OpenAI 发布 ChatGPT 5.4

OpenAI 刚刚推出了 ChatGPT 5.4。

OpenAI has just rolled out ChatGPT 5.4.

这里其实有几个我很兴奋的酷炫功能，我以前一直希望 ChatGPT 能做到这些，现在终于发布了。

There's actually a couple cool features in here that I'm really excited about, that I've been wishing ChatGPT has been able to do in the past, and they finally launched it.

当然，如果你看看他们的所有营销宣传，无非就是说这是他们迄今为止最强大的模型。

And, of course, if you look at all of their marketing, it's going to just basically be them saying, this is our most capable model yet.

当然，这是他们最强大的模型。

And of course, it's our most capable model.

如果不是的话，那他们更新的意义何在呢？

If it wasn't, I mean, what would they even be making an update for?

所以，我会跳过他们发布时所有的炒作和喧嚣，直接告诉你一些真正有趣的使用场景，以及我认为 GPT 5.4 真正有用的地方。

So, I'm just going to get past all of the hype and all of the buzz from what they said in their launch, and I'm gonna tell you some really interesting use cases and some ways that I actually think this is useful, GPT 5.4.

在进入这些内容之前，如果你想尝试所有最新的模型，去看看我的初创公司 AI Box dot AI 吧。

Before we get into all of that, if you wanna try all of the latest models, go check out my startup, AI Box dot AI.

我们汇集了来自前 15 家顶级 AI 公司的最新模型，从 Grok 到 Gemini，从 Anthropic 到 OpenAI，还有 11 Labs 的音频模型，以及大量出色的图像生成模型。

We have the latest models from the top 15 different AI companies, everything from Grok to Gemini to Anthropic to OpenAI, 11 Labs for audio, tons of cool image generation models.

我认为平台上总共有50多个模型。

I think there's over 50 models on the platform total.

你可以同时试用所有这些模型，每月仅需8.99美元。

You could try all of them side by side, and it's only $8.99 a month.

比ChatGPT便宜多了，但你能获得更多的模型。

So much cheaper than ChatGPT, but you get way more models.

当然，你还可以用它自动生成AI工作流，自动完成各种任务。

And, of course, you can also use it to automatically create AI workflows that can complete tasks for you that are automated.

所以有很多精彩的功能，如果你想以每月8.99美元的价格访问所有顶级模型，快去试试aibox.ai吧，如果选择年费计划还能再享20%折扣。

So there's a ton of cool stuff going on, but go check out aibox.ai if you want to get access to all of the top models for only $8.99 a month, and it's 20% off if you get an annual plan as well.

那里还有很多有趣的功能。

So, there's a lot of cool stuff there.

好了。

Alright.

我们来聊聊现在的情况。

Let's get into what's going on.

我想首先提到的是，这被称为GPT 5.4 Thinking。

The first thing I want to mention here is that this is called GPT 5.4 Thinking.

他们有一个性能更高的版本，称为GPT 5.4 Pro，但这两者共同设计用于处理从复杂分析到各种任务的一切。

They have a higher performance variant that is known as GPT 5.4 Pro, but both of these together are designed to kind of handle everything from some complex analysis.

它们能完成大量编程工作，跨越多种专业软件工具执行长时间运行的工作流，他们甚至将此称为自己的专业工作工具。

They do a lot of coding, a lot of long running workflows across a lot of different professional software tools, and they're kind of dubbing this as like their their professional work tool.

他们正试图让更多的职场专业人士使用它。

They're trying to get into, you know, into the hands of more working professionals.

这紧随他们与多家咨询公司达成的一系列协议之后，这些公司据称将把ChatGPT引入更多企业及专业环境。

And this is coming right on the backs of them signing a whole bunch of deals with a bunch of different consulting firms that are going to allegedly get ChatGPT into more businesses and in kind of the professional environment.

但与此同时，他们正陷入一场竞争，就连谷歌也参与其中，但主要是与Anthropic的Claude Code及其Codex工具竞争，他们正大力推动AI模型在软件中的应用以及计算机使用方式的变革。

Yet at the same time, they're having kind of this, you know, they're locked in a battle with, even Google's in this right now, but really with Anthropic for Anthropic's Claude Code, their Codex tool, they're really trying to push forward in kind of how software is using AI models and how computer use is going on.

所以，这才是他们真正专注的领域。

So, is where they're really focusing.

关于这一点，最大的变化之一就是规模。

One of the most one of the biggest changes basically about this is the scale.

因此，在API中，GPT 5.4的上下文窗口最大可达一百万个token，这使得它们能够处理巨大的文档、非常长的对话和大型数据集。实际上，如果你仔细想想，一个巨大的优势在于编程领域——你可以查看更大的代码库，真正地进行代码分析和操作。

So, in the API, GPT 5.4 has a context window of up to a million tokens, which basically lets them work with huge documents, really big conversations, big data sets, and really, I mean, if you think about this, a huge benefit is going to be coding where you can look at bigger code, you know, code bases to actually work with.

之前，Anthropic在这方面表现得非常出色，而现在OpenAI正试图迎头赶上。

So, this was something that Anthropic was really crushing it at, and then OpenAI is trying to get into this.

OpenAI还表示，他们的模型在token效率上更优，这一点其实让我非常期待。

OpenAI also says that their model is specifically more, what they're saying is token efficient, which I this is actually one thing that I'm excited about.

基本上，它可以用比GPT 5.2少得多的token来解决同样的问题，因此你的成本会大幅降低。

Basically, it can solve the same problems using a lot less tokens than GPT 5.2, so your costs are going to come down.

如果你已经在使用5.2的软件，这其实挺酷的——即使你没有，你所使用的很多软件都会用到它，成本会显著下降，而且速度也会快很多。

It's actually kind of cool if you already had 5.2 running in a software, which even if you don't, a lot of the software you use will, the costs come down a lot for that, and it also gets a lot faster.

所以，成本降低了，速度提升了，对我来说，这确实是一件让我非常兴奋的事情。

So, the costs come down and the speed goes up, and so, yeah, for me, this is something that I'm actually excited about.

至于基准测试的表现，我知道，我并不是想在这里挑刺那些百分比数字，但我确实想聊聊一些有趣的应用场景，以及为什么这些表现值得关注。

So, as far as how the benchmarks look, I know, you know, I'm not trying to like sit here and nitpick the benchmark percentages, but I did want to talk about some interesting use cases and reasons why these are why they're good.

具体来说，它在多个知名基准测试中都处于领先地位。

Specifically, it's, it's kind of leading on a bunch of the better known benchmarks.

其中之一是编程，当然我们知道为什么这现在很重要，还有计算机操作。

One of those is for coding, of course, we know why that's important right now, but also computer use.

这正是我现在感到兴奋的事情。

And this is something I'm excited about right now.

我觉得Anthropic在计算机操作方面做得非常出色。

I feel like Anthropic is really crushing it with computer use.

基本上，它能查看你电脑上的所有内容，点击各种东西，帮你完成任务。

Basically, it you know, can look at everything on your computer and go click on stuff and get stuff done for you.

这是一个我经常使用的用例，就是通过Claude的Anthropic浏览器，也就是Claude的Chrome扩展程序。

This is a use case that I've been using a lot with Claude's Anthropic browser, the Claude Chrome browser extension.

基本上，你点击一个按钮，它就会打开一个侧边聊天栏。

Basically, it's a button you click, it opens a side chat bar.

我会去一些非常复杂的用户界面或复杂网站，我不是开发者，但比如最近我需要在Google Cloud上做一些操作，来设置我在Lovable上构建的一个工具，我需要增强我的后端，以便能够做一些更高级的功能。

I go to really complex UI or complex websites, I'm not a developer, but if I'm going into like, for example, recently I had to do some stuff on Google Cloud to set up a tool that I was vibe building on Lovable, and I needed to beef up my back end so it could, you know, do some extra fancy stuff.

我其实完全不明白Lovable告诉我需要做些什么。

I didn't really understand anything that Lovable was telling me I needed to be able to do.

所以我打开了Claude侧边栏，告诉它：我现在在Google Cloud账户上，这是Lovable给我的操作说明，你去帮我设置一下。它就自己点击操作，帮我设置好了某些东西。

So I opened up the Claude sidebar, told it, look, I'm on my, you know, my Google Cloud account, go and here's the instructions from Lovable, and it clicked around and set up some stuff for me.

不过，我是不是应该让真正的开发者检查一下这些设置？

Now, should I have a real developer look over this?

我的意思是，现阶段我们就先不管那么多了，我都能听到开发者们此刻正对着耳机大喊大叫。

I mean, we're going to throw caution to the wind for the time being, and I hear all the developers screaming into their headphones right now.

但归根结底，事情办成了，我的软件现在能正常运行了，而且我不用花大量时间看那些冗长的YouTube教程，学习如何设置那些对我来说非常复杂的Google Cloud操作——毕竟我对编程一窍不通。

But, at the end of the day, it got it done and my software is now functioning, and I have, I I did not have to, watch a whole bunch of long YouTube tutorials on how to set up some complex I mean, for me, complex, because I have no idea how to code, Google Cloud stuff.

所以，这个应用场景在很多方面都令人惊叹，我认为OpenAI在计算机使用能力上的提升非常令人兴奋，因为他们将开始与Anthropic展开更直接的竞争。

So this is a really incredible use case for a lot of reasons, and I think OpenAI beefing up their capabilities in Computer Use is really exciting because they're going start competing more directly with OpenAI.

我的意思是，这并不是说只有Anthropic在做这件事。

I mean, this is not like or with Anthropic.

并不是只有Anthropic在研究这个方向。

It's not like Anthropic is kind of like the only one working on this.

OpenAI其实早就开始用智能代理做这些事了，但最近感觉他们的表现提升了很多。

OpenAI has been doing this for a long time with agents, but it feels like it's getting a lot better.

好的。

Okay.

另一个让我兴奋的是，它们在知识型工作方面变得越来越出色，也就是说，我认为每个人都在用它，所以我们只会看到一些渐进式的改进。

The other one that I'm excited for is they're getting a lot better at knowledge work, so, I mean, are kind of things that I think everybody uses it for, so this is something we're just going to see some incremental improvements on.

在OpenAI的GDP阀门基准测试中，该测试主要检查任务，涵盖了44种不同的职业。

On OpenAI's GDP Valve Benchmark, which basically checks tasks, It has like 44 different occupations.

它有点像向你展示如何将这些技术应用于不同领域的专业人士。

So, it's kind of like showing you how you can use this for different professionals.

在83%的对比中，它的表现超过了行业专业人士。

It is exceeding industry professionals in 83% of comparison.

所以，他们说：看，这些都是各行各业专业人士正在做的所有任务。

So, they're like, look, these are all the tasks that people in all of these different professional industries are doing.

在83%的情况下，它的表现优于行业专业人士，或者说，它在这些案例中给出的结果比行业专业人士更好，特别是针对知识型工作。

It is better than 83%, or it's, you know, it's beating what an industry professional might give you in 83% of these cases, specifically, I think for knowledge work.

而且它的表现相比GPT-5.2的71%有了显著提升。

And it has a really big jump from achieving about 71% that GPT 5.2 is getting.

所以从GPT 5.2升级到GPT 5.4，我们从71%提升到了83%。

So upgrading this to now GPT 5.4, we're getting from 71% to 83%.

它在知识工作方面将会好很多。

It just basically is going to be a lot better for knowledge work.

我的意思是，所谓好很多，是指我们看到了大约10%甚至12%的提升，这已经相当显著了。

Mean, and by a lot better, I mean, we're seeing, you know, a 10% jump here or, you know, 12% jump here, which is pretty significant.

在一些编程基准测试中，比如SWE Bench Pro，这是一个软件工程基准测试，模型比上一版略有改进。

On some of the coding benchmarks, so SWE SWE Bench Pro, this is a, know, software engineering bench pro, the model is getting slightly better than the last version.

所以，这当然不错，但除了略有改进之外，它的速度实际上快了不少。

So, I mean, this is good, but but, you know, beyond just getting slightly better, it is actually a bit quite a bit faster.

如果有人用过很多这些软件工具，特别是我们使用Cloud Code AI Box，我的开发者经常会发给我截图，因为AI在我们的后端代码库上执行了非常冗长复杂的任务。

So, if anybody has used a lot of these software tools, specifically, we use Cloud Code AI Box, and my developer sends me screenshots of like, because of these really long elaborate tasks that it's doing on our, our backend, our code base.

我发誓，他有个目标，就是看看能让他交给Cloud Code的项目连续运行多久而不中断。

And he, I, I swear it's like a goal for him to see how long he can get Cloud Code to run continuously without stopping on a project he gives it.

这挺有趣的，因为我在Lovable上写代码时，通常一两分钟内就能收到Lovable的回复。

It's funny because I'm, you know, vibe coding stuff on Lovable, and I usually get a Lovable response back to me in like, you know, a minute or two.

他做一项任务能连续运行三个半小时。

He has a go for like three and a half hours doing a task.

所以这个模型速度更快了，我感到兴奋，因为希望我们正在做的某些工作，这三个半小时能大幅缩短。

So, this model gets faster, I'm excited because hopefully that three and a half hours gets cut down on, you know, some of the stuff that we're working on.

我认为它还非常擅长真正的计算机交互。

I think one of the things that it's also very good at is for real computer interaction.

有一个叫 OS World Verified 的评估，它主要衡量 AI 操作桌面环境的能力。

There's an OS World Verified, it basically evaluates how well an AI can operate a desktop environment.

它基本上就是截个图，然后通过键盘和鼠标命令去点击各种东西。

It's it's, you know, pretty much just like takes a screenshot, and then it uses the keyboard and mouse commands to go and click stuff.

目前的成功率大约是百分之七十五。

Right now, has about a seventy five percent success rate.

我用过 ChatGPT 的代理工具。

I've used ChatGPT agents.

它并不完美。

It's not perfect.

这其实不是我的首选。

It's actually not my go to.

我用得不多。

I don't use it that much.

我希望我能多用一些。

I wish I could use it more.

我认为Anthropic在这方面做得更好，但75%的成功率说明他们确实在进步，成功率有所提升。

I think Anthropic is doing better in this, but 75% success rate, like they are improving, their success rate is up a bit.

它比GPT 5.2更好。

It's better than GPT 5.2.

但我仍然不觉得它是最好的。

I still don't think it's the best.

目前非常注重它在专业场景中的应用。

There's a major focus on kind of how it is being used professionally.

OpenAI表示，他们的模型目前在提供人们实际工作中真正使用的成果方面显著更出色。

OpenAI says their model right now is significantly better at, at basically giving the kind of deliverables that people use in real work.

所以像电子表格、演示文稿、财务模型、法律分析这些，它们已经完成了大量不同的任务，其中一项是由一名初级投资银行分析师完成的。

So, things like spreadsheets, presentations, financial models, legal analysis, all of those, they've done a bunch of different tasks, and they had one performed by a junior investment banker analyst.

它的得分是87%，而GPT 5.2的得分是68%。

It got 87% compared to 68% that GPT 5.2 got.

一些人类评估者也更喜欢它，大约68%的时间如此。

Some human evaluators also preferred it, about 68% of the time.

他们说它的视觉效果更好，结构也更清晰。

They said it had better visuals and better structure.

所以，有一些很酷的东西。

So, there's some cool stuff.

好的。

Okay.

一些你今天可能实际会用到的酷功能，这个是我非常期待的。

Cool features that you might actually use today, this is the one I'm very excited about.

它有一个他们称之为‘可引导性’的功能，但基本上，当你与ChatGPT对话时，这个功能也已集成在API中，我觉得这太疯狂了，但它确实已在ChatGPT中上线。

It has what they're calling Steerability, but basically, when you're you're talking to ChatGPT, it's it's available in the API too, which is I think crazy, but it's on ChatGPT.

如果你正在和ChatGPT对话，能隐约看到它的思考过程，比如它一步步推演，你发现它走偏了——比如说，你其实想找适合冲浪的顶级海滩，它却在分析考艾岛的海滩，你马上意识到：糟了，我在加州，根本不想看考艾岛！这时候你可以直接在对话中途输入：‘请只关注加州’。它会立刻吸收你刚说的内容，把新信息融入它的推理过程，并给出更新后的回应，这就是所谓的‘可引导性’。

If you're talking to ChatGPT and you can kind of see it's reasoning, right, like it's thinking through some stuff and it puts a couple steps down and you realize it's going in the wrong direction, you know, maybe you're like, hey, I'm trying to visit like, the best beach, for surfing, and it's like, okay, looking at beaches in Kauai, and you're like, oh crap, like I'm in California, I don't wanna see Kauai, and you're like, then you can set type a message like specifically in California and mid, like prompt, mid response, it actually takes into account what you just said and is, you know, steerability, it's going to go and incorporate that into its, into what it's looking at and into its reasoning and and give you an updated response.

所以，本质上你可以在对话中途输入提示，它会据此调整自己的推理，从而在过程中给出更优的回应。

So, basically, you can do mid response prompts and it's going take that into account and change its prompt and give you better prompt mid response.

这挺有意思的，因为我觉得他们在这里做了几件聪明的事，其中之一就是：当你提问时，你必须等待它思考、等待它回复，你只能干坐着等。

So, it's kind of interesting because I I think they did a couple clever things here, but one of them is when you ask a question, you have to wait for it to think, you have to wait for it to reply, you sit there and you wait.

我们都讨厌等待，但如果你在等待时能读到它的推理过程，并能随时补充输入和反馈，就会感觉等待的时间变短了。

We all hate waiting, and so, if in the middle of waiting, we're reading its line of reasoning, and we're giving it more input and more feedback, it feels like we did a lot less waiting.

你其实只是在浏览它的思路，顺便插一句建议，它就能更快、更好地完成任务，而不是非要等它把整个回答吐出来，你再回头说：‘不对，问题在这儿，应该这么改’。

We're really just kind of reading and trying to throw in something, and it can get it done faster and better, rather than having to wait for it to spit out the whole thing and you'd be like, okay, this is wrong, and here's why it's wrong, and here's what you should do instead.

你可以在对话进行中随时介入修正，我觉得这真的非常酷。

Like, you could do that in the middle of the chat conversation response, which is really cool in my opinion.

我现在还特别关注的一点是在线研究功能。

Something else I've kind of focused on right now is online research.

据说它能同时在更多网络来源中搜索，不再像以前那样：‘先看这个网站，获取点数据，再去看另一个网站’，而是能一次性在全网广泛搜索，再跨页面追踪线索。

Apparently, it can search across like a a greater number of sources on the web, so it can kind of, instead of just like, okay, we're looking at this website, getting some data, now we're going look at this website, It's going to go in and search just like a ton at the same time across the web, and then it's going to follow leads across different pages.

所以，它可能会从一篇文章中获得一个想法，然后去追踪另一篇文章，频繁地跳转。

So, it might get an idea from something to read on one article, it's going go follow that to another article, bounce around a lot more.

所以，它实际上是在进行更深入的研究，我知道我们之前就有深度研究功能，但现在它会把获取的所有信息整合成一个连贯的答案。

So, it's kind of doing, like, I know we have had deep research for a while, but it's doing deeper research, if that's the thing, and it's going to combine all of the information that it gets into one coherent answer.

因此，这在处理一些更复杂的问题时会更有用，因为这些信息往往分散在多个不同的网站上，而不是集中在一处。

So, basically this is going to be more useful for some of the more complex questions, where the information is kind of scattered across a lot of different sites instead of sitting in one place.

现在，并不是你问的每个问题都会用到这个功能，但当你遇到复杂的问题或需要深入回答时，它能更快地为你提供更连贯的答案。

Now, not every question you ask is going to be relevant, but sometimes when you have a complex answer or question, it's going to be able to go get you a more coherent answer quicker.

所以，这很棒。

So, this is great.

他们在发布时大肆宣传它幻觉更少、事实错误更少、准确性更高之类的说法。

They have all this like kind of, I don't know, fluff in their launch about how it hallucinates less, and it has less, you know, it has more, it has less factual errors and all this kind of stuff.

我觉得这并不是特别重要。

I don't think that's super important.

我们还听说，它会更少地拒绝你的请求。

One thing that we also heard about it is that it is gonna, it's gonna turn you down less.

所以，如果你问一个问题，它们可能会说：嘿，我不知道。

So, like, if you ask a question, and they're like, hey, you know, I don't know.

你问一个问题，据蒂姆·奥林说，它不太可能不作回答。

You you ask a question, it's gonna be it's less likely, allegedly, according to Tim Olin, to, like, not answer.

然而，我们的好友康纳·格兰南——他和我一起主持《AI应用》播客——他测试了这个功能。

However, our good friend, Connor Grennan, who hosts the AI Applied podcast with myself, he was testing it.

我看到他在领英上发了一条帖子，问他：‘真的吗？静脉输液中的气泡会让我死亡吗？'

I saw a post he made on LinkedIn where he asked it, is it true that, air bubble inside of an IV can cause me or can, you know, could kill me?

它似乎给他打出了完整的回答，就像我们之前看到的DeepSeek和中国审查模型一样：如果你问DeepSeek关于天安门广场的问题，它会先打出来，然后消失，再回复‘抱歉，无法回答这个问题’。

And it said, you know, apparently it typed out the whole response to him, kind of, and just like we saw with, like, DeepSeek and the Chinese censored model, if you ask anything about Tiananmen Square to DeepSeek, it like types it out and then it disappears and it's like, sorry, can't answer this.

显然，ChatGPT也说了完全相同的话。

Apparently ChatGPT said the exact same thing.

而且，现在正好是个敏感时刻，因为纽约目前正在试图通过一项立法，声称AI模型不能回答任何关于医疗、健康、法律的问题。

And also, this is kind of at a tricky moment because we're seeing New York right now is trying to pass some legislation where they they're saying, hey, we don't like, they're basically trying to pass legislation saying AI models can't ask, answer any questions about medical, health, legal.

它们涵盖了这些不同的领域。

Like they have all of these different areas.

我觉得就连美发师也在试图把它们加进去。

I think even hairstylists, they're trying to put in there.

这基本上是所有受到监管俘获的行业。

It's basically all of the, all of the different industries with regulatory capture.

他们就是不希望人们能免费获得这些答案。

They just don't want people to be able to get the answers for free.

所以，我对这项立法感到挺沮丧的，而且居然还有人认真考虑它。

So pretty, I don't know, kind of bummed about that legislation, and people like seriously considering that.

不过，看起来它似乎也没好到哪里去，但也许正朝着好的方向发展。

However, so it doesn't seem like it's that much better, but maybe it's moving in a good direction.

我也不完全确定。

I'm not a 100% sure.

感觉还是有一些模型更像‘屋子里的成年人’，但这些模型也有各自的优缺点。

It still feels like there's other models that are more of the adult in the room, but you also get pros and cons with those models.

Grok 以能回答你关于几乎所有这些话题的问题而闻名，但你知道，Grok 也可能存在其他一些缺点。

Grok famously is going to answer any question you have about basically any of those topics, but you know, there's a lot of different, there might be some other cons with Grok.

所以，每个模型都有其优缺点。

So, pros and cons to all of the models.

非常感谢您今天收听这个播客，如果您喜欢这一集，能在您收听播客的平台留下评分和评论，将对节目有很大帮助。

Thank you so much for tuning into the podcast today, If you enjoyed the episode, it would really help the show a ton if you left it a rating review wherever you listen to your podcast.

随便给我发个消息，说说您是否喜欢，来自哪里，以及对哪些话题感兴趣。

Just drop me a note, say if you enjoy it, you know, say where you're from, say what topics are interesting to you.

我都会阅读所有的评论和留言。

I read all the reviews and all the comments.

这真的非常有帮助。

It helps a ton.

另外，如果您想在一个地方访问所有这些最新的模型，别忘了去查看 aibox.ai，这样就不必为十个不同的平台分别支付20美元的订阅费。

Also, make sure you go check out aibox.ai if you want to get access to all of these latest models in one place, so you don't have to pay a $20 subscription to 10 different platforms.

每月只需8.99美元，您就能访问40多个不同的AI模型。

It's $8.99 a month, and you get access to over 40 different AI models.

快去试试吧。

So go check it out.

链接在描述中，aibox.ai。

Link in the description, aibox.ai.

我们下一期再见。

I'll catch you guys all in the next episode.