Anthropic launches Claude Science and a new model

如果你想免费收听这一集以及我接下来的30集，全程无广告，那就去听听我的播客《AI Chat》吧。

If you wanna get access to this episode and my next 30 episodes all ad free, so there'll be no ads on them, go check out my podcast AI Chat.

你可以在Spotify或者Apple上搜索这个名字。

You can go search for that on Spotify or Apple.

就叫《AI Chat》。

It's AI Chat.

我会把所有这些新闻类节目都发出来，还会发布一些访谈内容。

I'm gonna post all of these news episodes, and I'm also posting interviews.

我刚采访了Coher的首席执行官。

Like, I just interviewed the CEO of Coher.

他们为自己的AI模型融了超过十亿美元，聊了钱打算怎么花，AI行业未来会怎么发展，还有各种新动向。

They've raised over $1,000,000,000 for their AI model, talking about what they're gonna be spending the money on and and the direction of the AI industry along with all of this new stuff.

所以如果你想去听，完全免费还无广告，那就是《AI Chat》。

So if you wanna go check it out with no ads for free, it is AI Chat.

Anthropic正在推出Claude Science。

Anthropic is launching Claude Science.

这将为研究人员提供一个工作平台。

This is gonna be a workbench for researchers.

Base44 正在推出自己的AI模型。

Base forty four is launching their very own AI model.

他们想保护每年一亿美元的持续收入，来自他们的 vibe 编程业务；而 X 刚刚上线了托管的 MCP 服务器。

They're trying to protect their a $100,000,000 in annual recurring revenue for their vibe coding business, and X has just launched a hosted MCP server.

所以他们会向 Claude、Cursor、Grok、Build 开放 API，如果你了解 X 和他们 API 的背景，这可是个不小的动静。

So it's gonna open its API to Claude Cursor Grok Build, which if you know anything about the the backstory of X and their API, this is quite, quite a big deal.

Anthropic 正在发布 Claude Sonnet 5。

Anthropic is shipping Claude Sonnet five.

每百万输入 token 收费两美元，同时 Google 也在向开发者推出 Nano Banana 2 Lite 和 Gemini Omni Flash。

It's gonna be $2 per million input tokens, and Google is shipping Nano Banana 2 Lite and Gemini Omni Flash to developers.

如今几乎所有主流 AI 模型都在密集发布新版本。

So tons coming out from basically every top AI model today.

我们接下来会一一细说这些内容。

We're gonna get into all of that.

希望我这边别太吵。

Hopefully, it's not gonna be too noisy from my end.

我现在在北卡罗来纳，天气又热又晒，正抱着个宝宝呢。

It is a hot, sunny day in North Carolina where I'm at right now, and I am currently bouncing a baby on my lap.

我老婆昨晚没怎么睡，我得让她多休息会儿，所以抱着孩子。

My wife did not sleep a lot in the last night, so trying to let her get some rest, holding the baby.

你可能以前在播客里听过我说过这话。

You probably heard me say this before on the podcast.

如果你听到背景里有点咕噜咕噜的声音，那就是宝宝在那儿呢。

But anyways, if you can hear a little, gurgling and gooing in the background, that would be what it's from.

如果你有重复性的AI任务，比如写通讯、发邮件、生成图片之类的，欢迎去试试我的创业项目 aibox.ai。

If there's any repetitive AI tasks that you do, like writing newsletters or emails or generating graphics or anything you do on a repeated basis, I would love for you to go check out my startup, aibox.ai.

我们有个零代码的AI应用构建器，你只要说清楚你想做什么，系统就会自动把不同的AI模型连起来，帮你搭建一个能反复使用的流程。

We have a no code AI app builder where you basically just explain what you want it to do, whatever your workflow is, and it automatically links together different AI models and creates a workflow for you that you can use on a repeated basis.

这帮我省了不少时间，AI确实能大幅提升效率，但如果你每天都在重复做同样的事，不如干脆把它自动化了。

This saves me a lot of time because AI definitely is a huge unlock for being more productive, but at the same time, if you're just using it to do the same thing over and over every single day, you should probably just automate that.

如果你想试试，就去 aibox.ai/builder 看看。

So if you wanna give it a try, go to aibox.ai/builder.

我会在简介里放个链接，方便你直接访问我们的构建平台。

I'll leave a link in the description to go check out the builder platform that we have.

我们提供了超过80种不同的AI模型，可以自由组合，帮你打造超强的自动化流程，省下大量时间。

We have over 80 different AI models that you can link together and build some really incredible automations to hopefully save you a ton of time.

起步价每月只要8.99美元。

And it's only $8.99 a month to get started.

简介里我会再放个链接。

I'll leave a link in the description.

先从Anthropic开始聊起。

Kicking it off with Anthropic.

他们刚推出了Claude Science。

They have just launched Claude Science.

这会是一个工作台。

So this is gonna be a workbench.

它连接了60多个科学数据库，还整合了一整套预置的科研工具，涵盖基因组学、蛋白质结构、化学等领域，全部运行在Claude Opus 4.8上。

It's connecting 60 different scientific databases and a bunch of prebuilt research tools for genomics, protein structures, chemistry, and all of this is gonna run on Claude Opus 4.8.

抱歉，Fable Five还没上线。

So sorry, Fable five, not there yet.

等它推出后，我们可能也会把它整合进来。

I'm sure when that comes out, we may get that rolled in as well.

还会赠送三万美元的免费算力额度。

And it's gonna come with $30,000 in free credits.

他们打算把这些资源提供给50个学术项目，相信通过优化工作流程，比单纯堆算力更能吸引研究人员使用Anthropic。

They're gonna be giving this way to 50 different academic projects, and they're making this bet that basically workflow is going to beat raw capacity for getting these researchers to be using Anthropic.

其实没有推出新模型，Claude Science用的还是同一个Claude Opus 4.8。

So there's no new model or dating, but Claude Science uses the same Claude Opus 4.8.

任何人都能用，没有针对生物学做专门调优，也没有专用权重。

It's basically available to everybody, and there is no biology fine tuning or specialized weights.

他们已经和一些用户做了测试，效果相当不错。

They've been testing this with a few different people and have had some impressive wins.

他们有个来自加州大学旧金山分校脑肿瘤中心的团队，正在用它来压缩基因组和种系分析的工作流程。

They had UCSF Brain Tumor Center that was using it to compress genoma, germline analysis workflows.

还有艾伦研究所的杰罗姆·拉科克，他在搭建一个计算型文献综述流程，以及诺和诺德，他们是首批合作方之一。

We had the Allen Institute's Jerome Lacock, who is building a computational review pipeline, and we have Novo Nordisk, who's one of their launch partners.

它的实际功能是能快速启动多智能体架构，让科研人员同时创建多个并行的辅助任务，进行序列分析。

What it's actually doing is that it's able to spin up these multi agent architectures, and they're gonna let all of these scientific researchers create parallel sub assistance, and they're gonna be doing sequence analysis.

进行结构预测。

They're gonna be doing structural prediction.

进行脂肪含量检测。

They're gonna be doing fat checking.

这项技术的强大之处在于，所有这些操作都能在确保完全可重复性的前提下同时进行。

The reason why this is so powerful is because all of that can be happening while they're also keeping their full reproducibility.

代码、运行环境、对话历史，所有这些都能完整复现。

So the code, the environments, the message history, all of that they're gonna be able to reproduce.

科学家们需要很多定制化工具，而Anthropic已经把这些都内置进来了。

So there's a bunch of basically custom tools that scientists need, and Anthropic has built that in.

他们正在发放大量资助，希望让大家意识到，搞科研的话，Anthropic 是最好的选择。

They're giving away a bunch of grants hoping to, show people that this is the you know, Anthropic is the best place if you're doing science.

明白了。

Okay.

Base44 要推出自己的 AI 模型了，说起来我之前都不知道，Base44 其实是 Wix 旗下的。

Base forty four is launching their very own AI model, and the reason why is by the way, I didn't realize, but Base forty four is owned by Wix.

那是个特别棒的编程平台。

It's an incredible vibe coding platform.

我用过之后真的很 impressed。

I've been really impressed with it.

我平时主要用 Claude Code，几乎什么事儿都靠它，所以不是我首选。

It's not the one that I primarily use because I'm just using Claude code for mostly everything I do.

但如果你刚入门，我觉得 Level 和 Base44 这两个我都特别推荐。

But if you're getting started, I think anywhere between Level and Base 44, those are my two favorite.

Anywhere 也是个很棒的平台。

Anywhere is also a great platform.

不过话说回来，Base 44 年经常性收入高达一亿美元，他们刚推出了全新的模型 Base One。

But in any case, Base 44 is a $100,000,000 in annual recurring revenue, and they just launched their brand new model called Base one.

这是他们自家的大型语言模型。

So this is their very own LLM.

这其实和 Cursor 的做法很像，不过这个模型是用数千万次真实用户互动训练出来的。

This is actually very similar to what Cursor is doing, by the way, but this LLM is trained on tens of millions of real user interactions.

有意思的是，因为他们有数千万用户在用他们的工具，所以能收集到大量的对话数据。

So what's interesting is because they are able to, you know, have tens of millions of users using them to create their tools, they can take all of those conversations.

我不确定是不是所有用户都包含在内，还是说用户需要主动同意。

And I'm I'm not sure if it's every single user or if they had people opt in.

估计是默认所有用户都包含在内，毕竟这写在服务条款里了。

It's probably every single user, and it's just part of their terms of service.

他们把用户所有的操作和反馈都收集起来，用来训练模型，精准优化人们在开发网站、应用和软件时的真实需求。

But they're taking everything that those people have been saying, and they're training their model to basically fine tune what people need when they're building websites, apps, and software.

我觉得最厉害的是，他们清楚知道，当有人想开发某个功能时，比如 Anthropic 或者 OpenAI 的 Codex，理论上它们也该有这能力——但 Base 44 真实掌握了每次用户提出需求后，紧接着会追加哪五个操作。

Now what's interesting to me about this is they know exactly like, when someone's building software and, you know, theoretically Anthropic and Codex, like, for Open AI, I should have this as well, but they know every time that someone asks to build something and it does it, and then they give a follow-up, they're like, oh shoot, every time someone asks for this, you know, for x y z, they typically ask for these next five things.

这可能意味着，用户一开始并不清楚自己到底想要什么，或者后面总会跟着个bug，又或者有什么坑需要注意。

And that could either mean that when they first ask, they don't know exactly what they want, or there's traditionally a bug that follows, or there's a gotcha or something like that.

但关键是，他们能不断优化这个模型，让它提前预判用户真正的需求，甚至在用户还没开口之前就猜到他们的方向。

But essentially, they can fine tune this model that they have to know exactly what they want to what people are basically, the direction they're going before they even go there.

对吧？

Right?

然后用户说：

And they're like, hey.

我想做个健身应用。

I want like a fitness app.

模型就会想：大概率用户想要这五个功能。

Like, oh, these are probably the five features they want.

所以要么直接问清楚，要么直接开始搭建，要么先把架构搭好，随时能上手。

So either they'll ask them right away or they'll just start building stuff in, or they'll set up the architecture in a way that is ready to go.

用这种经过精细调校的模型，他们还能做不少有意思的事。

Now there's a couple cool things that they can do with these fine tuned models.

其中一个好处是，他们能让模型更高效，从而降低用户的使用成本，这当然很棒。

One of them being that they can make them much more efficient, and so they're able to cut the cost for their users, which of course is fantastic.

但另一方面，如果有人在和模型对话时说：‘你知道吗？’

But also, if someone's talking to their model and they're like, oh, you know what?

我们自己的基础模型，其实比调用 Opus 4.8 要强多了。

Our, you know, our base one model would be a lot better than sending it to Opus four point eight.

这样一来，他们就能把用户切换到更高效的模型，既帮用户省钱，又保留了专属数据。

All of a sudden, they're able to switch people over to the more efficient model, save the user money, and they're also keeping exclusive data.

对吧？

Right?

因为如果他们不把数据发给 OpenAI、Gemini 或 Anthropic，那这些公司就没办法用这些数据去训练模型了。

Because if they're not sending data over to, OpenAI or Gemini or Anthropic, all of a sudden, they're not able to train with all of that data.

我们也在 Cursor 上看到同样的情况。

So we're also seeing, we're seeing the same thing coming out of cursor.

很多不同的开发者工具在和代码互动时，其实都在悄悄捕捉代码和对话内容，理解用户真正想要什么，然后用这些数据训练自己的定制模型。

So a bunch of different players that are interacting with code are basically intercepting the code and and the conversations people are having and knowing what people want, and they're being able to train their own fine tuned models with it.

所以Base One的意思是，这个模型专门用来替代Anthropic的Opus，处理应用生成类任务，能让他们对延迟、成本和推理速度有更直接的掌控。

So Base One, basically, they're saying this is designed to replace Anthropic's Opus for app generation workloads, and it's gonna try to give them a really direct control over latency, cost, and inference speed.

而且他们再也不用依赖任何第三方API了。

And they're not gonna have to rely on any third party APIs.

另外顺便说一句，这对他们来说是一笔不小的开支，因为当你订阅Base 44时，你付的token费用里，很大一部分其实直接流到了Anthropic手里。

And also, by the way, this is a big cost for them because when you get a subscription to base 44, you pay them for tokens, but a huge chunk of your token bill just goes straight to Anthropic.

对吧？

Right?

相当于他们把一大块钱白白送人了，但如果能自己在服务器上跑，再找个成本更低的方式，他们反而能赚得更多。

It's like they're they're given a huge chunk, but if they can do it directly on their servers and if they have you know, a cost efficient way to do that, they're actually able to make more money.

我之前提过很多次的Lovable，是他们在瑞典的主要竞争对手，但他们到现在还在用外部的大模型。

So Lovable that I've mentioned many times is, of course, their rival in Sweden, and they still use external LLMs.

他们的年收入大约有5000亿美元，规模大概是我们的五倍。

They have about $500,000,000,000 in AR, so they're about five times bigger.

不过他们确实有很强的背景，因为Wix在他们才成立六个月、只有八个员工的时候，就花了八千万美元把他们收购了。

But they do have some really solid backing because Wix acquired them for about $80,000,000 when they were only six months old with eight employees.

他们的母公司现在正在裁员约百分之二十。

Their parent company is now cutting about 20% of their workforce.

所以你看，Wix 在缩减规模，但 Base 44 却在加速发展。

So, you know, Wix is cutting down, but evidently, base forty four is accelerating.

我的预测是，如果 Base 44 真能做成这件事，它最终的规模很可能会远超 Wix，这可真够惊人的，毕竟 Wix 当初只花了八千万美元就收购了它。

And my prediction here is if that if base forty four really pulls this off, base forty four will actually end up being much bigger than Wix, which is pretty wild considering they bought it for $80,000,000.

它现在的年经常性收入已经超过一亿美元了。

It's already doing more than a $100,000,000 in annual recurring revenue.

他们早就远远超过了当初的收购价。

So they were just, you know, blasted past even the acquisition price.

这对 Wix 来说真是笔超值的买卖。

This was an amazing deal for Wix.

好了。

Alright.

接下来我们要聊的是，X 刚刚推出了一个托管的 MCP 服务器。

The next thing we ought to talk about is that X has just launched a hosted MCP server.

如果你了解X，就会知道当初埃隆刚收购推特时，立马就关闭了API，因为我觉得OpenAI和其他一些公司当时就像接了一根消防水带，源源不断地用这些数据训练模型。

And if you know X, like they when they when Elon originally bought Twitter, he immediately shut down the API because I think OpenAI and a bunch of other people just had, like, a full fire hose blast of all of the API data, they were using it for training.

埃隆后来意识到这些数据有多值钱，就把它关了。

Elon shut that down realizing how valuable it was.

之后Reddit和其他平台也效仿了类似的做法。

We saw similar moves from Reddit and other players afterwards.

但现在X推出了新的MCP，他们打算开放API，让Claude、Cursor、Grok这些工具接入X的AI系统，甚至可以直接访问你个人的账号数据。

But now that X is kind of launching this new MCP, they're gonna let tools like Claude, Cursor, Grok, all tap into x AI's API and specifically your own account.

只要你有自己的账号和权限，就能允许MCP访问你的数据。

So if you have your own account and your own permissions, you can let the, you can let the MCP tap into that.

我敢说，用这个能做很多有意思的事，比如市场情绪分析之类的。

I'm sure there's a ton of, like, market sentiment analysis, you know, great things that you can do with this.

现在确实有公开的API了。

There is an API.

像对冲基金这种大机构，正想通过分析球迷对世界杯的情绪，来预测热狗销量，或者干一堆类似疯狂的量化交易计算。

So big firms, you know, like hedge funds that are trying to, you know, measure people's sentiment analysis on the FIFA World Cup to determine how many hot dogs are going to be sold and blah blah blah blah blah, and do all their, like, crazy, you know, hedge fund calculations.

这类功能本来就有API可用。

Those types of things, they already have an API.

他们接下来就会用上这个。

They're they're going to be using this.

这个主要是给普通用户准备的。

This is more for your average user.

MCP的意思是，你能把XAI的所有数据直接接入你正在用的Claude，适合普通用户使用。

An MCP means that you can get all of XAI's data into your Claude that you're actively using, so for regular people.

我个人很期待，真想把MCP连上，方便我在这里聊AI新闻的时候用。

I personally am excited and would love to connect the MCP for when I talk about AI news here.

我希望能每准备一个话题，就自动拿到相关的热门推文和高赞评论做个简报。

I would love to get, maybe, like, some sort of brief on every story I'm gonna cover with the top tweets and the top comments on those tweets.

要是MCP能帮我自动抓取这些，那就太有用了。

So if my MCP could go and pull that, that'd be really useful.

以前我总想让Claude去X上抓优质推文，但这事儿一直特别难搞。

I traditionally try to get, like, Claude to go and scrape x for good tweets is notoriously hard.

我一直很难做到这一点，只能自己手动刷X动态。

I've had a a hard time doing that, and I basically just have to go manually scroll x myself.

如果能省点时间，直接把最精彩的内容推给我，那就太棒了。

It would be fantastic if I could save some time and just, you know, surface the best things.

说白了，挺酷的是，如果你能把MCP接入X这样的平台——我知道我有点跑题了，但你可以想象，同样的方式也能用在Reddit或者其他社交平台上，你自己就能定制算法。我觉得Threads和扎克伯格现在也在试着让用户这么做，比如在Threads里，你可以让系统在接下来四周多展示某个特定话题的内容。

Basically, I mean, what's kinda cool is if you have an MCP into something like x, and I know I'm getting off the rails here, but you could imagine the same thing with Reddit or any other social platform, you can fine tune your own algorithm, which I think threads and Mark Zuckerberg's trying to let people kinda do that with threads, but it's like, in threads anyways, it's like, show more posts about this specific topic for the next like four weeks.

但这种设置又不是永久性的。

And then anyways, it's it's not like a permanent thing.

所以这个功能真的很吸引人。

So this would be cool.

我甚至都想重新打造自己的X信息流了，因为我虽然会互动，但每天刷到的政治内容实在太多了，根本不是我想看的。

I would honestly almost remake my own x feed because I probably get I I'm sure I interact with it, but I feel like I get, like, way too much politics in my thread, that I really wanna interact with in any given day.

这可能怪我自己，但要是能用MCP主动告诉X：我就想看AI新闻，那该多好。

So and it's probably my own fault, but it would be really cool to just intentionally be able to use this MCP on X and say, look, I just want AI news.

我就想关注这500家AI公司发布的动态，让信息流更聚焦、更干净，只留下真正相关的资讯。

I just want, you know, news from these, you know, 500 companies that I'm trying to follow in AI and get more of a more of a streamlined stream of content around that.

这对我来说挺棒的。

That'd be cool for me.

明白了。

Okay.

XAI 当然不是第一家搞 MCP 的大公司。

XAI obviously isn't the, you know, the first big company to do an MCP.

感觉现在几乎人人都在用 MCP，连我们公司 aibox.ai 也有了。

It feels like basically everyone has an MCP, even my own company aibox.ai.

我们也有个 MCP，以前大家会来我们网站，那里有八十多种 AI 模型，音频、文本、视频都有，用户可以在我们的测试平台里跟它们聊天，每月付九美元就能用所有模型。

We have an MCP, so, you know, traditionally people would go to our site, there's 80 different AI models, audio, text, video, and people would go chat with them in our playground and, you know, you pay like $9 a month and you get access to every model.

但很多人说：

But a lot of people are like, look.

我所有的活儿都是在 Claude 或者 ChatGPT 里做的。

All of my workload is is all of my work is happening inside of Claude or inside of ChatGPT.

所以我们搞了个 MCP，现在你可以直接通过它访问我们所有的工具。

So we created an MCP, Now and you can get access to all of our different tools.

你现在可以让Claude实现图像生成和音频生成，这些功能它以前是没有的。

You can get you know, like Claude can have image generation and audio generation, which it didn't have before.

所以MCP非常有价值，而XAI也不是第一个这么做的。

So MCPs are very valuable, and XEI is not the first one.

GitHub、Slack、Stripe、Salesforce，这些公司都已上线了官方的MCP接口。

GitHub, Slack, Stripe, Salesforce, all of them have official MCP endpoints.

如果你用x来发布内容，API的收费还是维持在每条公开帖子0.015美元，带链接的帖子大约20美分。

The API pricing, if you're gonna do that with x, is still gonna be 0.015 per published post and about 20¢ per post with links.

这个价格是今年早些时候定下来的。

That was set earlier this year.

他们明确说，这是为了大规模遏制垃圾信息。

They basically said they're trying to curb spam at scale.

但托管服务器现在展现出了一些新功能。

But the hosted server is showing some new capabilities.

现在支持搜索了。

There's search.

还能获取帖子内容。

There's post retrieval.

可以查询用户信息。

There's user lookup.

支持趋势分析。

There's trend analysis.

这些功能大大减少了开发者接入 X 平台的集成时间。

All of that is cutting down a lot of integration time for people trying to work with x.

我真的很期待亲自试试看。

So I'm gonna be excited to to test that out.

Anthropic 推出了 Claude Sonnet 5。

Anthropic is shipping Claude Sonnet five.

输入令牌的定价是每百万个 2 美元，比 Opus 4.8 还低，而且在代理式编程任务上的表现达到了 63.2%。

It's gonna be $2 per million input tokens, which is undercutting Opus 4.8, and it is about 63.2% on agentic coding.

它的表现跟 Opus 4.8 只差大约六个百分点。

So it's within about six points of Opus 4.8.

你可以想象一下，接下来的模型会是Fable五代吧。

You can only imagine, right, the next model, which is Fable five.

我的意思是，他们已经发布了，但性能会好上不少。

I mean, they already dropped it, but it's gonna be significantly better.

挺有意思的是，他们主打的模型反而被撤回了。

It's interesting because the you know, their headline model got pulled back.

现在他们只好把那些性能较弱、更高效的模型不断优化，结果这些次级模型反而慢慢追上了主模型的水平，毕竟主模型都没法再更新到最佳状态了，这真挺耐人寻味的。

So now they just have all of their, like, all of their their worse models or, you know, their more efficient models, their less powerful models, they're starting to catch up with their main model because they can't even, you know, update their main model to be its best, which is really interesting to me.

从下周二开始，这个模型就会成为所有免费用户和专业用户的默认选择。

So this model is gonna basically become the default for all free and pro users starting on Tuesday.

我觉得这说明，具备代理能力的AI现在已经渗透到各个层级了。

I think that is just showing that agent capable AI is now really important across basically every tier.

这不再是高端用户的专属功能了。

This isn't gonna be a just a premium feature.

我觉得这价格是促销价，输入每百万token两美元，输出十美元，会持续到八月三十一号。

And I do think that it's promotional pricing, the $2 per the inputs and the $10 for the outputs, that is gonna be running through August 31.

Anthropic 经常这么干。

And we see this from Anthropic a lot.

我猜我现在登录账户，他们会说，到七月初，还能享受两倍积分使用额度。

Like, I think if I go on my account right now, they're like, till the you know, at the beginning of July, it's two x credit usage.

说实话，我现在每周都尽量把额度用满，觉得这优惠太划算了，但也不清楚能持续多久。

And, I mean, basically, at this point, I'm just trying to get my max out my usage every single week because I feel like you're getting a really great deal, and I don't know how long that will last.

不过价格会从两美元涨到三美元，还是比 OPUS 4.8、GPT 5.5 和 Gemini 3.1 Pro 便宜。

But that's going to increase from $2 to $3, which is still cheaper than OPUS 4.8 and GPT 5.5 and Gemini 3.1 Pro.

虽然还是便宜，但毕竟要涨了。

So it's still cheaper, but it's yeah.

八月三十一号之前，还是两美元的价。

Until August 31, got the $2 price.

之后就会涨到三美元。

It's gonna then go up to $3.

说不定那时候又有新模型上线了，未来会怎样，谁也说不准。

And, you know, maybe we'll have new models coming out then, so who knows what happens in the future.

但在知识型任务的评测中，Sonnet 5 比 Opus 4.8 略胜一筹，这简直让人想不到。

But on knowledge work benchmarks, Sonnet five is slightly better than Opus 4.8, which is pretty crazy.

对吧？

Right?

Sonnet 本来是比 Opus 低一档的模型，结果因为 Opus 无法升级，反而这个‘低配版’现在比 Opus 4.8 还强。

Like, Sonnet is the, you know, the the lower quality model than Opus, but because they can't upgrade Opus, now the worse version is actually better than Opus 4.8.

我觉得这其实说明，单任务的成本效率现在变得特别关键。

I think that's just kinda showing though the cost per task efficiency is now really important.

它的重要性甚至可能超过了纯粹的推理能力排名。

It's more than more important than maybe the raw reasoning rank.

如果你愿意花大价钱，当然也可以用 Claude Code，开启他们的超强编码模式之类的。

And if you wanna spend a ton of money, you can go use something like Claude code and turn on like, you know, their super code mode or whatever.

那样能让你的模型多花点心思推理，但也会消耗多得多的令牌。

It gives you, you know, waste more reasoning, but it uses a lot more tokens.

很多早期使用者说，在一些多步骤的工作流里，中间卡顿的情况明显少了。

A lot of early testers say that there are fewer mid task stalls in some of the multi step workflows.

Zapier 的这个两步 Salesforce 自动化任务，之前好多模型都试过，但往往做到一半就放弃了，这次却顺利完成。

Zapier's two part Salesforce automation completed where I think a bunch of previous models, they were trying to do that two part Salesforce automation, and it was getting about halfway through and it abandoned it.

所以总的来说，这次的效率明显高多了。

So anyways, it seems like this is gonna be much more efficient.

说到新模型，可不止他们一家。

Speaking of new models, they're not the only one.

谷歌已经向开发者推出了 Nano Banana 2 Lite 和 Gemini Omni Flash。

Google is shipping Nano Banana two Lite and also Gemini Omni Flash to developers.

这是一个图文模型。

It is a text image model.

生成一张一千像素的图片大约需要四秒，费用只要 0.034 美分。

It's gonna generate in about four seconds, and it costs 0.034¢ for a one k image.

Gemini Omni Flash 是个视频生成模型，每秒输出大约 10 美分，说实话，这价格已经不错了。

Gemini Omni Flash, which is a video generation model, is about is about 10¢ per second of output, which honestly is not bad.

我知道，这类工具的开销有时候能高得离谱。

I have I you know, these things can be insanely expensive.

我们之前明明有OpenAI，结果直接取消了Sora项目，因为成本实在太高了。

We literally had OpenAI that canceled Sora and this or is discontinuing it because it was just so expensive.

所以，我们现在其实是有能力做AI视频的。

So, like, we have the capability to do AI video.

就是贵得离谱。

It's just really expensive.

现在Gemini能把每秒输出的成本压到10美分，这已经不错了，特别是如果你要做一小时的视频。

So the fact that Gemini is getting that down to 10¢ per second of output, not bad, you know, if you were doing a a one hour long video.

那费用还是会很高。

It's gonna be pretty expensive.

但这些模型生成的视频大多数也就五秒左右，算下来每条50美分，我知道听起来确实不便宜。

But most of the videos that these are generating are, like, five seconds long, so that's 50¢ a video, which I know actually sounds kind of expensive.

如果是十秒的视频，就是一美元。

If it's ten second video, it's a dollar.

要是百秒的视频，那就是十美元。

If it's a, you know, a hundred second video, it's $10.

但你知道，这就是成本。

But, you know, that's that's the cost.

明白了。

Okay.

Nano Banana Two Lite 将取代原来的 Nano Banana。

Nano Banana two Lite is gonna be replacing the original Nano Banana.

它将成为推荐的版本。

It's gonna be the as the recommended tier.

我觉得他们说的是，如果你以前的提示在原版 Nano Banana 上效果很好，那在 Nano Banana Two Lite 上照样能用，这很棒。

And I think they're saying that it's, like, if you have a prompt that worked really well for the original Nano Banana, it's gonna still work for Nano Banana two Lite, which is great.

最烦的就是模型一改，以前好用的提示全都不灵了。

I hate it when the model changes and the prompts are which kind of get off.

他们说会保持角色的一致性。

They said there's gonna be character consistency.

他们做了非常激进的速度和成本优化。

They have really aggressive speed cost optimization that they've applied.

这还挺有意思的。

So it's kind of interesting.

对吧？

Right?

就是 Nano Banana 2 Lite。

It's like two it's Nano Banana 2 Lite.

这本来该是个超快超轻的版本，但他们还想让它在性能上跟原版 Nano Banana 差不多，只是速度更快、成本更低。

Like, this is supposed to be like a super fast, super light version, and it's they're trying to get this thing to still be comparable to Nano Banana, but way faster and way cheaper.

每当我们对图像模型做这种优化时，我都特别兴奋。

And I'm really excited when we do these kind of optimizations on the image models.

他们还推出了 OmniFlash，定价跟 v o three one 一样实惠，同时还加入了对话式编辑和多模态输入混合功能，支持文字、图片和视频一起处理。

They also have OmniFlash, which is gonna match v o three one, fast pricing while it's also adding conversational editing, multimodal input mixing, so they could do text, image, and video.

它还能把屏幕上的图形内容跟文字操作同步起来，我觉得这简直太棒了。

And it's gonna do text to action synchronization for any sort of like graphics that you have on the screen, which I think is phenomenal.

对视频创作者来说，这真的特别实用。

So for videographers, this is actually really cool.

Nano Banana 2 Lite 将会应用在谷歌现有的九款产品中。

Nano Banana two Lite is gonna ship across nine of the different products that Google already has.

它会集成到 AI Studio 里。

So it's gonna be inside of AI Studio.

也会内置在 Gemini API 中。

It's gonna be inside of the Gemini API.

搜索中的 AI 模式也会用上它。

It's gonna be inside of AI mode in search.

当你在谷歌上搜索时，它会出现在 Gemini 应用、Notebook LM、Google Photos、Stitch、Google Flow，还有 Google 广告里。

So when you're on Google, just searching, it's gonna be on the Gemini app, Notebook LM, Google Photos, Stitch, Google Flow, and then it's also inside of Google Ads.

对吧？

Right?

他们总得靠这些广告赚钱吧。

They gotta gotta make money from the ads off of all of this.

我很期待谷歌接下来会推出什么。

I'm excited to see what Google comes out with.

今天看到这么多新模型，我真的很期待。

I'm excited for all the new models today.

各位，非常感谢你们收听这个播客。

Guys, thank you so much for tuning in to the podcast.

如果你想了解 AI Box 的 MCP，记得去 aibox.ai/mcp 看看。

Make sure if you wanna check out the MCP for AI Box that you go check out aibox.ai/mcp.

你可以在 Claude、Gemini 或 ChatGPT 里直接使用全部 80 个顶尖 AI 模型，看你平时最常用哪个。

Basically, you're gonna get all 80 of the top AI models right inside of Claude or Gemini or ChatGPT, whatever you work with the most.

里面还集成了其他所有模型，随时都能调用。

You can get all of the other models inside of there so you can actually call them.

真的特别实用。

Really, really useful.

和往常一样，如果你想每天收到我在这档播客里提到的最新动态，直接去 aichatdaily.com 订阅。

And as always, if you wanna get all of these new stories that I talk about on the podcast straight into your inbox every day, go to aichatdaily.com.

这是和这个播客配套的网站。

That's my website that goes along with this podcast.

右上角有个明显的订阅按钮，点一下就行。

There's a big subscribe button you can hit in the top corner.

点完之后，留下你的邮箱，我会把所有这些故事和更深入的分析文章发给你，把每个细节都讲清楚。

And if you hit that, you can put in your email, and I'll send you all of these stories and, more in-depth articles where it breaks down all of the information.

这些内容你都能看到。

You can see all of that as well.

非常感谢大家收听这个播客。

Thank you so much for tuning to the podcast, guys.

祝你今天过得特别棒，我们下一期再见。

I hope you have a fantastic day, and I'll catch you in the next episode.