第一次真正推出了一款爆火的产品——NotebookLM。

But they also, for the first time in a very long time, had a genuine breakout product hit in the form of NotebookLM.

特别是NotebookLM的音频概要功能，是在秋季推出的，你可以把一堆资料丢进去，让AI生成一个像播客一样的语音讲解。

Specifically, the audio overview feature of NotebookLM, came in the fall, where you could have a synthetic AI podcast discuss whatever set of resources you had dropped into Notebook.

很多人发现，这真是学习新知识、备考、了解新闻的绝佳方式，于是纷纷开始使用。

People saw that this could be a really cool way to learn new material, study for exams, familiarize themselves with the news, and lots and lots of people did it.

到了2025年，谷歌的AI业务已经积累了不错的势头，而I/O大会也证实了这一点。

Google's AI efforts then headed into 2025 with good momentum, and I o twenty twenty five validated that.

他们推出了V3，这是首个原生支持音频的视频生成模型。

They premiered v o three, which was their first video generation model to have native audio.

整个2025年，Gemini模型真正成了能和ChatGPT一较高下的选择，用户增长也明显提速，逐渐逼近ChatGPT的水平。

And throughout 2025, Gemini models were treated as a genuine option and contender, and you saw that in the growth of Gemini, which started to get up to chat GPT style numbers.

谷歌还推出了几个真正拓展了生成式AI可能性的产品，最突出的就是Nano Banana和Nano Banana Pro。

Google also had a couple of moments where something that they released genuinely expanded the set of things that were possible with generative AI, most notably, Nano Banana and Nano Banana Pro.

Nano Banana在八月底发布，最初官方名字叫Gemini 2.5 Flash Image，后来发现大家更喜欢叫它Nano Banana，就干脆改了名。它厉害的地方不在于图片质量有多惊艳，而在于它提供了前所未有的精细编辑功能，其他模型根本做不到。

Nano Banana, which came out at the end of August and which was technically called Gemini 2.5 Flash Image before they realized that they should just call it by the name that people liked for it, wasn't remarkable because the base level quality of the images was so differentiated but unlocked a ton of value in the new types of fine grained editing controls that it gave you that were simply not possible with other image generation models.

到了十一月，Nano Banana Pro进一步升级，加入了对提示词的推理能力，还有强大的文字渲染功能，第一次真正让信息图这类内容在线上变得实用起来。

That expanded in November with Nano Banana Pro, which added reasoning over the prompt, and significant text rendering capabilities that brought things like infographics online in a real way for the first time.

所有这些都让人觉得，谷歌在进入2026年时，AI方面的势头是前所未有的强劲。

All of this led to a feeling that Google was headed into 2026 with the best AI momentum that they had ever had.

然而，悄悄地，一些事情正在发生，至少从叙事角度看，让它们再次大幅落后了。

Quietly, however, something was happening that would push them from a narrative perspective at least significantly back behind again.

2026年是编码代理、知识型代理和工作流工具的年份。

2026 has been the year of coding agents, agents for knowledge work, and harnesses.

而这一切的种子，早在2025年初就已经埋下了。

And the seeds of that were laid at the beginning of 2025.

其中一部分是推理模型，各大实验室在2025年全年陆续发布了这类模型。

One part of that was the reasoning models, which all of the major labs released throughout the course of 2025.

但当Claude Code在去年二三月推出时，很多人还没意识到工作流工具的重要性。

But the importance of harness wasn't exactly clear to everyone when Claude Code launched back in February and March of last year.

Codex技术上晚了几个月，到五月才上线，而整年里，这些变化都在默默酝酿。

Codex would technically come a few months later in May, and for much of the year all of this was quietly brewing.

Anthropic的开发者生态明显引起了OpenAI的注意，到了八月，大家已经清楚，他们开始认真对待这个威胁了。

Anthropics developer devotion clearly started to be noticed by OpenAI, and by August, it was clear that they were taking the threat seriously.

他们试图把GPT-5的发布包装成完全围绕编码应用，首次把编码相关用例放在发布内容的首位，但最终这部分被GPT-5实际表现平平，以及GPT-4.0即将停用所引发的用户强烈反弹给边缘化了。

They tried to position their g p t five announcement as all about that, putting coding based use cases up top of their announcement for the first time, but ultimately, that part of the announcement was significantly sidelined both by the genuinely underwhelming performance of GPT-five as well as by the planned deprecation of GPT-four point zero which led to effectively a consumer rebellion.

不过，即便在应对这些风波的同时，到年底时已经很清楚，OpenAI已经全面押注在代码模型和编码相关技术上。

Still, even as they worked through that, by the end of the year it was very clear that OpenAI was all in on codecs and coding related models.

我们每推出一个新版本——GPT-5.1、GPT-5.2、GPT-5.3——都会配套一个专门优化过的代码版本。

Each incremental new release we got, GPT-five point one, GPT-five point two, GPT-five point three, had a codecs optimized version along with it.

正如大家所知，所有这些进展在Opus 4.5和GPT-5.0.2这个阶段汇聚到了一起。2026年初，尤其是进入一月后，人们突然意识到，局面已经彻底改变：用代码构建东西的能力跟以前完全不一样了，那些曾让人兴奋多年的智能代理，不再只是未来的设想，而是实实在在地来到了眼前。

And as we well know, all of these things came together around the Opus 4.5, GPT-five 0.2 era, where all of a sudden around the beginning of twenty twenty six and especially into January, everyone was realizing that something big had shifted, that the capacity to use code to build things was totally different than it had been before, and that the promise of agents, which had been so exciting for so many years, was no longer just something for the future but was actually here.

当智能代理——尤其是编码代理——成为主流话题，当人们开始把编码代理的应用扩展到知识工作领域时，谷歌又一次被远远甩在了后面，至少在圈内人的讨论中是这样。

When the dominant theme became agents, specifically coding agents, and the reappropriation of coding agents for knowledge work, it absolutely left Google in the dust once again, at least from the insider AI conversation.

那么，会不会有人认为，圈内人过度聚焦在编码代理上，其实是只见树木不见森林？

Now is there an argument that the insider fixation on coding agents is missing the forest for the trees?

这会不会恰恰反映出AI发展的某种困境：我们对这个需求明确、市场契合度极高的领域过于痴迷，甚至把它的成功过度延伸到了那些远为困难的知识工作场景？

Is it actually potentially an indictment of AI progress that we're all obsessed with this one area of immense product market fit and, in fact, overinterpolating that product market fit to other areas of knowledge work that are going to be much harder.

这种可能性确实存在。

It is certainly possible that that is the case.

当然，作为十二月那次重大战略调整的一部分，OpenAI 决定，这个方向太重要了，以至于他们第一次放弃了其他与谷歌竞争的领域，尤其是 Sora 这个视频模型和应用，基本上彻底搁置了。

Certainly, as part of their big code red in December, OpenAI made the decision that it was significant enough that they were willing to, for the first time, abandon other areas of their ambition that were competing with Google, most notably around Sora, the video model and app that they basically completely gave up on.

但与此同时，很明显，不光是那些深度使用 AI 的用户，大家都觉得有什么大事发生了。

And yet at the same time, it is very clear that it's not just the extremely enfranchised AI users that think that something significant has changed.

企业界已经全面押注这种新的 AI 方式，这给其他实验室带来了问题，尤其是谷歌。

The enterprise has gone all in on this new approach to AI in a way that opens up questions for the labs that aren't there, most notably Google.

这并不是说谷歌在 2026 年上半年就没有优势。

Which is not to say that Google didn't have any big advantages in the first half of twenty twenty six.

谷歌的 TPU 芯片因为算力瓶颈日益凸显，变得比以往任何时候都更重要。

Google's TPU chips have become a bigger deal than ever as compute constraints have come home to roost.

而且，谷歌第一次开始把 TPU 当作不只是内部资源，而是一个对外的业务方向。

And for the first time, Google is starting to think about that as not just as an internal capacity, but as an external business line.

更关键的是，OpenAI 决定全力聚焦企业市场，完全忽略消费者端——当然，他们自己不会说得这么直白——这反而给谷歌腾出了一条越来越宽的路。

And what's more, OpenAI deciding to focus on the enterprise to the exclusion of consumer, although, of course, they wouldn't put it quite that dramatically over at OpenAI, gave Google what appears to be an increasingly open lane.

所以，带着这些背景，大家在 IO 大会前都在问：第一，谷歌会不会推出一个能跟得上前沿的模型？

And so with all of that, the questions coming into this IO were, one, would Google release a competitive state of the art model?

第二，他们会发布、更新，还是整合出一个真正能用于编程或知识工作的智能代理系统吗？

Two, would they release, update, or consolidate around a real agentic coding or knowledge work harness?

我们该用 Antigravity 吗？

Are we supposed to use Antigravity?

我们该用 Google AI Studio 吗？

Are we supposed to use Google AI Studio?

对 OpenAI 来说，是 Codex。

For OpenAI, it's Codex.

对 Anthropic 来说，是 Claude Code 或 Claude Cowork，这两个都在同一个应用里。

For Anthropic, it's Claude Code or Claude Cowork, both of which live in the same app.

对 Google 来说，工具太分散了，从来没人说得清到底该重点投入哪个系统。

For Google, it's a sprawl, and it's never been exactly clear which of these harnesses we should invest in.

除了模型和工具的问题，Google 会明确表态在消费端和企业端之间怎么选吗？还是会抓住这个新AI权衡时代的机会，比如主打低成本或更高效的模型？

On top of questions of model and harness, would Google clarify their position on consumer versus the enterprise, or would they lean into this new AI trade off era, for example, by honing in on lower cost or more efficient models as one of their unique opportunities?

结果呢？说实话，有点乱七八糟。

What we got was, honestly, a little bit of a confusing mess.

我们接下来要聊的关键发布包括Omni、Spark、Antigravity 2.0和Gemini 3.5 Flash，这些只是公布内容的一部分，但我们要看看它们如何拼出一个更大的图景。

The key announcements we're going to talk through are Omni, Spark, Antigravity 2.0, Gemini 3.5 Flash, which together only represent a part of what was announced, and try to understand how they all add up to something bigger.

光看推特和社交媒体，你可能会觉得Omni只是一个视频模型，但实际上，谷歌把它定位为新一代生成式AI模型，最终目标是实现真正的全模态——任何输入都能对应任何输出。

Just by looking at Twitter and social media, you would think that Omni was just a video model, but in fact, Google is positioning it as a new family of generative AI models that are eventually an anything to anything truly multimodal medium.

它的理念是，不再局限于用视频输入做视频、用文字输入做文字，而是让Omni在完整形态下，能接受任何输入形式，甚至多种输入组合，然后按你的需求生成任意格式的输出。

The idea is that instead of being constrained to a video model using a video input or a text model using a text input, Omni at full strength will be able to take any input or any combination of inputs to produce whatever output you need in whatever format you need.

这种能力其实早就被承诺了很久，但一直没真正实现过。

Again, this is a type of thing that's been promised for a long time but has never fully come to bear.

这次发布的版本被明确标为视频模型，所以大多数人一开始都把它当成了一个视频工具。

The version that was released was specifically positioned as a video model, and that's how most people initially judged it.

研究者、AI谣言澄清者安迪·马斯利总结了很多人最初的感受，他在推特上说：谷歌的Omni挺好玩，但对我来说算不上重大突破。

Researcher and AI water myth dispeller Andy Massley summed up the first impression feelings of many when he tweeted, Google's Omni is fun, but definitely not a huge step up for me.

在X平台上的Heisenberg说得更直接：本来期待的是VO4，结果等来了Gemini Omni。

Heisenberg on x went farther, saying, was expecting v o four got Gemini Omni instead.

Seed Dance 2.0肯定能把这个当早餐吃掉。

Seed Dance two point o is going to eat this for breakfast.

在评论里，他还说，Grok Imagine 更厉害。

In the comments, he even said, Grok Imagine is better.

然而，很多人指出，拿它和 CDance 直接比较其实并不恰当。

And yet, as many jumped in to point out, the comparison to CDance itself was actually not particularly apt.

卡洛斯·桑塔纳写道：这是一款前所未有的视频编辑工具，就像 Nana Banana 那样。

As Carlos Santana wrote: This is a model for editing videos like Nana Banana like we've never had before.

事实上，当你仔细读完博客文章后，就会发现谷歌正是这么定位它的。

And indeed, when you actually read the blog post, that's how Google is positioning this.

重点不在于视频生成的初始画质——虽然画质确实不错，而在于你能否轻松地对其进行编辑。

It's less about the initial quality of the video generation, although that's strong, and more about your ability to easily edit it.

没过多久，就有人明白了其中的玄机（双关语），安迪·马斯利后来也回过头说：好吧，我收回对 Omni 的评价。

It wasn't long before some people grokked this, pun intended, with Andy Masley actually coming back later and saying, okay, I take it back on Omni.

用它来编辑真实视频，简直太疯狂了。

Editing real videos with it is crazy.

亨利·多布雷斯写道：Omni 对视频来说，更像是一次 Nana Banana 式的突破，而不是传统的影视级生成模型。

Henry Dobres writes, Omni feels much closer to a nano banana moment for video than a traditional cinematic generation model.

你这是在拿苹果和梨比。

You're comparing apples and pears.

我可以让Omni轻松改掉整个镜头的场景，把白天变成黑夜，同时保留原有的镜头结构。

I can easily ask Omni to change the entire setting of a sequence and turn day into night while preserving the shot structure.

这意义可太大了。

That's a massive deal.

高效的视频到视频转换，很可能成为未来娱乐和混合制作的核心部分。

Efficient video to video is probably going to become a huge part of the future of entertainment and hybrid productions.

伊森·莫洛赫写道：它在理解指令方面表现极佳，有了一个能处理视频的智能模型，你的创作空间一下子扩大了很多。

Ethan Moloch writes: It's really good at instruction following, and having a smart model capable of doing video expands what you can do by a considerable margin.

其他人还分享了不少例子，说明用它剪辑视频有多简单，比如把网红风格的视频里的网红本人去掉，换背景、换衣服，全都不在话下。

Other people shared a bunch of examples of how easy it was to edit video with character consistency, taking an influencer style presentation and making the influencer invisible, changing the background, and changing the person's outfit.

还有一个例子，视频里是伦敦的街景，镜头特写大本钟，Omni把原本普通的午后阳光场景，直接改成了跨年夜的烟花场景，连钟表时间都同步更新了。

In another example, there's a video scene of London with a close-up of Big Ben where Omni changed the scene from a generic afternoon daylight scene to New Year's Eve with fireworks with the clock updated as well.

我个人觉得，就我目前看到的来说，我们往往高估了基础模型的升级，却在一开始低估了模型可操控性带来的改变。

Now I tend to think that based just on what I've seen, we tend to overestimate the importance of base level model upgrades and underestimate and undervalue, at least initially, changes to the steerability of those models.

自从 Nano Banana 之后，情况有点不一样了，你能明白他们为什么拿这个当参照点，但很明显，这种层级的编辑能力会打开以前根本不可能实现的视频生成新用途。

That's changing a bit post Nano Banana, and you can see why they use that reference point, but it certainly seems like this layer and depth of editability will unlock new use cases in video generation that weren't possible before.

不过在我看来，自从 OpenAI 决定关闭 Sora 项目后，现在比几个月前有更多人在问：这东西到底给谁用的？

Still, it seems to me, like in the wake of OpenAI deciding to close the Sora experiment, there are more people asking the question of who this is actually for than there might have been just a few months ago.

Google 是不是觉得普通用户也会用这个？

Is this something that Google imagines general consumers using?

如果是的话，他们是用在社交媒体上，还是别的地方？

And if so, are they using it for social media or something else?

这会不会只是面向半专业用户的工具，只有特定的一小撮创作者才会用？

Is this just a prosumer type of feature where a very specific subset of creators are going to be using it?

它到底是专为职业视频工作者设计的实用工具，还是仅仅为了展示一种更大的模型范式转变，而这个功能只是最直观的演示方式？

Is it in fact an entirely professional focused tool made for people who actually use video for a living, or is it just meant to be a preview of a larger paradigm shift in models with this being the easiest way to demonstrate how those types of models will change things?

Google 并没有明确回答，也许只是因为他们自己也不确定。

Google didn't really answer that, which may simply reflect the fact that they just aren't sure.

说到那些还搞不清目标用户是谁的酷炫功能，接下来是 Gemini Spark。

Now speaking of cool things that are a little confused about their audience, next, we got Gemini Spark.

谷歌把它描述成一个全天候的个人助手，能根据你的指令帮你打理数字生活。

Google describes it as your twenty four seven personal agent that helps you navigate your digital life, taking action on your behalf and under your direction.

但这次依然不清楚他们具体指的是什么。

This once again was unclear exactly what the reference point is.

有些人直接认为这是他们版的OpenClaw。

Some jumped to the idea that it was their version of Open Claw.

《The Verge》确实就这么称呼它。

That's in fact what The Verge called it.

而且他们描述的某些方式，确实往这个方向靠。

And certainly, some of the ways that they described it point in that direction.

他们在发布帖子里说，这个助手基于Gemini 3.5运行，还用了‘反重力’技术，能在后台长时间执行任务。

Their announcement thread, they said it runs on Gemini 3.5 and is built on antigravity so it can perform long running tasks in the background.

他们特别提到，因为它运行在谷歌云的虚拟机上，你不用一直开着笔记本电脑——这明显是在调侃那些带着电脑到处走、随时用Claude Code或Codex的人。

They called out the idea that because it runs on virtual machines on Google Cloud, you don't have to keep your laptop open, which is clearly a reference to the people who are walking around Claude Code or Codex ing with their computers open.

他们还提到了通过MCP与谷歌工具和第三方服务的整合，暗示这个产品很可能瞄准的是那些已经在用Claude Code、Codex，甚至Hermes或OpenClaw的进阶用户或专业人士。

And they also even reference integrations with Google tools and third parties through MCP, suggesting that this might, in fact, be designed for that more prosumer or professional type of audience that's currently using Claude Code, Codex, or even something like Hermes or Open Claw.

但他们举的例子却更偏向于普通消费者的使用场景。

And yet the examples they give seem to point it a little bit more in the direction of a consumer use case.

Sundar Pichai 说，这是你的个人AI助手，帮你打理数字生活，替你做事，听你指挥。

Sundar Pichai said, it's your personal AI agent that helps you navigate your digital life, taking actions on your behalf and under your direction.

谷歌副总裁Josh Woodward说：‘需要给老板发封邮件汇报进展吗？’

Google VP Josh Woodward said, Need to send an email to your boss with a status update?

Spark能从你的邮件、文档、表格和幻灯片里提取所有信息，帮你起草好内容。

Spark can pull all the facts from your emails, your docs, your sheets and slides, and write the draft for you.

他还提到，小企业已经在用Spark了。

He also said: Small businesses are using Spark.

它们可以帮您盯着收件箱，确保不会漏掉任何客户的问题。

They can watch over their inbox so they never miss a question from a customer.

一方面，它显然不是专为非专业人士设计的纯个人助手；但另一方面，它也没把自己跟那些工具完全对立起来。

So on the one hand, it's clearly not fully a personal agent for a nonprofessional use, but it's also not exactly positioned in opposition to one of those other tools.

对谷歌来说，也许这本来就很直观——它就是要填补一个空白：那些觉得Claude Code太难上手，甚至觉得Claude Cowork太专业的人，却依然想要这类功能。

And maybe to Google, this is just supposed to be intuitive, that it is supposed to fill a gap that they see for users who are not comfortable with something like Claude Code or maybe even Claude Cowork, but who want this sort of capabilities.

但对他们来说显而易见的事，对别人却未必清楚。

But if that is obvious to them, it's not obvious to others.

人们不知道该怎么描述这个产品，这恰恰反映了谷歌眼下普遍存在的产品定位混乱。

And the way that people didn't know exactly how to frame this is an example of the broader product confusion which I think is everywhere with Google right now.

产品虽然发布了，却连具体什么时候能用都还不明确，只说今年夏天某个时候会推出，这让人们更难把握它的进展。

Making it harder to get a handle on, the product was announced but not only is it not actually available it's not even clear when it will be available, saying it's coming just sometime this summer.

眼下最重要的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擅长写提示词的人，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，并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?

这些能力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转向真正掌握AI能力，KPMG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，就是 kpmg.com/us/sophisticated。

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

问个问题：你上一次亲自上网站查东西，是什么时候？

Quick question: When was the last time you actually visited a website to research something?

如果你跟我一样，现在大部分搜索和调研工作都已经交给AI了。

If you're like me, AI pretty much does that work for you now.

这自然引出一个新问题：如果AI替你完成了发现、研究和决策，你的网站到底为谁而建？

That of course raises a new question for brands: If AI is doing the discovering, researching, and deciding, who is your website really for?

用户行为的这种转变——AI助手正成为你最重要的新访客——正是我的赞助商Scrunch正在全力应对的挑战。

That shift in user behavior the rise of AI bots becoming your most important new visitors is what my sponsor, Scrunch, is taking head on.

Scrunch 是一个AI客户体验平台，帮助市场团队了解AI代理是如何使用你们网站的，它们在哪些AI回答中出现、哪些没出现，以及是什么阻碍了它们被检索、被信任或被推荐。

Scrunch is the AI customer experience platform that helps marketing teams understand how AI agents experience their site, where they show up in AI Answers, where they don't, and what's preventing them from being retrieved, trusted, or recommended.

这不仅仅是曝光度的问题。

And it's not just visibility.

Scrunch 能帮你发现那些真正影响结果的内容缺口、引用缺失和技术障碍，并指导你修复它们，让你的品牌在 AI 回答中被找到、被选择。

Scrunch shows you the content gaps, citation gaps, and technical blockers that matter, and helps you fix them so your brand is found and chosen in AI Answers.

针对我们的听众，Scrunch 现在提供免费的网站诊断，帮你看清 AI 是怎么看待你的网站的，哪里存在缺口，以及你的品牌在 AI 回答中和竞争对手相比表现如何。

Now, for our listeners, Scrunch is providing a free website audit that uncovers how AI sees your site, where there's gaps, and how you're showing up in AI versus the competition.

访问 scrunch.com/aidaily，立即为你的网站做一次检测。

Run your site through it at scrunch.com/aidaily.

你肯定知道 AssemblyAI，他们的实时语音转文字准确度是行业顶尖的。

You know AssemblyAI for having the most accurate streaming speech to text out there.

但他们又往前走了一步，推出了完整的语音助手 API。

But they just went a step further and launched a full Voice Agent API.

原理很简单：只需一次连接，剩下的——听、想、说——全由他们搞定。

The idea is simple: one connection and they handle everything the listening, the thinking, the speaking.

你只要把音频流传进去，就能直接收到助手的语音回复。

You just stream audio in and get your agent's voice response back.

比如外呼销售电话能自动筛选潜在客户，客服能处理复杂问题而不依赖脚本，预约助手说话像真人助理，这些功能，用一个 API 五分钟就能搭建出来。

We're talking about things like outbound sales calls that actually qualify leads customer support that handles complex requests without a script, scheduling agents that sound like a human assistant, and you can build one in five minutes with one API.

更重要的是，他们的流式模型在识别其他语音助手容易漏掉的内容上表现最佳，比如电话号码、邮箱、人名和医学术语。

And importantly, their streaming model is the best at catching all the stuff that breaks on other voice agents things like phone numbers, emails, names, and medical terms.

如果你还在试用阶段，他们提供无合同限制、并发无上限的方案，让你可以毫无障碍地亲自体验。

And for those of you who are still in experimentation mode, there are no contracts and unlimited concurrency, so you can actually test it out without any friction.

前往 assemblyai.com/brief，直接在网站上试用实时语音助手演示。

Head to assemblyai.com/brief and try the live voice agent demo right there on the site.

无需注册。

No sign up needed.

这里有个残酷的事实：你的公司很可能正在花成千上万甚至上百万美元购买AI工具，但这些工具的利用率却低得惊人。

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

虽然有一半的企业已经引入了AI工具，但只有12%真正用它们创造了商业价值。

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

大多数员工仍然只用AI来总结会议纪要。

Most employees are still using AI to summarize meeting notes.

如果你在公司里负责推动AI落地，那你一定需要Section。

If 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企业AI转型的平台。

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

它指导员工掌握真实的应用场景，追踪谁在用AI创造业务价值，并清晰展示AI在哪些地方真正产生了效果，哪些地方没有。

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.

结果是什么？

The result?

你不再只是部署工具，而是真正推动可衡量的AI价值。

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

你的员工将不再只做会议纪要，而是去解决实际的业务问题。

Your employees move from meeting summaries to solving actual business problems.

你还能清楚证明投资回报率。

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.

访问 sectionai.com 了解 Section。

Check out Section at sectionai.com.

那就是 sectionai.com。

That's sectionai.com.

那大公司呢？

But what about the biggies?

我一开始就说，可能最值得关注的两个问题是：谷歌会不会推出一个顶尖模型，以及在代码助手这一块会不会出现真正能跟 Codex 和 Claude Code 抗衡的竞品。

I said coming into this that maybe the two biggest questions were whether Google was going to release a state of the art model and whether there was going to be any clarity on the harness side of things and a legitimate competitor to Codex and Claude Code.

说白了，连这个问题都存在，本身就是对谷歌的一种批评。

To sum it was an indictment of Google that this was even a question.

NVIDIA 的分析师泰·金写道：媒体圈普遍对德米斯·哈萨比斯推崇备至，但谷歌在最具潜力的智能体 AI 市场上，正全面落败。

NVIDIA focused analyst Tay Kim wrote, the media establishment consensus is enamored with Demis Hassabis, but Google is soundly losing in the biggest product market fit Agentic AI market.

那我们到底发现了什么？

So what did we discover?

我们确实看到了谷歌智能体交互界面的更新——Antigravity。

Well, we did get an update to Google's agentic coating surface, Antigravity.

团队写道：推出 Antigravity 2.0，一款全新的独立桌面应用，真正实现了当初那个智能体优化体验的愿景，从零重建，支持多智能体协作、任务调度、原生语音交互，以及一键接入其他谷歌产品。

The team writes, Introducing Antigravity 2.0, a new standalone desktop application that delivers fully on that original glimpse of a truly agent optimized experience, rebuilt from the ground up with multi agent teams, scheduled tasks, native voice and one click integration with other Google products.

所以，看起来这应该是他们推出的智能编码竞品。

So this, it appears, is meant to be the Agentic coding competitor.

但有点让人困惑的是，他们同时发布了Google AI Studio的若干新Vibe Code功能，我又觉得，内部可能很清楚：Antigravity对标的是Claude Code，而AI Studio更像Claude CoWork，或者可能是Lovable的版本。

Adding some confusion to that, though, was the fact that they also announced a number of new Vibe Code features for Google AI Studio, and I'm again left feeling like it might be clear ish inside Google that Antigravity is the Claude Code equivalent, while AI Studio is the Claude CoWork equivalent or maybe the Lovable equivalent.

不过，不管是不是这么想的，他们的公开宣传完全没说清楚。

Anyway, if that is the conception, it's certainly not clear from their public communications.

为了展示Antigravity 2.0的新能力，他们用93个子代理、处理了数十亿个token，重新构建了一个可运行的操作系统核心框架，整个过程花了大约十二个小时。

To demonstrate the new capabilities of Antigravity two point o, they rebuilt the core framework of a working operating system using 93 sub agents and processing billions of tokens with the entire operation taking about twelve hours.

开发者们看到这个演示后，第一反应有几个。

Now there are a couple first reactions when developers saw this.

系统短暂崩溃了，体验确实不太好，但考虑到任何产品发布当天都难免有网络限流，大家还是能理解的。

It went down briefly, which wasn't the best experience, but I think everyone can give them a pass for that given that any sort of launch day is always gonna involve some throttling the network.

但更让人质疑的是，很多人觉得Antigravity跟Codex这类工具太像了，感觉像是照搬的。

But even less positively was chatter around the similarity and derivative feeling of antigravity to other tools like Codex.

更尴尬的是，在Antigravity发布视频的第二分钟，屏幕上赫然出现了一个叫Codex的文件夹，大家一看都摇头，心想这怎么都没发现，结果反而让Antigravity 2.0和Codex之间的对比更明显了。

This was made worse by the fact that in the second minute of the launch video for antigravity, you can see a folder for Codex on the screen being demoed, which made people shake their heads with wonder that that wasn't caught and certainly invited even more comparison between Antigravity two and Codex.

Codex团队自己也不避讳，直接从他们那边传话出来：‘我真想知道Antigravity团队有没有设计师。’

The Codex team themselves weren't shy about sharing their feelings with TPU from that team saying, I wonder if the Antigravity team has designers.

今天简直不敢相信自己的眼睛。

Couldn't believe my eyes today.

这可真是对Codex团队极大的肯定。

Very flattering to the Codex team.

不过，那些不太在意和Codex做比较的人，倒是普遍觉得，单从第一眼来看，Antigravity确实朝着智能代理系统该有的方向进化了。

But I would say that others who weren't as concerned with the comparison to Codex did seem to think that, at least on first glance, antigravity had evolved in the way that an agentic harness needed to evolve.

马克·克雷奇曼写道：Antigravity 2.0挺有意思，因为它不再像Google做了一个AI开发环境。

Mark Cretchman wrote: Antigravity two point zero is interesting because it no longer feels like Google made an AI IDE.

Antigravity 1.0是完整的IDE，包含编辑器、终端、浏览器和代理工作区。

Antigravity one point zero was the full IDE, editor, terminal, browser, agent workspace.

说白了，就是Google对智能编码的全套解决方案。

Basically, Google's take on agentic coding as a complete environment.

而2.0版本的感觉更像是，他们把代理系统从IDE里抽出来，单独做成了一个产品。

Two point zero feels more like they pulled the agent system out of the IDE and made it the product.

它更像是一款综合应用，有桌面端、命令行工具、SDK、托管代理、定时任务、子代理，还整合了AI工作室、Android和Firebase。

It feels more like the Codex app, desktop app, CLI, SDK, managed agents, scheduled tasks, sub agents, integration with AI Studio, Android, and Firebase.

IDE还在，但已经不是重点了。

The IDE is still there, but it's no longer the main story.

重点是代理层。

The agent layer is.

我确实没看到有人认为它在任何方面超越了Claude Code或Codex。如果要打分的话，最多只能说我们迎来了一次切实的智能代理架构升级，至少让谷歌勉强跟上了潮流。

What I certainly didn't see is anyone arguing that it had surpassed Claude Code or Codex in any way, so if we're keeping track of the score, at best you have to say that we did get a meaningful Agentic harness upgrade that at least sort of brings Google into the realm of parody.

那模型这边呢？

So what then about the model side?

这次发布的新模型是3.5 Flash。

The new model premiered was 3.5 Flash.

没有Pro版，他们说Pro版以后会推出，这次只有Flash版。

There was no Pro version, they say that's coming later, just the Flash version.

你会注意到，这里的‘Flash’已经不是以前那个意思了。

And as you'll see, Flash doesn't exactly mean what it used to.

或者说，既然过去Flash意味着速度和成本兼顾，现在它显然更侧重于速度了。

Or at least, to the extent that Flash used to be about both speed and cost, it is definitely more about speed now.

单看评测数据，这似乎是谷歌目前最强的模型。

Now just going by the benchmarks, it does seem like this is now Google's most powerful model.

在Terminal Bench 2.0上，它得了76.2%，高于Gemini 3.1 Pro的70.3%，超过了Opus 4.7，但没赶上GPT-5.5。

It scored 76.2% on Terminal Bench two point o compared to 70.3% for Gemini 3.1 Pro, but beating Opus 4.7 but falling short of GPT five five.

在SuiteBench Pro上，它得了55.1%，比Gemini 3.1 Pro略高一点，但还是明显落后于5.5和Opus 4.7。

On SuiteBench Pro, it scored a 55.1%, which was a slight jump over Gemini 3.1 pro, but still significantly behind five five and Opus four seven.

在某些智能体评测中，Gemini 3.5 Flash的数值表现已经达到了前沿水平。

On certain agentic benchmarks, Gemini 3.5 Flash appears by the numbers state of the art.

在Computer Use基准测试OS World中，它的表现和5.5、4.7不相上下。

For the computer used benchmark OS World, the model is neck and neck with five five and four seven.

在GDPVal上——这个指标衡量的是有经济价值的实际工作任务——它还远未达到顶尖水平，但相比Gemini 3.1 Pro已经有了明显提升。

On GDPVal, which is a measure of economically valuable real world knowledge work tasks, it is not close to the state of the art but still a significant jump from Gemini 3.1Pro.

总体来看，根据这些评测，这个模型确实够用，但还没达到像Opus 4.7和5.5那样全面领先的顶尖水准。

Generally by the benchmarks, it looks like the model is going to be competent but not nestled up against Opus four point seven and five point five as truly cross the board state of the art.

当然，以前的Flash模型从来不是以纯算力为重点，我一直说，它们主打的是速度和成本。

Of course, raw intelligence was never the focus of previous Flash models, which were always, as I said, about speed and cost.

而这次，他们显然是想把速度推到极致。

And in this case, they definitely wanted to hammer the speed.

3.5 Flash的速度大概是3.1 Pro的三倍，但性能却相差无几。

3.5 Flash is around three times faster than 3.1 Pro while delivering similar performance.

它比之前的3 Flash还要快上大约60%。

It's also around 60% faster than three Flash.

问题是，它并不便宜。

The problem is that it's not all that inexpensive.

相比上一代Flash，价格涨了三倍，比起2.0 Flash更是翻了二十倍，而且运行一些大型基准测试——比如Artificial Analysis的那些——成本甚至比Gemini 3.1 Pro和GPT-5.0.5 Medium还高。

The cost is about a three x increase since the last Flash model and about a 20x increase since two point zero Flash, and the cost to run many of the big benchmark tests, like those from artificial analysis, was actually higher than other models like Gemini 3.1 Pro and even GPT-five 0.5 medium.

早期体验的人觉得这玩意儿有点怪。

People at early access described it as a little strange.

彼得·戈斯特夫说，这是一次挺奇怪的发布。

Peter Gostev wrote, it's a pretty weird release.

它做的远比你要求的多得多。

It does way more than what you asked for.

有时候能产出顶尖水平的结果，但有时候又会突然崩溃，干出些怪事来。

It sometimes generates best in class stuff, but sometimes crashes out and does something strange.

玩游戏还行，但比不上3.1和Controls。

It can be decent at games but weaker than 3.1 and controls.

虽然在三维世界这类任务上表现不错，但在网页界面方面却很差，甚至不如一些开源模型。

While it is good at things like three d worlds, it is really quite bad at Web UI, worse than open source models.

价格方面，当然最让人费解。

The pricing, of course, is the weird bit.

它直接升级成了专业版，结果丢掉了Flash原本让人喜欢和看重的那些特点。

It went up to pretty much become a pro model, so losing a lot of what Flash was known and loved for.

西蒙·史密斯写道：初用Gemini 3.5 Flash的感受。

Simon Smith writes: Initial experience with Gemini 3.5 Flash.

聪明，快得惊人，就是太啰嗦。

Smart, crazy fast, but quite verbose.

比如，我选中了自己写的一段话，问它：‘这段话通顺吗？’

Example, I highlighted some text I wrote and asked, Does this make sense?

结果它回了568个词，还附上了一整套法律分析。

The reply was five sixty eight words and included a legal analysis.

这显然不是个例。

And this doesn't seem to be an isolated experience.

当你去看那些用来跑AI指数测试的输出token数量——这基本就是衡量效率的指标——虽然3.5 Flash远不是最耗token的推理模型，但它比3.1 Pro效率还低，甚至比所有GPT-5 0.5模型都低，除了Extra High版本，它的消耗和那个差不多。

When you go look at the output tokens that were used to run the artificial intelligence index tests, which is basically a proxy for token efficiency, while 3.5 Flash is nowhere near the top of the most token hungry reasoning models, it is less token efficient than 3.1 Pro, and in fact less token efficient than any GPT-five 0.5 model except extra high, where it's in a very similar range.

在这些测试里，它用的token量是GPT-5 0.5 Medium的3.5倍，而这正像越来越多的人指出的那样，严重质疑了它‘速度快、成本低’这个卖点。

It used about 3.5 times more tokens for those tests than GPT-five 0.5 medium, which as many are starting to point out, is kind of an indictment of the value proposition of speed and cost.

如果你确实快，但要花3.5倍的token才能得到同样的答案，那速度优势就大打折扣，成本优势更是被彻底吃掉。

If you're really fast, but it takes you three and a half times as many tokens to get to the same answer, that cuts into a lot of the speed gains, and it also eats alive any cost gains.

Theo在t3上写道：结果就是，做类似任务时，它的运行成本是Gemini 3.1 Pro的两倍。

Theo from t three wrote, The result of this is it costs two times more to run than Gemini 3.1 Pro on similar tasks.

它的价格甚至比GPT-5 0.5 Medium还高。

It's more expensive than GPT-five 0.5 medium.

事实上，Theo用得越多，就越不喜欢这个模型。

And in point of fact, the more that Theo used it, the less he liked the model.

他上传了一段自己承认崩溃的视频，展示它在实际代理任务中的表现有多差，这无疑是最严重的问题。

He posted a self admitted crash out video about just how badly it performed on actual Agentic tasks that he tested it on, which is certainly the most problematic on top of all these other issues.

Tenebrous写道：到目前为止，对3.5 Flash的印象相当负面。

Tenebrous writes: So far, pretty negative impression of 3.5 Flash.

它在输出令牌的速度上确实很快，但根本没用，因为几乎每个任务都会疯狂地触发大量不必要的工具调用。

It is very fast in terms of token output, but this basically doesn't matter because it explodes in a huge avalanche of unnecessary tool calls on basically every task.

一旦遇到卡壳，它几乎从不暂停或寻求帮助，就是一味地硬冲，乱撞一气。

When it gets stuck on something it seems to pretty much never pause or ask for help, it just kinda keeps steamrolling ahead and flailing.

经常虚构不存在的缩写含义，文字质量中等偏下，满屏表情包，代码质量也就和SONNET差不多。

Frequently hallucinated fake acronym expansions, writing quality as mid to bad, tons of emoji slop, actual code quality as SONNET tier.

这还只是初步的感觉，可能我漏掉了什么。

This is a very early vibe check, I could be missing things.

但就连‘超快速代码探索’这个子代理的初始用途，对我而言也很快变得不靠谱了，因为它根本不够聪明，快不起来。

But even the initial use case of Super Quick Codex Exploration sub agent is is pretty quickly dissolving for me because it's not actually smart enough to be quick about it.

总的来说，这绝对不是谷歌该推出的版本。

All in all, definitely not what Google needed to drop.

有趣的是，如果把视角拉远一点，把宣传重点全放在速度上，其实和开发者当前真正关心的问题脱节了。

And interestingly, zooming out a bit, the decision to focus so much of the messaging around speed seems kind of out of sync with the reality that speed is not the big thing that's an issue for developers right now.

真正的问题，当然是成本。

That issue is, of course, cost.

我一开始就在想，谷歌会不会干脆全力押注一个不追求顶尖性能、但更节省token、价格更低的模型？

One of the questions that I had coming into this was whether Google would lean all the way into the potential to compete not on the full state of the art, but on a much more token efficient, lower priced model.

按理说，3.5这种所谓的‘闪存’模型就该是这个方向，但现实情况却不是这样。

And although that's kind of what you would assume a 3.5, quote unquote, flash model would bring, that's not exactly how it appears in the real world.

不过，谷歌也悄悄承认了，其他行业正在发生的那些商业模式变化，同样影响着他们。

Now there was some quiet acknowledgment of the same sort of forces that are changing business models everywhere else.

他们大肆宣传把Ultra套餐从每月250美元降到200美元，还推出了新的100美元档位。

Google made hay about the fact that they had dropped the price of Ultra from 250 to $200 a month, as well as offering a new $100 plan.

但作为Ultra用户，我在一封官方邮件里看到，他们也提到，Ultra套餐现在会引入基于计算量的使用限制，会综合考虑你的提示复杂度、使用的功能以及对话长度。

But in an email describing the changes, which I got as an Ultra user, they also wrote that the Ultra plan would now include, quote, compute based usage limits that factor in the complexity of your prompt, the features you use, and the length of your chat.

此外，包括Flow Design Platform和Antigravity在内的智能工具，现在也将转为按使用量计费的模式。

In addition, Agentic tools, including the Flow Design Platform and Antigravity, would now switch to a usage limit model.

说白了，虽然他们降低了基础价格，但在某些情况下，你现在得为以前免费的使用行为额外付费了。

Basically, while they're bringing down the base price, you are now, in certain instances, going to have to pay on a usage basis when you didn't before.

不过我觉得这并不能说明谷歌做错了什么。

Now this, I don't think, is an indictment of Google or anything.

事实上，这恰恰反映了整个行业的发展趋势。

In fact, it's an acknowledgment of where everything is going.

但这也让谷歌把速度作为3.5版本最主要的竞争卖点，显得更加值得商榷了。

But it does make the decision to focus so much on speed as the big competitive differentiator of 3.5 all the more questionable.

我并不觉得巧合的是，就在同一天，OpenAI推出了他们的新保障算力计划，本质上是让客户在算力短缺加剧的未来，也能确保长期稳定的资源访问权。

I don't think it was any accident that on the same day, OpenAI introduced their new guaranteed capacity program, which is basically a way for customers to guarantee long term access to compute even with even worse compute shortages on the horizon.

虽然他们在IO大会上表现得不够出色，但我还是认为，这对谷歌来说依然是个机会。

And I will say that even though they didn't seem to nail it at IO, I do think that this is still an opportunity for Google.

Box的亚伦·莱维提到，未来企业应用AI时，令牌成本将成为一个核心议题。

Box's Aaron Levy writes, token costs will become a dominant topic in enterprises going forward with AI.

刚和一群《财富》500强企业的CIO吃晚饭，讨论得最激烈的就属这个话题了。

Just got out of a dinner with many Fortune 500 enterprise CIOs, and this was the most heated topic.

大家用的策略五花八门，但说到底，没人觉得自己找到了对的办法。

A mix of strategies are being employed, but basically no one feels like they have the right solution.

大家都在摸索怎么分配任务给不同模型，按用户类型开放不同能力的代理，给各个团队设预算上限，让团队用实际场景证明AI的价值，也有人干脆放任自由使用。

It's gonna be a mix of figuring out how to prioritize workloads to different models, giving out access to better or worse agents by user type, setting different spend caps by team, having teams justify AI by their use case, and some just having unfettered access.

现在所有人都在试着找一个半稳定、可预测的模式，可底层技术和计费方式却一直在变。

Everyone is trying to figure out a semi or predictable model right now in a world where the underlying tech and cost models are constantly evolving.

抛开对某个具体产品或刚发布模型的第一反应，我注意到大家的讨论主要集中在两个方向上。

Overall, holding aside first reactions to any one product or the model that was announced, there were two big strands in the sentiment that I saw.

第一个是：天哪，工具也太多了吧。

The first was OMG product sprawl.

西蒙·史密斯又提到：今天谷歌这一连串发布让人兴奋，但也让人晕头转向。

Simon Smith again writes, The barrage of Google announcements today are exciting, but also confusing.

现在有Google Images、Google Photos、Google Pics，还有Antigravity，再加上Spark。

So now there's Google Images, Google Photos, and Google Pics, and there's Antigravity, but also Spark.

他在另一篇帖子中补充说：谷歌I/O让我头都大了。

In another post, he followed up: My head is spinning from Google IO.

这让我想向OpenAI和Anthropic提个请求。

It leads me to a request for OpenAI and Anthropic.

请别搞得太分散了。

Please avoid Sprawl.

就给我一个强大的智能工具，像Codex或者CoWork那样，让我能做完所有事。

Just give me a single powerful agentic tool like Codex or CoWork through which I can do everything.

我不想每次都纠结该用Spark、Antigravity、AI Studio、Flow、Pomelli还是Pix。

I don't want to have to think about whether to use Spark or Antigravity or AI Studio or Flow or Pomelli or Pix.

我明白这些产品可能适合不同用户，谷歌也擅长做生意，但 personally，我就想要一个统一的界面，我愿意为这份简单买单。

I accept that these may meet the needs of different users and Google knows how to run a killer business, But personally, I just want one interface to rule them all, and I'm willing to pay for that simplicity.

再往外看，不只是AI圈内的人，Marques Brownlee也说：现在要记住所有新发布的AI产品名字，真的越来越难了。

Moving outside of just the pure enfranchised AI circle, Marques Brownlee writes: It's getting genuinely difficult to keep track of all the names of AI products being unveiled.

就在刚才一小时里，谷歌发布了Google Pics——这可不是Google Photos，还更新了Google Flow、Nano Banana、VO，全是做媒体生成的，还有Google Antigravity、Gemini Spark、Gemini Omni、Gemini 3.5 Flash。

In the last hour, Google unveiled Google Pics, which is not Google Photos, and updates to Google Flow, Nano Banana, VO, which are all media generation, Google Antigravity, Gemini Spark, Gemini Omni, Gemini 3.5 Flash.

内森·克拉克一脸调侃地总结道：这功能就在吉米尼里。

Nathan Clark summed up with his tongue firmly buried in his cheek, It's in Gemini.

刚在AI工作室里创建的。

Just created in AI Studio.

哦，这个是给你的个人Google One账户用的。

Oh, that's for your personal Google One account.

如果是Workspace，那你得用吉米尼商业版。

For Workspace, you need Gemini Business.

不是。

No.

也不是吉米尼高级版。

Not Gemini Advanced.

现在叫AI Pro了。

That's AI Pro now.

除非你需要AI Ultra。

Unless you need AI Ultra.

哦，智能代理？

Oh, agents?

其实你在 Spark 里就能做这个。

You do that in Spark, actually.

写代码的话，用 Joules。

For coding, use Joules.

除非你说的是那个智能开发环境。

Unless you mean the Agentic IDE.

那个叫反重力。

That's antigravity.

不是。

No.

那是旧版的反重力。

That's the old antigravity.

去下载新版的。

Download the new one.

实际上，Gemini 命令行工具已经被弃用了。

Actually, Gemini CLI is being deprecated.

用 Antigravity 的命令行工具吧。

Use antigravity CLI.

不对。

No.

Flash 模型比 Pro 模型更聪明，除非你非得用 Pro。

The Flash model is smarter than the Pro model unless you need Pro.

如果是处理视频，就用 Flow。

If it's video, use Flow.

不对，Flow 用的是 VO。

No, Flow uses VO.

其实，那个功能现在已经在 Gemini 里了。

Actually, that's in Gemini now.

除非你在用搜索功能，那时候用的是 AI 模式。

Unless you're in search, then it's AI mode.

不，研究功能是集成在笔记本里的。

No, research is notebook.

模型。

Lm.

总之，这一切其实都很简单。

Anyway, it's all very simple.

所以一方面，即使这些产品都不错，用户还是感到一头雾水。

So that is one take that even if these products are good, everyone is left confused.

但另一方面，这可能也没那么重要。

The other take though is that it might not matter.

谷歌拥有庞大的用户基础，我们日常与谷歌的数字互动无处不在，这种广泛的存在感，意味着只要把合适的产品版本放在用户已经活跃的地方，对普通用户来说就足够了。

The sheer surface area that Google has with its users, the total amount of digital interactions where we already interface with Google, may just mean that product sprawl that puts the right version of the thing in the right place where someone's already interacting is for the average consumer all that's going to matter.

在这方面，我们昨天看到的那些数据确实很难反驳。

And on that front, it's hard to argue with some of the numbers we got yesterday.

Gemini 应用的月活跃用户数，从去年五月的四亿，增长到了今年四月的九亿。

The Gemini app has jumped from 400,000,000 monthly active users back in May last year to 900,000,000 users last month in April.

在同一时间段里，谷歌所有平台处理的月度令牌数量从48万亿飙升至3.2千万亿。

In that same period of time, the number of monthly tokens processed across all of their surfaces has jumped from $480,000,000,000,000 to 3.2 quadrillion per month.

还有一点关于开放通道的问题。

And there's also that point about the open lane.

彼得·杨写道：我觉得谷歌会在消费级AI领域胜出。

Peter Yang wrote: I feel like Google is going to win consumer AI.

它是唯一一家在开发视频模型的美国实验室，而消费者就是爱视频。

It's the only US lab that's building video models and consumers love video.

比如，TikTok和YouTube的受欢迎程度远超文字类平台。

EG, TikTok and YouTube is far more popular than text based platforms.

真正唯一的竞争对手是Seedance这类不care版权的视频模型。

The only real competition is Seedance and other video models that don't care about copyright.

法尔扎德写道：我觉得谷歌已经赢下了消费级AI市场。

Farzad writes: I think Google just won the consumer market for AI.

我觉得接下来会发生的是：Anthropic擅长做企业级模型，而谷歌擅长做普通人和创作者用的模型。

Here's what I think is going to happen: Anthropic, the best models for running businesses Google, the best models for everyday people and creatives.

SpaceX 是远超其他公司的超大规模服务商。

SpaceX, the biggest hyperscaler by far.

OpenAI 恐怕要凉了，除非他们能突然拿出什么惊人之举。

OpenAI probably cooked unless they pull a rabbit out of a hat.

现在有几个人问，他到底有没有实际使用这些产品，还是只是发了个博取关注的推文。

Now a couple people asked whether he actually used these products or whether this was just an engagement tweet.

但说实话，就连一条把 Google 当成消费级 AI 市场赢家的推文，都足以说明问题。

But honestly, the fact that an engagement tweet placing Google as the default winner of the consumer market says something in and of itself.

Hater 说，Google 已经牢牢掌控了自己的生态体系，正在下长远的棋。

Hater writes, Google has secured its ecosystem and is playing the long game.

Gemini 已经支撑着搜索、Workspace 等多个产品，所以 Google 不再需要补贴模型来吸引用户。

Gemini already powers search, workspace, and more, so Google no longer needs to subsidize models to attract users.

Gemini Omni 可能不是最强的编程模型，但它会成为未来通用多模态模型的基础。

Gemini Omni may not be the best coding model, but it will become the base for future universal multimodal models.

而且说实话，当我们逐渐接近尾声时，我想再把话题拉回到 Omni 上。

And actually, as we start to round the corner, I want to bring it all the way back to Omni.