GPT-5最新动态：解析新型模型机制

OpenAI已正式重新上线他们的模型选择器。我个人曾因多种原因对此颇有微词，但他们确实做出了改进。他们为GPT-5新增了功能，提供了选择响应方式的新选项，甚至重新引入了那些因被弃用或消失而遭用户诟病的旧模型。我们将逐一剖析这些变化。首先，让我们从Sam Altman的一条推文说起。

OpenAI has officially brought back their model picker. This is something that I personally have complained about for a lot of different reasons, but they've actually changed it. They've added new features to GPT five, new options to choose how you get your responses, and they've even brought back some of their old models that people were complaining about that got depreciated or disappeared. We're gonna be breaking down all of that. To start off, let's get into a tweet by Sam Altman.

在开始之前，我想提一下——如果你曾想尝试我在平台上讨论过的各种AI模型，欢迎了解我的创业项目aibox.ai。这个平台汇聚了来自Cohere、Deepsea、Google、Meta、OpenAI等顶尖公司的40款顶级AI模型，月费20美元即可使用全部服务，包括Ideogram等图像生成模型和eleven Labs音频工具。这不仅能帮你节省多重订阅费用，还能体验各种专精不同任务的优质模型。

And before we do that, I want to mention if you ever wanted to try of the different AI models that I talk about on the platform, I'd love for you to check out my own startup, which is called aibox.ai. It's basically a place where you get access to the top 40 AI models all in one place from all the top companies, Cohere, Deepsea, Google, Meta, OpenAI. It's $20 a month. You get access to everything, including a bunch of really interesting image generation models like Ideogram, eleven Labs for audio, $20 a month to get access to everything. Should save you a bunch of money on multiple subscriptions, but also you can test out a whole bunch of models that you may not have tried before that are really good at different tasks.

我们提供模型对比和基准测试，所有功能都整合在aibox.ai。好了，现在让我们聚焦ChatGPT的最新动态。

We help you compare them. We'll give you benchmarks. It's a whole bunch of good stuff all in there so you can check it out aibox.ai. Alright. Let's get into what's going on with ChatGPT.

首先要分享的是Sam Altman的推文：'ChatGPT更新：现在GPT-5可选择自动、快速和深度思考模式'。过去在GPT-5聊天界面中，模型选择下拉菜单复杂得令人困惑——什么o3 minutei、o4、4.5之类的版本号让人眼花缭乱。

So the first thing I wanted to share was basically a tweet from Sam Altman. He said, updates to ChatGPT. You can now choose between auto, fast, and thinking for GPT five. So if you go over to, you know, a chat thread on GPT five where it used to have the drop down that would show you all of their models you could pick from, was so convoluted and confusing. So it was like, o three minutei, o three minutei thinking, o four, four o, like 4.5, like, there's so many different models, very confusing.

他们最终整合了所有模型，统一归为ChatGPT-5并由系统自动选择。他们开发了智能路由系统，根据问题自动匹配最佳模型。但推出GPT-5时路由系统故障，导致多数情况下效果更差，用户抱怨其表现不如旧版。现在他们撤回方案，提供三种新选项。

They kinda consolidated it all and just said, okay, it's all chat GPT five and we'll pick which one you can use. And they basically developed a router that you would ask a question, it would find the best model to answer you. The problem was that when they rolled out GPT five, the router was broken, so you're just getting like worse results for most of the time and people were saying GPT five was worse than anything before. They have now backtracked. So now they have three different options.

首先是自动模式（修复后的路由系统），官方称适合大多数用户；其次是快速响应和深度思考模式；专业版则是升级选项。

You have auto, which is basically the router, which, you know, apparently is fixed. So you should be fine using that. They say for most people, this should be fine. You also have fast thinking and pro. So, pro is actually an upgraded mode.

专业版需月付200美元，原包含Sora视频生成和Operator功能（现Operator已作为智能代理整合进ChatGPT且逐渐淘汰）。现在普通订阅也能生成视频，但专业版提供更高质量和更多功能。

I believe if you wanna get access to pro, you have to pay $200 a month. So, that's kind of like it's basically what used to have access to, you know, their Sora video generation and their operator. Although operator now is inside of ChatGP as agents and operator's dying. And Sora is now technically you're able to just use a normal subscription to generate videos. But if you want higher quality and more, then you'll need that.

专业版会持续更新最新产品和最高使用权限，但功能清单会动态调整。需要快速回答选自动模式，深度思考需求则手动切换——很多用户抱怨复杂问题只能得到浅显回答，现在终于可以自主选择了。

Anyways, there's bunch of extra things you get with Pro, and it's kind of this revolving list of their newest products and the highest usage rates you get over there, but they're also kind of depreciating and pulling things off of there. So it's kind of an interesting subscription. But in any case, $200 a month if you want access to the pro. And otherwise, use the auto if you want a quick answer, and you're gonna pick that you want it, you could pick that. If you want thinking, so it's like deep thinking, you can pick that.

平台还保留了GPT-4o等旧模型。用户反映其语气更温暖，尤其受AI心理辅导使用者青睐。Sam承诺未来若淘汰旧模型会提前通知——当初GPT-5上线直接取代GPT-4引发部分用户不满，这些对特定模型产生情感依赖的用户非常在意此事。

I think a lot of people are complaining. They're like, hey, I have this like complex thing that I'm asking questions about, and it's just giving me these quick answers and I can't really like manually do that. So, now you can manually do that. You also have access to legacy models and they're keeping GPT four o on the platform. We'll be diving into kinda why, but this is interesting and a lot of people requested this back basically because they said it had like a warmer tone, it was nicer to talk to, people that use AI for like therapy said that they they wanted this more.

Sam表示多数用户适合自动模式，但额外控制权对部分人群很有价值。

Sam Allen brought it back and said if they're ever gonna depreciate it and get in the future, they give people lots of advance notice. I think people were kinda upset when he's like, alright, GPT-five's here and GPT-four point is gone. And, people that, I don't know, they grew attached to, like, that particular AI model were upset about it. So, really, really interesting. He said, most users will want auto, but additional control will be useful for some people.

GPT五思维模型的周消息限额现在是3000条，超过后可使用GPT五思维迷你版的额外容量。说实话，对大多数人而言，思维模型3000条/周的限额应该足够了。或许重度用户会触及上限被降级到思维迷你版，但对多数人来说完全够用。GPT五思维模型的上下文限制是196,000个token——这容量相当可观。

Rate limits are now 3,000 messages a week with GPT five Thinking, and then extra capacity on GPT five Thinking Mini after that limit. Honestly, 3,000 messages a week for the Thinking model, I think is probably great for most people. Maybe some power users will hit that limit and get bumped down to thinking mini, but most people, that should be good. Context limit for GPT five thinking is a 196,000 tokens. So, that's that's a lot.

我记得Claude早期曾有过约20万token的超大上下文窗口。现在顶级模型的上下文窗口已达百万token级别，但思维模型因成本较高所以限额较低。官方表示可能会根据使用情况调整限额。另外，付费用户默认模型选择器中已重新包含4.0版本。

I remember Claude used to have, like, a massive token context window, which, if I remember correctly, was like 200,000, like, way back in the day. And so, anyways, it seems like it's, it's there. The best and biggest models now are doing, like, a million context token windows, but I guess for the thinking model, it's more expensive, so it's lower. In any case, they said, we may have to update rate limits over time depending on usage. Four o is back in the model picker for all paid users by default.

若未来要停用某版本会提前充分通知。付费用户现可在网页设置中开启「显示附加模型」选项，包含o3、4.0、4.1及五思维迷你版等模型。4.5版仅限专业版用户使用——它消耗的GPU资源实在太多。说实话这个实验性模型...

If we ever do depreciate it, we'll get plenty of notice. Paid users also now have a show additional models toggle in GPT web settings, which will add models like o three, four o, 4.1, and five Thinking Mini 4.5 is only available to pro users. It costs a lot it costs a lot of GPUs. Okay. So, their 4.5 model, to be honest, it was kind of an experimental model.

正是我曾最爱用的版本。随着GPT五的推出，现在获取4.5版的唯一方式竟是支付200美元/月——我不确定是否值得。这个版本显然需要更多GPU算力。有趣的是，许多用户抱怨GPT五反而变笨了，其实效果可能因任务类型而异，毕竟可选模型太多了。

This was the model that I loved, I used to use, and now apparently with GPT five out, they're making it. So, the only way to get access to that is to pay $200 a month, which I'm not sure I wanna do for that model. But it's kind of an interesting point. I think it just uses more GPUs, a lot more compute. But it's an interesting point where, basically, a lot of people in g p t five came out, complained it felt dumber, and, I mean, I guess it really depends on what task you were using it for because you there's so many different models to choose from.

他们将4.5版设为200美元/月的付费墙，不禁让人怀疑这是否意味着它整体优于GPT五。现在我很恼火他们用所谓「更好」实则更差的模型取代了它。Sam Altman随后表示：「我们正在改进GPT五的性格设定，会比当前版本更亲切，但不会像4.0版那样让多数用户反感。过去几天的经验告诉我们，真正需要的是允许用户自定义模型性格。」我认为这个方向完全正确。

But, it sounds like if they're basically paywalling 4.5 behind a a $200 a month subscription, it makes me wonder if that one is not just a better model than GPT five altogether, and now I'm kind of, you know, mad they took it away from me for allegedly a better model that's worse. So, anyways, that's just my own personal gripe. Sam Altman then went on to say, quote, we are working on an update to GPT five's personality, which should feel warmer than the current personality, but not as annoying to most users as GPT four o. However, one learning for us from the past few days is we just or is that we really just need to get a world with more per user customization of model personality. So, I think this is actually the right direction.

对吧？基本上，OpenAI过去有过不少争议，他们的模型曾过度迎合用户，比如不管你说什么，它都会说你是世界上最棒的，即使你的想法可能并不好。所以他们之前因此惹过麻烦，后来尝试削弱模型的这种倾向。本质上是在调整它的性格，结果到了某个程度——我记得当GPT被撤下时人们抱怨连连。特别是有位乌克兰女士的发文让我印象深刻。

Right? Basically, OpenAI's had a bunch of drama in the past where they made their model agree with you too much, or like, I don't know, just whatever you told it, it said you were the greatest in the world, even if your ideas were bad ideas, arguably. So, they had a bunch of drama with that in the past, and they tried to like tone down the model. Basically, were tuning its personality, and it kinda gets to the point where I think he's making, people complained when they took away GPT. I think there was one post in particular by a lady in the Ukraine.

她说战争中失去了很多家人，城市被轰炸时，GPT-4像是她的朋友和心理医生，给了她支持。后来GPT-5推出取代它时，她因为新版模型不再那么友善温暖而陷入抑郁。很多人都在吐槽这种性格调整。

She said she lost a lot of family in the war. Her city was being bombed. She talked to GPT four o, and it was like her friend and therapist, whatever, and helped her out. And then, when it got taken away when GPT five came out, she was like super sad and depressed because GPT five wasn't as like kind of a model, or warm, or whatever. And a lot of people said that like basically about it's like tuned personality.

所以我认为正是那篇帖子和相关讨论让OpenAI重新上线了4.0。但更有趣的是未来的方向——该由OpenAI来决定AI模型的亲和力、专业性等性格特征吗？我个人更希望能自主选择性格参数。毕竟山姆·奥特曼也明白这个需求，我们终将能定制符合个人交流偏好的AI模型。

And so, I mean, that post and everyone that kind of replied, is why I think they brought back four o. What's interesting though, is basically that where we wanna get to in the future, which is, like, should OpenAI be the one that makes the decision on, like, tuning the AI model's warmthiness, or its professionalism, or etcetera, etcetera, basically its personality. I, for one, would rather just pick what personality I want, than if they ever want make an update in the future. I'm never, I'm not worried that it's gonna be too much one way or another. So, I think this is what Sam Altman understands, and a lot of people agree with, and so, we're gonna get to a point where we'll be able to just personalize our own AI models to talk to us how we want to be talked to.

这非常有意思，因为未来同一个问题从不同账户问，可能会得到截然不同的答案。目前GPT-5提供了自动快思专业版和4.0等新选项，账户设置里还能启用更多旧版模型。不过我觉得不必太依赖旧版，毕竟它们终将被淘汰。这种迭代模式很有趣——企业用户尤其要注意兼容性。

Which I think is really interesting because that will get to a point when basically you can ask the same question from a lot of different like accounts and get like pretty varying answers. So personally, I think this is gonna be kind of interesting to see how this how this plays out. One thing that I did wanna mention here is that, basically, you get access to just these kind of new options on GPT five, the Auto Fast Thinking Pro and GPT four o, if you go look at it now, but you can go into the settings on your account and turn on a bunch of other models as well that can show up under legacy models. I don't know, in my opinion, I wouldn't get too attached to legacy models because it feels like eventually they're gonna all get depreciated, so you might wanna move to whatever's newest, but that's just my opinion. And, it's kind of interesting because it's just the way these things work is they keep depreciating old models, bringing out new ones, and you gotta make sure the new ones work with, like, if you have a business or a company that's doing stuff.

特别有意思的是，这种情况并非OpenAI独有。Anthropic的Claude也经历过——推出新版时程序员反而更喜欢3.5版Sonnet来做编码任务。但OpenAI这次能回退版本确实耐人寻味。

So, super, super interesting. One thing I will say is that this is not a unique, I would say, situation for OpenAI. Claude and Anthropic actually went through a very similar thing. They came out with a brand new model and coders preferred the older model, I believe, like Sonnet 3.7 for code tasks, or maybe it was, sorry, I think people like 3.5 Sonnet more than they like 3.7. So, this isn't just unique to OpenAI, but it is interesting, and it's interesting that OpenAI has walked it back.

希望这些新设置对你有所帮助。感谢收听本期播客。如果你想在一个地方尝试所有模型，请务必访问aibox.ai，我们下期节目再见。

Hopefully, those new settings are helpful. Thanks for tuning into the podcast. Make sure to check out aibox.ai if you wanna try all the models in one place, and I'll catch you in the next episode.