你应该使用哪个AI？Claude vs ChatGPT vs Grok vs Gemini（2026指南）

在过去的九十天里，Frontier Labs发布了超过15款模型。

Over the last ninety days, Frontier Labs shipped 15 plus models.

OpenAI发布了三款，Anthropic发布了四款，Google发布了三款，连Meta也发布了。

OpenAI shipped three, Anthropic shipped four, Google shipped three, even Meta.

还有Elon Musk也发布了前沿模型。

And Elon Musk shipped Frontier models.

中国那边也发布了不少。

The Chinese shipped a bunch as well.

所以如果你在听这个节目，你可能会想，我现在到底该用哪个模型呢？

And so if you're listening to the show, you're probably wondering which model should I be using right now?

你们大多数人可能已经订阅了GPT或Claude，但你们也在想，我是不是该试试这些不同的模型？

Most of you likely have a GPT or Claude subscription, but you're wondering, should I be using these different models?

实际上，AI模型的前沿领域，最初以为只有一两个，现在已经扩展到成百上千个模型了。

The truth is the Frontier landscape of AI models, when it was originally thought to be one or two, has now expanded to hundreds and hundreds of models.

光是在Open Routo上，你就能访问超过400个模型。

In Open Routo alone, you can access 400 plus.

所以在这一集里，我们会深入探讨你应该用哪个模型，针对什么用途，以及什么时候用最合适。

And so on this episode, we're gonna dig into which model you should use, for what particular use, and when makes the most sense.

你会发现，争论的重点已经变了，不再是选最聪明的模型，而是选更便宜的，或者最适合你的那个。

And you'll realize that the argument has shifted not from using the most intelligent model, but maybe using the most cheaper model or using the model that's specific for you.

你有得选，伙计。

You got options, baby.

AI领域现在热闹得很，我们会帮你理清这个领域，因为选择太多了，确实容易让人头大。

There's a lot going on in the AI space, and we're gonna help navigate that space because there is a lot of options and it does get overwhelming.

我觉得咱俩，Ejaaz，基本上每个模型都碰过，至少试过几个。

I think between the two of us, Ejaaz, we've probably touched every single one of these, at least a couple of

我推荐太多了，Josh。

I have too many suggestions, Josh.

所以，对，我们那75个订阅加起来，这些模型都覆盖了。

So, yeah, between our 75 different subscriptions, we have covered these.

我们有一些建议，关于哪个模型在什么情况下用，哪个最适合哪些场景。

We have some feedback about which one to use for when, which is best for which use cases.

我想先说说，我们这里有一个挺有用的可视化辅助材料，可以帮我们梳理一下我们的想法，以及我们是怎么考虑这个问题的。

And I guess to start, we have this, like, pretty helpful visual companion artifact here that can walk through kind of what we're thinking, how we think about this.

首先就是分类。

The first is the split.

在考虑用哪个模型的时候，其实可以分成两大类。

There are two distinct classes of model when it comes to considering which ones to use.

第一类是开源的。

The first is open source.

这些主要是中国的模型。

These are Chinese models predominantly.

大部分都是像Kimmies、DeepSeek这样的。

Most of these are your Kimmies, your DeepSeqs.

这些模型都是中国的开源模型。

Like, all of those models are Chinese open models.

然后就是美国的模型，它们基本都是闭源的。

And then we have the actual US models that are all closed source.

这两者之间差别很大，我们往下滚动一点就能看到使用上的差异，因为2025年6月时，美国实验室占了生成token的70%，现在降到了30%。

There's a big difference between the two, and we could see if we scroll down a little bit the difference in usage between these two because in June 2025, US Labs accounted for 70% of the tokens generated, and now they're down to 30%.

经济格局已经发生了巨大变化。

The economics have changed widely.

所以现在那30%的价值远超之前的70%，但值得注意的是，过去几个月里，中国模型确实有了一种复兴的势头。

So now that 30% is worth much more than 70%, but it is interesting to note that there has been this kind of reinvigoration of Chinese models over the past couple of months.

当然，这是基于Open Router的数据。

Now granted, this is based on Open Router data.

Open Router只是一个单一的模型路由实例。

Open Router is a single model router instance.

这并不能代表整体情况，但值得指出的是，一些开源模型在处理当前任务上确实相当强大。

This is not reflective of the norm, but it's just worth noting that it's like some of these open source models are pretty powerful for the task at hand.

对。

Yeah.

我觉得没人能否认，大家确实大量转向了这些开源模型，而且原因多种多样。

I think no one can debate the fact that people have shifted a lot using these open source models, and and it's for a variety of different reasons.

人们想拥有自己的数据。

People want to own their own data.

他们想在家里私下运行。

They wanna run it privately at home.

但最大的转变，最主要的原因是因为这些模型便宜得多。

But the biggest shift, the biggest reason has been because these models are a lot cheaper.

你提到的Open Router数据其实只代表了很小一部分软件工程师，他们想尝试各种模型。

And open Rated data you mentioned only really represents a very niche sector of software engineers that want to experiment with a bunch of these models.

但即使在企业领域，OpenAI和Anthropic在模型份额上占主导地位，它们也在失去token市场份额，因为企业正在测试不同模型以节省数千万甚至数亿美元。

But even in the enterprise world where OpenAI and Anthropic are pretty dominant with their own model share, they've been losing the token market share to enterprises that are trialing and testing different models to save tens to hundreds of millions of dollars as well.

当我看到主要原因时，不是因为开源模型最聪明。

And when I look at the main reason why, it's not because the open models are the most intelligent.

你马上就会看到，在评分卡上，中国的GLM 5.3是个很棒的模型，但不如Fable 5。

You'll see in a second, as we go through the scorecard, GLM 5.3 from China, amazing model, not as good as Fable five.

你看Kimi K2.7也会有同样的感觉，但这些模型已经足够好，能完成你大部分的工作。

You'll look at Kimi K 2.7, you'll think the same thing, but those models are good enough to do the bulk of your own work.

听起来不太像AI生成的那种内容。

It doesn't sound too much like AI slop.

它实际上就是正常地跟你说话。

It actually just speaks to you normally.

最大的优势是它的限制更少，这有利有弊，但基本上你让它做什么它就做什么。

And the biggest advantage is it has fewer safeguards, which has its pros and cons, but it basically does the task that you ask it to.

现在，总的来说，在节目里聊一堆模型没问题，但最好还是先大致了解一下这些模型的智能程度或实际效果。

Now, overall, it's fine if we talk about a bunch of models on this show, but it's good to get an overview essentially of how intelligent or how effective these models are.

我们屏幕上显示的是所谓的“人工分析指数”或“智能指数”。

And what we have on the screen here is something known as the artificial analysis index or intelligence index.

这基本上是测试这些不同模型通用智能的最佳基准。

And this is basically the best benchmark to test the general intelligence of these different models.

你可能不会感到意外，Claude模型，也就是Anthropic的模型，在这项排名中位居榜首。

Now it may come as no surprise to you, but the Claude models, Anthropic models top this.

最上面的是Opus 5，这是他们最近发布的模型。

We've got Opus five at the top, which is their most recent model launch.

还有Claude Fable 5。

We've got Claude Fable five.

然后让我惊讶的是，Josh，Elon的Grok 4.6简直——又回来了，宝贝。

And then surprising to me, Josh, but Grok 4.6 from Elon is just- back baby.

太棒了。

It is fantastic.

真的。

Come on.

SpaceX是怎么做到从在模型竞赛中远远落后，到现在这么厉害的？

How effectively SpaceX has been able to pivot from being way behind the model race.

他们解雇了整个AI团队，招了一批新人，花600亿美元收购了Cursor，然后在过去六周左右就推出了一款能和顶尖模型抗衡的模型。

They fired their entire AI team, hired a bunch of new folk, acquired Cursor for $60,000,000,000 and put out a model in, I think it's like the last six weeks that was able to contend with the top.

最棒的是，他拥有训练这些模型所需的全部GPU和计算资源。

And the best part is he has all the GPUs and compute to train them.

再往下看，你会看到OpenAI的GPT 5.6 Sol，以及一堆其他中国的开源模型。

And then if you go further down the stack, you'll see the likes of OpenAI, GPT 5.6 Sol, and a bunch of other Chinese open source models.

问题是，你怎么看这个？

The question is like, what do you make of this?

作为一个只是用AI的人，你怎么看这个？

Like as someone who is just a user of AI, what do you make of this?

你怎么去应对，嗯，

How do you kind of navigate Well,

让我试试这个，Josh。

let me try this, Josh.

让我问你一个问题。

Let me ask you this question.

对吧？

Right?

屏幕上这些模型，你用的是哪些，具体用来做什么？

Of these models that you see on the screen here, which ones are you using and for what specifically?

有意思的是，大概90%的使用量都来自前两个，剩下的10%来自其他所有模型。

It's funny because about 90% of the usage comes from the top two, and then 10% comes from everything else.

这大概就是我想表达的意思。

And this is kind of what I was going for.

我其实挺好奇的，你那边是不是也差不多？

I'm actually curious, is it a similar thing for you?

你主要用哪些模型？是的。

What kind of models are Yes.

你

You

不过对我来说，大概是75%和25%吧。

But okay, so I would say it's around, I'm gonna say like 75, 25.

我之所以有这个微调，是因为我倾向于用最方便的AI模型，不管是从哪里得到信息或者想到要用AI模型。

And the reason why I have that slight adjustment is because I kind of use the AI model that's most convenient to me wherever I'm getting the information or the inclination to use an AI model.

比如X就是个很好的例子，对吧？

So X is a perfect example, right?

我在刷X的时候，看到某个AI大牛发的东西，我就想，这到底在说啥啊。

I'm scrolling X, I read something from one of these genius AI researchers and I'm like, I have no idea what the hell this means.

我就点一下那个“问Grok”的按钮。

I tap the ask Grok button.

我在用谷歌。

I'm on Google.

我在搜东西，对吧？

I'm searching something, right?

我在用笔记本电脑。

I'm on my laptop.

我就想，好吧，那我就跟Gemini说说这个。

I'm like, okay, I'll just speak to Gemini in this.

所以从这个角度来说，哪里方便用AI，我就在哪里用。

So in that sense, wherever the AI is conveniently placed, I use it there.

而且我觉得，等苹果过几个月或者可能就几周后发布他们的新AI模型，我可能也会在手机上用它，因为那些模型基本上都能很好地回答我的基本问题。

And I have a feeling when Apple releases like their new AI model in a couple of months or maybe a couple of weeks actually, I'll probably use it on my phone as well because the models just generally are good enough to answer my basic questions.

对。

Mhmm.

对。

Yeah.

好的。

Okay.

所以我们看这些图表，看到的是，好的，这里是智能所在。

So we're looking at these charts and we're seeing, okay, here's where the intelligence is.

这里智能稍微弱一些，但便宜得多。

Here's where things get a little less intelligent, but much cheaper.

如果你是普通消费者，我觉得成本不是什么大问题，因为成本差异主要体现在规模上。

And if you're a normal consumer, I have a feeling that cost isn't that big of a deal because the cost differences are really at scale.

很多公司提供的，比如Anthropic、OpenAI、Grok和Gemini，都是每月20美元、100美元或200美元的套餐。

What a lot of these companies offer, what Anthropic offers, what OpenAI offers, what Grok and Gemini offer is just the $20 month, $100 month, $200 month plan.

你只需支付固定费用，就能获得这些模型的大量使用权限。

You could just pay a fixed amount and get access to a lot of usage of these models.

所以如果你是一般消费者，能每月花20美元订阅，那么去那些主要的AI实验室可能更适合你。

So if you are a general consumer who is able to spend $20 per month on a subscription, you're probably better served going to one of the major IO labs.

产品更好。

The product is better.

你会得到更多的智能。

You're gonna get more intelligence.

你会得到一个更连贯的产品。

You're going to get a more coherent product.

我发现，根据我的经验，很多和我交谈的人，使用这些开源中国模型的唯一真正原因，是你需要消耗极其大量的令牌，比如用于代理任务。

I find that a lot of the people who I speak to, lot of the time in my experience, the only real reason to use these open source Chinese models is if you're consuming tremendously large amounts of tokens for, say, agentic tasks.

比如，你在运行一个Open Claw实例，或者任何带有大量自动化工作流、大量代理的Claw，你可能会想用这些更便宜的，也许是中国的开源模型，来节省成本。

Like, you're running an Open Claw instance or any sort of Claw with a lot of automated workflows, a lot of agents, you may want to defer to some of these cheaper, perhaps the Chinese open source models just to save on costs on the low end.

说到日常使用的高端场景，我觉得看这个视频的人中，90%以上最好还是订阅你最喜欢的AI公司，用他们的产品。

When it comes to day to day use on the high end, I'd say probably 90 plus percent of the people watching this are best served just getting a subscription to your favorite AI company and using that product.

我就是这么用的。

That's kind of how I use it.

有趣的是，剩下的10%的提示词不是通过Anthropic完成的，而是主要通过Grok完成的，就像你提到的，当我在x上全天候在线的时候。

It's funny that the remaining 10% of prompts that aren't done through Anthropic are done mostly through Grok, like you mentioned, when I'm on x twenty four hours a day.

它是个好搭档，因为它能访问

It's a good companion because it has access

那个数据库。

to that database.

然后

And then

另一个其实是ChatGPT，因为他们的图像生成模型真的很棒，我喜欢用它来制作表情包或者给朋友发图片，就把它当作视觉工具。

the other one is actually ChatGPT because they have a really amazing image generation model, and I love creating memes or sending images to my friends who are just, like, using that as the visual component.

我觉得这特别有意思。

I find that really compelling.

所以我的组合大概就是，主要用Anthropic，有时候用Grok，有时候用ChatGPT，特别是用来生成图片的时候。

So that's kind of my stack is, like, mostly Anthropic, sometimes Grok, sometimes ChatGPT, particularly for the ImageGen.

我觉得说到图像生成，我们以前一直用Nano Banana和Gemini。

I think when it comes to ImageGen, we used to use Nano Banana and Gemini all the time.

那时候它们可是最厉害的。

That was, like, the top dog.

我觉得自从ChatGPT发布了2.0版的图像生成功能，确实非常吸引人。

I think ever since ImageGen two point o release from ChatGPT, it is very compelling.

我喜欢的是它背后真的有思考过程。

What I like is that it does get actual thinking behind it.

以前你让它生成图片，它就按字面意思理解你的话，直接生成。

So previously, when you asked them to generate an image, it would take your words at face value, generate the image.

现在它会做一些推理，做一些思考，理解你的意思，也理解图片里的内容，然后生成更好的结果。

Now it it actually does some inference, it does some thinking, it kind of interprets your message and interprets what it's seeing in the image, and then creates a much better output.

所以我觉得这就是我处理这个问题的思路。

So that's kind of how I would think of navigating this.

比如，中国模型确实很厉害，但除非你是那种高级重度用户，消耗大量token，不然可能用处不大。

Like, Chinese models are are certainly a big deal, but unless you are kind of advanced super user consuming a ton of tokens, perhaps they're not that valuable.

我来说说我的个人经验吧。

I will speak from my own personal experience.

我基本上一直用Claude模型，但我现在用Opus 5比Fable 5更多了。

I use the Claude models pretty much all the time, but I've started using Opus five more than Fable five.

这听起来可能有点傻，但原因就是它跟我说话像个正常人。

And this might sound like a silly reason, but it's because it speaks to me like a normal human being.

我不知道Opus五和Fable五在个性设置上有什么不同，但他们得赶紧给Fable五也弄成这样，因为Fable五老是绕来绕去，说话还文绉绉的。

I don't know what the personality dials are on Opus five versus Fable five, but they need to do that to Fable five pronto because Fable five just waffles and speaks in like archaic pros.

我就想说，听着，我只要我问的信息，别的都不用给我。

And I'm just like, listen, just give me the information that I've asked for and do not give me anything else.

我可不想每次用都得读一篇作文，对吧？

I don't want to have to read an essay every time I use this, right?

现在说说Grok 4.6，Josh，我确实注意到它很聪明。

Now Grok 4.6, I've actually noticed the intelligently, Josh.

我不知道你有没有在X上试过，可能你在研究趋势的时候会用到。

I don't know if you've tried this on X or you probably do when you're like kind of like trying to figure out what trends.

它跟你说话特别聪明，而且没那么粗俗了。

It speaks to you super intelligently and there's less crass about it.

如果你记得早期版本的Grok，它总是会加一些奇怪的词、俚语和短语，现在它就像个聪明人一样跟你说话了。

If you remember the earlier versions of Grok used to kind of like throw in some kind of like questionable words, slang, and phrases, and now it just speaks to you like an intelligent human being.

我觉得这是因为马斯克，你知道，SpaceX已经上市了。

And I think this is because Elon's like, you know, SpaceX has IPO ed.

他有了Cursor，还有非常专业的研究人员参与，他们正试图瞄准更偏向企业级的受众。

He's got Cursor, very professional researchers involved, and they're trying to target a more enterprise oriented audience.

他之前说过，Grok 4.7将在几周内发布。

And he said before that Grok 4.7 is on its way to being released in a few weeks now.

他的目标是年底前推出Grok 5。

He's aiming for Grok five by the end of the year.

那其实也就只有几个月的时间了。

That's literally only in a couple of months.

他确实在瞄准企业市场，因为他看到Anthropic和OpenAI都在往这个方向走。

He is really targeting the enterprise landscape because he's seen Anthropic and OpenAI go after it.

现在，跟你一样，对于中国的模型，我一个都没用。

Now, same as you on the Chinese models, I am not using any of them.

我承认这可能是我自己的偏见，因为主要是我觉得，我不确定，我是不是要费劲去注册账号，用那些模型，然后还得想，嗯，我不知道这给我的信息是不是对的。

And I admit this might be my own bias because mainly I'm like, I'm not sure, like I want to go to the efforts of signing up on an account using those models and then being like, well, I don't know whether this has given me the right information.

我不知道。

I don't know.

我对美国品牌有点奇怪的信任问题，可能这是我自己的问题。

I kind of have this weird trust thing with The US and American brands and maybe that's like my own fault.

然后我想指出的是，所有这些不同的模型，我们在屏幕上展示的，我觉得大部分公司，特别是Anthropic和OpenAI，都专注于让他们的模型对企业用户非常友好。

And then the one thing that I'll point out across all these different models that we make the point on, on the screen here is I think a bulk of the companies Anthropic and OpenAI especially, have focused on making their models really good for enterprises.

但就像我之前提到的，真正的转变已经转向了开源，因为更多可用的模型更便宜，更容易让普通人使用。

But as I mentioned earlier, the shift has really happened to open source because more available models are cheaper and more accessible to people to use.

他们可以私有化部署。

They can run privately.

但Anthropic和OpenAI对此的回应是，我们只需要蒸馏我们的主模型，然后给你更便宜的模型。

But Anthropic and OpenAI's response to this is we'll just distill our main model and give you cheaper models.

所以这就是OpenAI所做的，对吧？

So that's what OpenAI has done, right?

你有Jupey 5.6，有三个版本。

You've got Jupey 5.6, there's three versions of it.

有Sol、Terra和Luna三个版本。

You've got Sol, Terra, and Luna.

Luna的价格，我觉得，比一周前已经降了80%，对吧？

Luna is, I think, slashed 80% of the price than it was before already a week ago, right?

所以现在它跟GLM 5.3竞争很激烈，Anthropic也在做同样的事。

So now it's like competing very much with GLM five point three and you've got Anthropic doing the same.

我觉得Opus 5的成本大概是Fable 5的一半，但能力跟Fable 5一样强。

I think Opus five is like half the cost of Fable five whilst being as capable Fable five.

而且他们SANA的定价也保持得很低，是最低的。没错。

And they also kept the SANA pricing low too, which is the lowest Exactly.

就是说，对，有很多

That it's like, yeah, there's plenty of

选择。

options there.

选择很多。

Plenty of options.

我觉得这归根结底就一件事，我不想把话题扯到算力上，但我不得不提，Josh。

And I think this comes down to one thing, and I don't wanna make this about compute, I have to, Josh.

我觉得不管是哪家实验室，不管你是中国的，还是美国那种低端实验室，或者是高端实验室，算力都会决定你能不能把这些模型以更便宜的价格提供出去，而这基本上会决定哪些客户会使用你的产品。

I think whichever lab, whether you're Chinese, whether you're like a low tier US American lab, or whether you're the high tier labs, Compute will dictate whether you can serve these bottles cheaper, which will basically decide or determine which customers use your product.

对。

Yeah.

而且我觉得，这最终还是要看你的使用场景。

And I mean, this is ultimately coming down to your use case.

比如，很多人根本不会遇到这个问题，因为他们永远不需要用到所谓的API密钥。

Like, a lot of people won't ever run into this problem because they will never need this thing called an API key.

他们永远不会按token付费。

They'll never actually pay per token.

他们通常只是通过订阅来付费。

They'll oftentimes just pay through the subscription.

如果你用的是订阅，那你有几个选择。

If you're using a subscription, you have a couple of options.

也许现在是时候来看看这些订阅服务，以及各家都提供什么了。

Maybe this is a good time to go through these subscriptions and what are the offerings of everybody.

我们有Gemini，也许我们可以从它开始

We have Gemini, which perhaps we could start

因为我觉得我们最近对Gemini有点苛刻。

there because I feel like we've been mean

我们最近对Gemini有点苛刻。

to Gemini recently.

我们还没怎么聊过

We haven't talked so

那关于

much What's about

那个什么？

that what?

就是那个。

That like, that one.

那是谁？

Who's that?

他们现在用的是Gemini 3.1 Pro，说实话，已经远远落后于前沿水平了。

They're currently at Gemini 3.1 Pro, which is nowhere really near the Frontier anymore, unfortunately.

他们确实有Nano Banana Pro，这个图像生成功能挺好玩、挺酷的。

They do have Nano Banana Pro, which is like this fun, cool image generation process.

我觉得Google有意思的地方在于，如果你对这些工具感兴趣，订阅可能会很值。

They have what I find interesting about Google, and if you are interested in these tools, it may be interesting for a subscription.

Google有一些比较冷门的工具，但它们的应用方式特别有意思。

Google has these, like, weird edge case tools that have really interesting harnesses.

其中一个就是音乐制作方面的。

One of them is for music production.

我记得我们在节目里用过很多次。

I remember we've used this on the show many times.

它在生成歌词和音乐方面真的很厉害，听起来效果特别好。

It's really good at generating lyrics and generating music in a way that sounds very good.

而且这种体验我只在谷歌的产品上遇到过。

And I've only had that experience on a Google product.

我在其他地方都没法用上这个功能。

I haven't been able to use that anywhere

对。

else.

Lyria。

Lyria.

Lyria。

Lyria.

没错。

That's right.

是的。

Yes.

有Lyria。

Have Lyria.

这个项目是为节目制作了开场曲。

Project for created the jingle for the for the show.

太棒了，兄弟。

So good, dude.

所以，像Gemini和谷歌，就适合处理那些奇怪的东西。

So, like, Gemini and Google is, like, good for the weird stuff.

比如，它能给你生成相当不错的图片。

Like, it'll generate you pretty good images.

它能给你生成有趣的音乐。

It'll generate you fun music.

他们还有另一个工具。

They have this other tool.

这个在谷歌实验室套件里，能帮你生成营销材料。

This is in, like, the Google Labs Suite where it'll generate you marketing material.

所以如果你有很多品牌素材，或者有一堆标志，想要营销材料，比如把你的标志印在布置得很好的东西上，或者想要定制商品，它都能帮你搞定很多。

So if you have a lot of material from your brand or you have a bunch of logos and you want marketing material, the kind of say your logo's printed on something that's staged nicely or you want custom merchandise, it'll do a lot of that for you.

所以，对于实验室产品那些特定的使用场景来说，还是挺好玩的。

So it's fun for those kind of narrow use cases of where the labs products lie.

除此之外，我就不太清楚了。

Outside of that, I don't know.

我不太推荐它。

I can't really recommend it that much.

它不是最强的，离最强还差得远。

It's not the strongest, not close to the strongest.

我觉得，除此之外，还有Grok，它正在快速追赶上来。

I'd say probably on top of that, we have Grok, which is very quickly catching up.

如果你像我们一样花很多时间在X上，Grok就是你的好朋友。

If you spend a ton of time on X like we do, Grok is your best friend.

Grok能访问X上实时可用的数据。

Grok has access to that data that is available in real time on X.

如果你想了解几乎任何事情的实时新闻，Grok就是你的首选。

If you're looking for current up to date news on pretty much anything, Grok is your go to.

它能实时整理并引用特定时刻的推文。

It can sort and cite specific tweets from specific moments that are happening in real time.

而且总的来说，就像你提到的，Ejaaz，它非常擅长直接和你交流。

And just generally speaking, like you mentioned, Ejaaz, it's very good at just communicating with you directly.

没有太多废话。

There's not a lot of fluff.

它是一个科学和技术模型。

It is a science and technical model.

它直截了当。

It is direct to the point.

如果你愿意，它还会跟你爆粗口。

It'll get vulgar with you if you want.

跟它玩很有意思。

It's fun to play around with.

还有语音模式，真的很有趣，我觉得在所有语音模式里，那个是我玩得最开心的，因为它实在太搞笑了。

There is the voice mode, which is really fun that I've I'd say of all the voice modes, that's the voice mode I've had the most fun with is because it's just so ridiculous.

而且很多时候，你在网上看到的那些梗图，背后就是Grok在配音。

And oftentimes, if you see memes online, it is because Grok is the voice behind it.

GPT的语音模式真的让我很不适应。

The GPT voice mode just jars me so much.

我也不知道为什么。

I don't I don't know.

你有没有跟它聊过，就是最近那个……

Have you have you spoken to it with the recent

聊过。

I have.

对。

Yes.

你这么说，具体是什么意思呢？

So when you say that, what do you mean?

就是它听起来更像真人，对吧？

And and so it sounds more human, right?

它会停顿，但听起来就像某个很拽的人在说话。

It like pauses, but it sounds like it's like some sassy person that's talking

跟我们说话。

to us.

比如我问了个问题，它就说，对，对。

Like I asked this question and he goes, yeah, yeah.

然后说，对，听着，我懂。

Like, yeah, like, listen, like I get it.

然后说，你知道，今天天气就是这样。

And you know, this is what the weather is.

我就想，喂，我不需要这个。

And I'm like, yo, I don't need this.

我只是问你天气怎么样。

I just asked you what the weather is.

直接告诉我就行了。

Like, just tell me what's up.

不过大体上我同意你对谷歌和SpaceX的看法。

But I agree with you largely on like Google and SpaceX.

乔什，关于谷歌，我想说的是，我觉得他们和微软的情况很像。

What I will say about Google, Josh, is I think they have a similar profile to Microsoft.

我很好奇你是否同意我的看法。

I'm interested if you agree with me here.

微软并不是我会拿来和顶级前沿AI模型实验室相提并论的名字。

Microsoft is not a name that I would draw against the top Frontier AI model labs.

事实上，我觉得他们错失了大好机会，但他们几乎渗透进了每一家财富500强公司，不管我们喜不喜欢。

In fact, I think they fumbled the bag massively, but they are embedded across pretty much every single Fortune 500 company, whether we like it or not.

我们只是生活在自己的科技圈子里，但在科技圈之外，人们都在用微软Teams和微软办公套件。

Like we just live in our tech bubble, but outside of the tech bubble, people run Microsoft Teams and the Microsoft Suite.

我觉得谷歌基本上也处于类似的位置。

And I think Google's in a similar position essentially.

很多新兴公司都用谷歌文档、谷歌套件，他们也会像我一样直接点Gemini按钮，因为方便嘛。

Like a lot of new upstarts use Google Docs, Google Suite, and they're just gonna click the Gemini button in the same way that I do so because it's just convenient.

你同意这个观点吗？

Do you do you agree with that?

对。

Yeah.

我觉得作为全球最大的公司，能接触到所有这些传统企业客户，这本身就是一种护城河。

I guess there's just a moat to being like the biggest company in the world and having access to all of these kinda, like, legacy enterprise companies.

感觉过去二三十年里，凡是创业做生意的人，都在用微软，也在用谷歌。

It feels like anyone who has built a business over the last twenty or thirty years has been using Microsoft, has been using Google.

他们会继续利用这个优势。

They're going to continue to leverage that.

如果让我猜的话，谷歌似乎比微软稍微领先一点。

It seems like Google is slightly more ahead than Microsoft, if I had to guess.

比如，我对谷歌这家公司比对微软更感兴趣，但他们都受益于那些老客户群。

Like, I'm I'm much more excited about Google as a company than Microsoft is, but they both benefit from that legacy customer base.

嗯，两家都是很棒的VC。

Well, both are amazing VCs.

对吧？

Right?

两家都是很厉害的风投。

Both are incredible VCs.

是的。

Yes.

我觉得评判它们不应该看产品好不好，而要看它们对那些将来会打败它们的竞争对手的投资怎么样。

I think they should be judged not on the quality of their product, but the quality of their investment in the other competitors that are going to crush them.

那倒是，但是……

That is the so but

我喜欢这个观点。

I do like that.

对吧？

Right?

就像它们有对冲一样。

It's like they have a hedge.

微软其实持有OpenAI很大一部分股份。

Microsoft owns, like, a large part of OpenAI.

这太厉害了。

That's incredible.

谷歌在SpaceX和Anthropic都持有大量股份，这两家公司将来都会是巨大的IPO，其中一家已经上市了。

Google owns this, like, massive stake in SpaceX, in Anthropic, in both of these companies that are going to, like, be huge IPOs, one of which already did.

所以它们是很棒的风险投资基金，但作为AI实验室可能稍微差一点。

So they are amazing venture funds, perhaps slightly less a better AI labs.

我们可能会推荐的那两家，恰恰是不用依赖这种传统软件优势的。

The two that don't have to benefit from this kind of legacy software are the two that we're probably going to recommend.

实际上，如果你再往下滚动一点，就能看到价格和智能度的对比图，以及每个模型大概在什么位置。

And in fact, if you scroll down a little bit further, we could see the box of price against intelligence and kind of where each of these models sit.

对。

Yeah.

就是这张图，我们看的是右上角那些。

This chart right here, and we're looking at those in the top right.

这是大多数人会想用的地方。

This is where most of the people are gonna wanna live.

我觉得我们大部分时间都会花在这里。

This is where I think we spend all of our time.

你要么用ChatGPT，要么用Claude会员，基本上这就是大多数人需要的。

You're either getting a ChatGPT or you're getting a Claude membership, and like that's that's for the most part, that's what most people need.

差别不大，但值得注意，主要是模型方面的差异。

The differences are small, but noteworthy, mostly as it relates to the models.

我觉得，这些里面，我强烈推荐。

I mean, of them, I can't recommend enough.

就花那20美元一个月。

Like, spend the $20 a month.

试试看。

Try it.

就试一个月。

Just try it for a month.

如果你还没有用过，试试看感觉怎么样。

If you don't have it, see what you think.

如果额度用完了，可以升级到每月100美元的套餐。

If you run out of tokens, upgrade to a $100 a month.

这能让你接触到那些研究人员用来取得数学突破的同样智能，真的非常厉害，能力也很强。

This is access to the same intelligence that all these researchers are making math breakthroughs with, and it's really impressive and really capable.

不过，每个模型都略有不同。

Now each one of these is slightly different.

我觉得Claude更偏向编程和写作，而且作为一个通用模型，它也非常强大。

I'd say Claude is a little more tailored towards coding, writing, and it's a really strong general purpose model.

我发现GPT 5.6，特别是Soul，特别擅长做高层级的操作。

I find with GPT 5.6, Soul in particular, it's very good at being the high level operator.

所以如果你在处理复杂的任务，它很擅长制定规格说明。

So if you're working on, like, complicated tasks, it's good at creating spec sheets.

它也很擅长检查你做的事情。

It's good at running checks against the things that you do.

而且很多时候，我发现Claude和那些传说中的模型，还有Opus模型，在具体执行方面特别强，而且它们不太需要太多指导。

And then oftentimes, like, I find that that Claude and the fabled models and the Opus models are very good at implementation, and they're very good with less guidance.

所以我最近用这些模型的方式，就是越来越少地给它们指导。

So the way I've been using these models recently is with less and less guidance.

我觉得早前，我有一整套技能表，大概有15种不同的技能，用来帮我自动化处理各种事情。

I think earlier on, I had this whole skill sheet, and it was like 15 different skills that would automate different things for me.

但慢慢地，那个技能表变得越来越短，越来越短。

And slowly over time, that skill sheet has gotten lower and lower and lower.

其实我们在节目开始前也聊过，要从这些模型里榨出最大价值，最好的办法就是别挡道，直接告诉它最终目标，然后说，嘿。

And we were talking actually before the show, the best way to extract the most value out of these models is just to get out of the way, give it the end goal, and say like, hey.

帮我去做这件事。

Go do this thing for me.

相信你有这个能力去完成。

Trust that you have the intelligence to do this.

我不会给它设什么条条框框。

I'm not gonna put guardrails on it.

你有上下文。

You have the context.

你可以帮我去做这件事。

You can go and do this for me.

我觉得这主要是我个人使用的方式，我用的是Fable。

And I think that's mostly where I find myself using this personally, is I use Fable.

我绝不会浪费任何一个Fable的token。

I I do not leave any Fable tokens unused.

每周的token都非常宝贵。

Very valuable tokens each week.

然后Opus作为通用主力模型的备选。

And then Opus is the fallback for the general workhorse model.

这个组合真的很强大，基本上能完成你想做的任何事情。

And it's been a really powerful combo of just kind of being able to accomplish anything that you want.

它连接了我所有的服务。

It connects to all of my services.

我把它连到了我的邮箱和谷歌云端硬盘上。

I have it connected to my email, to my Google Drive.

它可以访问我的文件、文件夹和联系人，真是一个全能助手。

It it has access to my files, the folders, the contacts, and it's just a really helpful all in assistant.

它用起来非常顺手。

It works really well.

我想大家可能也在想，好吧。

I think people are also probably wondering, okay.

如果我不太在意通用智能呢？

Well, what if I don't really care about the general intelligence as much?

如果我想做某件具体的事情呢？

What if I'm trying to do a specific thing?

那我该用哪个模型？

Which model should I use then?

我觉得这里有两件事需要考虑。

I think there are two things to consider here.

第一点是，如果你是软件工程师，AI模型在编程能力上一直很火，但可能你们很多人其实在公司里是重度代码使用者，想知道该用哪个模型。

Number one is if you are a software engineer, like AI models have been all the rage for their coding capabilities, but maybe a bunch of you actually like code hardcore at your companies and you wanna figure out which model you should use.

我觉得这个问题上，Claude模型和GPT 5.6模型都挺靠前的。

I would say in that question, the Claude models and the GPT 5.6 models are pretty high up there.

Codec的用户量在一个半月内从500万涨到了1500万。

Codec's usage has gone from 5,000,000 users to 15,000,000 users in about a month and a half.

我每次看到OpenAI的Codex负责人Thibault在时间线上发用户量破百万的更新都烦了，但这些模型不仅擅长理解和阅读你的代码库，还能直觉地判断你接下来该开发什么产品或功能。

I am tired of seeing Thibault, who is the head of Codex at Open AI's tweets on my timeline every single time he gets a million user update, but these models are very powerful at not only understanding and reading your code base, but intuitively figuring out what product or feature you should build next.

我有很多朋友在科技相关甚至完全不是科技行业的公司工作，他们用这些模型来构建他们的高级功能，反馈很多。

I have a ton of feedback from friends that work at companies that are tech adjacent and maybe not even in tech at all, which use these models to build their premium features.

另外我想说的是，我们这里讨论的右上角那个框，基本上就是帕累托前沿。

The other thing I'll say is, you know, we're talking about the top right box over here, which is essentially the Pareto Frontier.

所以如果你必须在成本、智能度或其他方面做出牺牲，你仍然会使用这些类型的模型。

So if you had to sacrifice something of cost or intelligence or blah, blah, you would still be using these types of models.

但如果让我下注的话，我会说Grok 4.6、Quinn 3.8和Kimmy在几个月内会进入这个框。

But if I had to take a bet, if I was a betting man, I would say Grok four point six, Quinn 3.8, and Kimmy will be inside this box within a couple of months time.

这会改变我们使用这些不同类型模型的方式。

That's going to change the way we use these different types of models.

我在节目开始前就说过，虽然这么说不太吸引人，但我认为未来十二个月里，最顶尖的AI公司会是某种聚合平台。

I said before we started this show, it's unsexy to say, but I think the number one AI company that will come out in the next twelve months will be some form of aggregator platform.

我们看到Stripe据说以七十亿美元收购了Open Router。

We saw that Stripe just acquired Open Router for $7,000,000,000 allegedly.

我们昨天就这个话题发了一期节目。

We put out an episode of this yesterday.

大家一定要去看看。

Definitely go check that out.

但我觉得我们会看到这些平台帮你挑选合适的模型，在合适的时间使用。

But I think we're gonna see these platforms that help you pick and choose which models to use at the right time.

它会整合你的记忆，所以它已经知道你想做什么。

It aggregates your memory, so it already knows what you wanna do.

这样你就不会有任何相关的麻烦。

And so you don't have any of the complications around that.

我觉得我们会看到很多这样的模型在这方面更进一步。

I think we'll see a bunch of these models kind of step up that.

现在，除了编程之外，如果你不是程序员，而是一个普通的知识工作者，你上班用邮件、用Slack，还有你提到的那些插件，有些模型特别擅长代理工具的使用。

Now, aside from coding, if you aren't a coder, but if you are, let's say a general knowledge worker, you go to work, you use email, use Slack, you use a bunch of these other plugins that you also mentioned, there are models that are specifically good for AGENTIC tool use.

其实我最近最喜欢的是Grok 4.6，或者更具体说是Grok Bot。

And actually my most recent favorite is Grok 4.6 or Grok Bot specifically.

这是埃隆·马斯克和SpaceX上周发布的，我记得。

This was released from Elon Musk and SpaceX, I think last week.

它基本上就是一个代理，或者多个代理，在云端的虚拟机里运行。

And it basically is an agent or multiple agents that spins up in a virtual machine like in the cloud.

你不用担心它在你的笔记本上运行，然后黑掉你的所有东西。

You don't have to worry about it running on your laptop and hacking all your stuff.

你可以让它访问任何工具，它能直观地理解你想做什么。

And you can give it access to any and every tool and intuitively understands and gets what you want to do.

它会随着时间学习你的工作方式，然后不断改进。

It could learn what you do over time and it improves.

所以我觉得当你看到MetaMuse Spark 1.2的时候，你可能会想，哦，那是Meta的模型。

So I think when you look at the MetaMuse Spark 1.2 and you might think, oh, that's a Meta model.

我从来不用Meta。

I don't ever use Meta.

这类模型会越来越普及，用于日常使用或知识工作。

These types of models are gonna become more available for general usage or knowledge work.

而且我觉得如果你倾向于使用这类模型，又不太在意谷歌搜索那些功能，反正你会用Claude来处理那些。

And I think if that's something that you're inclined to use and you don't really care about the Google search stuff, you'll use Claude for that anyway.

这些模型我可能会去了解一下，因为它们在很多方面都很优秀。

These are models that I'll probably look into because they are superior in many ways.

对。

Yeah.

选一个然后坚持用下去，挺有意思的。

It's fun to pick one and stick with it.

我发现，很多时候用那些聚合工具，适合在大量工作、生成大量token的时候用。如果你只是个普通人，想用AI来处理日常任务，比如列购物清单、做菜、去健身房，或者给你一些健身课程和想法，那么选一个单一的模型、一个单一的服务，然后深入使用，会更有帮助。

I find that, like, oftentimes with the aggregators, the the right time to use that if you're doing lots of work, if you're generating lots of tokens, if you are just a person who wants to use AI for their day to day task to help you come up with, like, a grocery list, to help you cook specific things, to help you go to the gym and give you workout classes and ideas, it's helpful to pick a singular model, a singular service, and then just go deep with that.