Anthropic的Opus 4.6与OpenAI的Codex：新一代AI模型角逐

Anthropic 刚刚发布了他们最新的顶级模型 Opus 4.6，我认为最重要的是，这是对它们最受欢迎的产品 Claude Code 的巨大升级。

Anthropic has just released their latest top of the line model, Opus 4.6, and I think most importantly, this is a huge upgrade to their most popular product, which is Claude Code.

就在他们做出这一重大宣布后的十五分钟，OpenAI 立即推出了自己的竞争产品，通过更新来应对这一 AI 模型。

Now, fifteen minutes after they made this big announcement, OpenAI immediately dropped their own competitor, to this to this AI model with their own update.

OpenAI 正在推出大量新功能，试图吸引那些使用 Anthropic 的企业用户。

OpenAI is doing a whole bunch of new, features to try to get to the enterprise users that use Anthropic.

过去几天，网上关于 Anthropic 投放广告的争论非常多，据说他们即将推出一条针对 OpenAI 的超级碗广告，而 Sam Altman 也在 Twitter 上做出了回应。

There's been a ton of beef online over the last couple days from open a from Anthropic running a ad, and allegedly, they're gonna have a Super Bowl ad that is coming after OpenAI, and Sam Altman is replying on Twitter.

今天这一期节目，我们将深入剖析这些纷繁复杂的戏剧性事件，并展示最新模型的一些非常酷的使用场景。

There's so much drama that we're gonna break down in this episode today and show some of the really cool use cases of the latest models.

所以有很多内容值得探讨。

So there's a lot to get into.

在开始之前，我想提一下，如果你想测试 Anthropic、OpenAI、Google、Grok 以及所有顶尖公司的最新模型，包括 Eleven Labs 的音频模型和众多出色的图像模型，请前往 aibox.ai 查看。

Before we do that, I wanted to mention if you wanna test out any of the latest models from Anthropic, from OpenAI, from Google, from Grok, from all of the top companies, including eleven Labs for Audio and tons of really cool image models, Go check out aibox.ai.

我们以每月 20 美元的价格让你在一个平台上访问所有顶级 AI 模型，并且你的所有文件都会集中保存。

We'll let you access all the top AI models in one place for $20 a month, and you get all of your files in one place.

你不需要订阅一大堆不同的服务，只需说明你想要什么类型的工具，我们的AI构建器就能将不同的模型连接起来。

You don't have to have tons of different subscriptions, and you can also vibe build tools by just explaining what type of tool you want, and our AI builder will link together different models.

所以你可以做很多有趣的事情，但我认为很多人喜欢的一点是，你可以在一个地方访问所有AI模型。

So there's a bunch of cool different things you can do, but I think a lot of people love the fact that you get access to all of the AI models in one place.

我会在描述中附上链接。

I'll link it in the description.

网址是 aibox.ai。

It is aibox.ai.

好的。

Okay.

让我们来看看刚刚发布的内容。

Let's get into what has just been released.

我认为在Opus 4.6之前，Opus 4.5是在去年十一月发布的。

I think prior to this kind of Opus 4.6, Opus 4.5 came out in November.

现在，Anthropic显然正努力扩大其模型的用户群体，不再局限于开发者、专业人士和企业这类小众用户。

And so I think, and right now Anthropic is obviously trying to get the reach of their model to to kind of get out of a small segment of users, which is mostly developers and professional and enterprise.

当然，用户群体小并不是坏事，他们赚的钱根本不是一点，因为大多数用户每月都要支付数百美元。

Now, not that a small segment is a bad thing, like they're not making a lot of money, they're making an insane amount of money because most of their users are paying hundreds of dollars.

当你谈到开发者时，API积分简直高得离谱。

The API credits are crazy when you're when you're talking about developers.

我个人每月在云积分上花费数百美元，用于开发和构建各种东西。

Personally, I pay hundreds of dollars a month in in, you know, clawed credits to to for development and and to build things.

所以，他们赚的钱肯定远超我每月20美元的OpenAI ChatGPT订阅费，但他们正努力扩大用户覆盖面。

So it definitely is making a lot more money than just my $20 a month OpenAI ChatGPT subscription, but they're trying to get a broader reach.

我认为其中一个最大的新增功能是他们现在称为‘智能体团队’的功能。

I think one of the biggest additions is what they are now calling agent teams.

与其依赖单个智能体依次完成任务，智能体团队能将大型任务拆分成多个小部分，根据Anthropic的说法，多个智能体可以同时处理这些部分。

So, rather than just relying on a single agent to work through different, you know, a task sequentially, agent teams allowed really large jobs to be broken into smaller pieces that multiple agents, according to Anthropic, can then tackle at the same time.

这是他们最重要的更新之一。

This is one of their biggest updates.

我们来听听他们是怎么说的。

Let's hear what they said about it.

他们说：‘与其让一个代理按顺序处理任务，你可以将工作分配给多个代理，每个代理负责自己的部分，并直接与其他代理协调。’

They said, quote, instead of one agent working through tasks sequentially, you can split the work across multiple agents, each owning its piece and coordinating directly with the others.

这是Anthropic在一份大型新闻稿中提到的。

This is Anthropic was saying this in a big press release.

Anthropic的产品负责人斯科特·魏斯表示，这个功能就像拥有一支非常优秀的团队，通过分工协作来完成任务。

Scott Weiss, who's, Anthropics head of product, basically said that this feature is like having a really talented team of humans working together, by segmenting responsibilities.

因此，不同的代理可以并行协调，从而实现更快的进度。

So, the different agents can actually coordinate in parallel, and then they can move much faster.

代理团队目前作为研究预览版向API用户和订阅者开放。

Agent teams are currently available as a research preview for API users and subscribers.

因此，这个新的Opus 4.5还将引入大幅扩展的上下文窗口。

And so this new Opus 4.5 also is gonna be introducing some really big context windows, that have been expanded.

我认为，很多人曾将这一点作为关键优势，比如长期以来，Anthropic的上下文窗口一直比OpenAI大，尤其是在Sora刚推出时，这就是很多人选择它的原因。

Now this is something that I think a lot people have used kind of as a leverage point, like classically or like famously, Anthropic was had a bigger context window than OpenAI for many years when ShyGPT first came out, that's why a lot of people would use it.

对我来说，这基本上是我最初使用Anthropic的主要原因——它的上下文窗口要大得多。

And this is, I think, basically the first use case for myself with Anthropic, why I was originally using it was, it was just a much bigger context window for things.

然后，当然，OpenAI 也开放了它们的上下文窗口，接着 Google 推出了超大上下文窗口。

And then, of course, OpenAI kind of opened them up, then Google came out with a massive context window.

因此，Anthropic 迎头赶上，强势出击。

Because of this, Anthropic has come out swinging.

它们的新模型现在支持高达一百万令牌的上下文，这正是 Google 凭借 Gemini 推出后获得大量使用的功能。

Their new models now support up to a million tokens of context, which is what Google famously did with Gemini that got it a lot of usage.

而这与 Anthropic 已经在 Sonnet 4 和 4.5 中提供的能力相匹配。

And this is matching what Anthropic already offers with Sonnet four and four point five.

如此庞大的内存使得处理大型代码库、大文档变得容易得多，让你能在单次对话中完成更复杂的多步骤工作流，这真的很酷。

So, that really massive memory makes it much easier to work with kind of huge code bases, large documents, and it lets you do a lot more complex multi step workflows just with a single, kind of in one single conversation, which is really cool.

此外，Anthropic 正在深化 Claude 与日常生产力工具的集成。

So, Anthropic is also deepening Claude's integration with everyday productivity tools.

彭博社有一篇文章称，最近几天市场上发生的 SaaS 低迷，是因为 Anthropic 推出了与众多 SaaS 工具更深层次的集成，人们开始意识到，也许我们根本不需要这些 SaaS 产品了。

There was a post on Bloomberg that said, basically, the SaaS crash in the market that's been happening over the last couple of days is due to Anthropic releasing some of these deeper integrations with a bunch of SaaS tools, and people are realizing maybe we don't even need the SaaS products at all.

也许，Anthropic、Claude 以及它们所替代的这些工具，因为推出了大量新工具，正在冲击市场，因为人们将不再为其中一些工具付费。

Maybe, Anthropic and Claude and what they're doing and some of these tools that they're replacing, because they came out with a bunch of different tools, are crashing the markets basically, because people are gonna stop paying for some of these tools.

因此，有了 Opus 4.5，Claude 现在可以直接作为侧边栏嵌入 PowerPoint 中。

So, with Opus 4.5, Claude now is directly inside of PowerPoint as a side panel.

以前，用户可以要求 Claude 生成演示文稿，但之后你仍需导出、编辑，它是一个独立的文件，非常麻烦。

Previously, users could ask Claude to generate a presentation, but then you still have to go export it, you had to edit it, and it was kind of a separate file, which was a pain.

现在，你的演示文稿可以直接在 PowerPoint 中构建，并借助 Claude 的帮助进行定义和优化。

Now, your presentations are gonna be built and you can actually go define them and refine them directly inside of PowerPoint with Claude's assistance.

这是一个很酷的功能。

So that's a cool feature.

此外，我安装了 Claude 的 Google Chrome 扩展程序，它也是一个侧边栏。

In addition, I mean, I have the Claude I have the Claude Google Chrome extension, which, you know, is is a side panel.

如果我在 PowerPoint 里，我可以让它去进行编辑，它就能直接接管我的屏幕完成操作。

And if I was on PowerPoint, I could tell it to go and make edits, and it could just like take over my screen and do that.

这同样是一个很酷的功能。

So that's also kind of a cool feature.

但根据 Anthropic 的说法，所有这些变化都是为了展示 Opus 已经超越了其最初面向开发者的定位。

But according to Anthropic, all of these changes are trying to basically show how Opus has evolved beyond just kind of its original niche, was developers.

最初作为主要以软件开发闻名的模型，现在已经发展成能够支持更多人完成各种不同专业任务的工具。

What started as a model, which was basically known for like software development, has now grown into something that can do a lot more it it can support a lot more people and doing a lot more different professional tasks.

以下是他们最近的表述。

Here's what they recently said.

他们说，我们注意到许多非专业软件开发人员也在使用Cloud Code，主要是因为它是一个极其出色的任务执行引擎。

They said, we noticed a lot of people who are not professional software developers using Cloud Code primarily because it was a really amazing engine to do tasks.

他们表示，Anthropic的用户群体现已不仅限于工程师，还包括大量产品经理、金融分析师以及众多其他领域的专业人士。如今，竞争显然正在加剧。

They said that Anthropic has now seen adoption, not just from engineers, but from a lot of product managers, a lot of financial analysts, a lot of professionals in a lot of different Now the competition is definitely heating up.

Anthropic并不是唯一一家瞄准开发领域的公司。

Anthropic isn't the only one going after the development area.

OpenAI正大力推广其产品，我最近一直在测试Codex，这是OpenAI推出的一款工具。他们其实早已拥有Codex代码模型，但最近推出了一款Mac应用程序，我下载后一直在体验。

OpenAI is trying to really go pitch heavily for I've recently been testing out Codex, which is a tool that OpenAI has they've they've kinda had this Codex code model for a while, but they recently came out with a Mac app that I downloaded and they've been playing around with.

他们正在试图与Claude Code竞争。

They're trying to compete with Claude Code.

有趣的是，在Anthropic此次最新发布中，据称OpenAI和Anthropic原本计划在同一时间发布各自的新型编码模型。

And what's interesting is in this brand new release that Anthropic did, apparently OpenAI and Anthropic were scheduled to release their new coding models at the exact same time.

Anthropic 提前了十五分钟发布，抢在 OpenAI 前面。

Anthropic bumped it up by fifteen minutes to beat OpenAI to the punch.

所以，在 Anthropic 发布之后，十五分钟不到，GPT 5.3 Codex 就推出了。

And so after, you know, Anthropic did this whole release, fifteen minutes later, GPT 5.3 Codex was launched.

因此，OpenAI 表示，这个新模型将把 Codex 从一个只能编写和审查代码的工具，转变为能够处理开发者在电脑上做的几乎所有事情的工具，包括专业人士的工作。

So OpenAI says that this new model is gonna turn Codex from a tool that can write and review code into one that can handle almost anything developers do on a computer, including professionals as well.

他们说，在用一系列内部基准测试后，他们声称 GPT 5.3 Codex 能在数天内从零构建出功能完备的复杂游戏和应用程序。

They said that after they tested a bunch of against a bunch of different, like, internal benchmarks, they say, I mean, this is their claim that GPT 5.3 codecs can build, quote, highly functional complex games and applications from scratch over the course of days.

目前，OpenAI 称他们的模型比 GPT 5.2 快约 25%。

OpenAI right now is saying that their model runs about 25% faster than GPT 5.2.

他们还表示，这是他们首次创建出在开发过程中大量用于自我调试和评估的模型。

And they also said that this is the first model that they have created that was used, like, they use this a ton to help them debug and evaluate itself as they were building it.

所以，他们用 GPT 5.2 来开发 GPT 5.3，这对他们来说意义重大。

So they were using GPT 5.2 to develop GPT 5.3, which is a big deal for them.

我认为，这个时间点绝对不是偶然，针对 Anthropic 的吧？

I think the timing definitely wasn't an accident opening an Anthropic, right?

把这些产品相隔十五分钟发布，其实OpenAI本想同时发布，以抢走对方的风头。

Releasing these things fifteen minutes apart, and really OpenAI was trying to release it at the same time to try to steal some of their thunder.

目前，OpenAI非常有雄心，而且其视野已经超越了编程领域。

OpenAI right now is very ambitious, but it's also looking farther beyond coding.

就在本周，他们还推出了OpenAI Frontier，这是一个端到端平台，旨在帮助企业构建、部署和管理AI代理。

Just this week, they said, they also unveiled OpenAI Frontier, which is an end to end platform designed to help enterprises build, deploy, and manage AI agents.

简而言之，Frontier是一个开放平台，意味着企业不仅可以管理OpenAI生态系统内构建的代理，也可以管理外部构建的代理，这相当酷。

Basically, Frontier is an open platform, meaning that companies can manage agents built outside of OpenAI's ecosystem as well as ones that are built inside of it, which is kind of cool.

但通过Frontier，不同企业可以将代理连接到外部数据和应用程序，然后定义这些代理可以访问的内容，并限制它们能执行的操作。

But with Frontier, different businesses can connect agents to external data and applications, and then they can kind of define what those agents are allowed to access, and they can limit what actions they can take.

OpenAI表示，该系统借鉴了企业管理员工的方式，以及入职流程和反馈循环，目的是随着时间推移不断提升代理的性能。

OpenAI says that the system is modeled after how companies manage human employees, and how that kind of competes with onboarding processes and feedback loops, which are intended to try to improve agent performance over time.

OpenAI还提到，他们展示了一些正在使用这一平台的客户。

OpenAI also said, you know, they highlighted a bunch of different customers that are using this.

他们表示，惠普、甲骨文、州立农场、优步等公司都在使用Frontier，而且显然非常满意。

They say HP, Oracle, State Farm, Uber, all of those are using Frontier and apparently, you know, loving it.

我不确定，如果你这是来自OpenAI的说法的话。

I don't you know, if you this is from According to OpenAI.

他们表示，定价尚未公布，也没有就与此相关的任何成本发表评论，这很有趣，因为这虽然是一次发布，但你还是想知道具体要花多少钱。

They said that the pricing hasn't been disclosed, and they haven't really commented on any sort of costs associated with this, which is interesting because it's, I don't know, it's like an announcement, but you kind of want to know what the damage is going to be.

我认为，对于这些AI公司来说，整个代理管理平台如今已经变得非常重要。

The whole agent management platforms, I think, have really become very important for these agent, these AI companies right now.

代理在实用性方面正在迅速崛起。

Agents are just surging in how useful they are.

而且，自2024年以来，Salesforce就推出了他们自己的产品——Agentforce。

And, you know, even since 2024, Salesforce launched their own product, which is Agentforce back in 2024.

我认为还有很多其他公司正在使用像LangChain和CrewAI这样的工具，这些工具也获得了大量风险投资，以在这一领域展开竞争。

I think there's a lot of other people that using tools like Lanechain and CrewAI that also have raised a ton of venture funding to kind of compete in that space.

在去年12月，高德纳将代理管理平台描述为‘AI领域最有价值的资产’，并称其为AI普及的关键基础设施。

And in December, Gartner described agent management platforms as both, quote, the most valuable real estate in AI, and also basically a really important piece of infrastructure for AI adoption.

我认为，如果你结合这一点来看，OpenAI今年年初大力进军代理管理领域也就不足为奇了。

I think if you look at that and you kind of look at that context, OpenAI kind of making this big push into agent management early this year is not very surprising.

今年，他们将企业采用作为重中之重。

They've made, you know, the enterprise adoption a huge focus this year.

他们已经宣布了与ServiceNow和Snowflake的一些重要合作关系。

They already have some really big partnerships with ServiceNow and Snowflake that they've announced.

因此，我认为如果OpenAI希望成为企业市场的长期主导力量，Frontier显然是朝着这个方向迈出的一步。

So I think if OpenAI hopes to become kind of the long term force in the enterprise market, Frontier looks like a really good step in that direction.

当然，竞争非常激烈。

And of course, there's a lot of competition.

我们在关注Anthropic，也在关注OpenAI。

I mean, we're looking at Anthropic, we're looking at OpenAI.

现在每个人都正在争夺成为引领智能体的平台、引领开发者的平台。

There's just everyone's battling right now to be the platform that leads agents, the platform that's, you know, leading developers.

我认为竞争非常激烈，但看到这两大玩家相互较量也很有意思。

I think that the competition is fierce, but it's interesting to see these two players kind of compete against each other.

非常感谢您收听今天的播客。

Thanks so much for tuning into the podcast today.

如果你喜欢这期节目，能给我留下一个评分或评论，我会非常感激。

If you enjoyed this episode, it'd mean the world to me if you left a rating or review.

说实话，这能大大帮助节目被更多人发现。

Honestly, it just helps the show out a ton to get found by people.

如果你喜欢这期节目，或者觉得它对你有帮助，或者希望我探讨其他主题、类型，欢迎在评论区告诉我，同时在节目下方留下评论或评分也会非常有帮助。

And I always love to hear what you think if you enjoyed an episode, or if, you know, this is something that's useful for you, or if you'd like me to cover different ideas, topics, genres, like, let me know in the comments, and in the as a review on the show helps a ton.

非常感谢你的收听，我们下一期再见。

Thanks so much for tuning in, and I'll catch you guys all in the next episode.

一如既往，请访问 aibox.ai 以一站式获取所有最佳模型。

As always, make sure to check out aibox.ai to get access to all of the best models in one place.

下期再见。

Catch you in the next episode.