How to Maximize GPT-6 Astra

如何最大化利用 GPT-6 Astra

GPT-6 Astra was recently released. I got access to it on Friday evening, European time, and have been using it extensively ever since. In this article, I’ll share my first impressions of the model and the techniques that I utilize to get as much as possible out of the model to maximize its effectiveness. GPT-6 Astra 最近发布了。我在欧洲时间周五晚上获得了访问权限，并从那时起一直在广泛使用它。在本文中，我将分享我对该模型的第一印象，以及我为了最大限度地发挥其效能而采用的技巧。

I’ll also share some of the downsides I’ve experienced with the model and how I’m attempting to deal with them in the coming weeks, when I’ll be using the model more and more, and we’ll be trying to maximize how productive I can be with the model. This infographic highlights the main contents of this article. I’ll discuss how to get the most out of GPT-6 Astra and my first experiences with the model. Image by ChatGPT. 我还将分享我在使用该模型时遇到的一些缺点，以及在接下来的几周里我打算如何应对这些问题。届时我将更频繁地使用该模型，并努力将其生产力最大化。这张信息图突出了本文的主要内容。我将讨论如何充分利用 GPT-6 Astra 以及我对该模型的初步体验。图片由 ChatGPT 生成。

Why use GPT-6 Astra

为什么要使用 GPT-6 Astra

Firstly, I always like to cover why you should care about the topic of an article. In this case, it’s because GPT-6 is the latest release from OpenAI, which is one of the frontier labs. This naturally makes it one of the LMs that you immediately want to try out to see how well it performs on my workflows. My workflows mostly consist of coding tasks, though I do have some other tasks such as using my computer, moving around in the browser, and doing research, so measuring how well it works for deep research and so on. 首先，我总是喜欢先说明你为什么要关注文章的主题。就本例而言，这是因为 GPT-6 是 OpenAI（前沿实验室之一）的最新发布。这自然使它成为你想要立即尝试的大语言模型之一，看看它在我的工作流程中表现如何。我的工作流程主要包括编码任务，尽管我也有一些其他任务，例如使用电脑、在浏览器中操作以及进行研究，因此需要衡量它在深度研究等方面的表现。

I use coding agents essentially every day, and thus I believe I have a very good baseline to compare new model releases against. GPT-6 is also a very hyped release, and it’s a new major release from a frontier lab, which is of course extremely exciting. 我几乎每天都在使用编码智能体，因此我相信我有一个非常好的基准来对比新发布的模型。GPT-6 也是一个备受炒作的版本，作为来自前沿实验室的重大新发布，这当然令人极其兴奋。

My first impressions of GPT-6 Astra

我对 GPT-6 Astra 的第一印象

First of all, let’s cover my first impressions of the model. Immediately, once I started using the model, I started testing a few things: Running it on tasks that I’ve done before; Running it on some new tasks (feature implementations and bug fixes); Starting to look for refactoring opportunities and, in general, opportunities to improve my code repositories. 首先，让我们谈谈我对该模型的初步印象。一旦开始使用该模型，我立即进行了几项测试：在之前做过的任务上运行它；在一些新任务（功能实现和错误修复）上运行它；开始寻找重构机会，以及总体上寻找改进代码库的机会。

My first impressions will thus be based on my experience when running these three tasks. I would like to note, however, that first impressions might not give the best picture of how good the coding model is. The reason I say this is that I remember when GPT 5.6 Sol was released, I got extremely good first impressions of the model, and I still would say it’s a very good model. However, over time, I started to notice some quirks with the model where it didn’t work as well, and these quirks were something I didn’t discover for a week or two by actively using the model. Thus, I wouldn’t have discovered such quirks yet when it comes to GPT 6. 因此，我的第一印象将基于我在运行这三项任务时的经验。不过，我想指出的是，第一印象可能无法最准确地反映编码模型的好坏。我之所以这么说，是因为我记得 GPT 5.6 Sol 发布时，我对它的第一印象非常好，而且我至今仍认为它是一个非常好的模型。然而，随着时间的推移，我开始注意到该模型的一些小毛病，表现得没那么好，而这些毛病是我在积极使用该模型一两周后才发现的。因此，对于 GPT-6，我目前还未发现此类问题。

Nonetheless, I’ll give my review. So when it comes to running tasks that I’ve done before and verified, or that I verified that both Claude Code and GPT-5.6 Sol were able to do, I would say GPT-6 was, of course, able to do them all. But one thing I actually noticed in favor of GPT-6 is that it achieved it way faster, and I don’t think this has anything to do with inference speed. In my experience, this just seems like the model is more effective at utilizing its tokens and is more able to complete tasks quickly. 尽管如此，我还是会给出我的评价。当运行我之前做过并验证过的任务，或者验证过 Claude Code 和 GPT-5.6 Sol 都能完成的任务时，GPT-6 当然都能胜任。但我注意到 GPT-6 的一个优势是它完成得更快，我认为这与推理速度无关。根据我的经验，这似乎是因为该模型在利用 Token 方面更有效，并且更有能力快速完成任务。

This impression was further verified once I started to run the model on new tasks as well, both finding bug fixes and feature implementations. It just seems like the model was able to complete tasks super quickly while also doing it correctly, at least when comparing it to both the previous generation of OpenAI models and when comparing it against Claude Fable 5, which is the other frontier model available on the market. 当我开始在新的任务（包括查找错误修复和功能实现）上运行该模型时，这一印象得到了进一步验证。该模型似乎能够超快地完成任务，同时还能保证正确性，至少在与上一代 OpenAI 模型以及市场上另一款前沿模型 Claude Fable 5 相比时是这样。

I found this very useful and thus immediately switched to GPT-6 as my main driver to perform coding tasks. Currently, I don’t see any area in which Fable 5 is superior, other than the fact that Fable is more effective at spinning up sub-agents. I do find that Fable is better at spinning up sub-agents. So if I wanna complete a bunch of smaller tasks, I do tend towards using Fable 5. 我发现这非常有用，因此立即切换到 GPT-6 作为我执行编码任务的主要工具。目前，除了 Fable 在启动子智能体方面更有效之外，我没看到 Fable 5 在任何方面有优势。我确实发现 Fable 在启动子智能体方面表现更好。所以如果我想完成一堆较小的任务，我倾向于使用 Fable 5。

Refactoring and Code Optimization

重构与代码优化

Now the last test that I also did was starting to look for refactoring opportunities with the model. This is overall a very fun exercise, and I do that with every new model release. For example, when Fable released, when Opus 5 released, when GPT 5.6 Sol released, and so on. And my general prompt is something like: 最后，我进行的测试是开始利用该模型寻找重构机会。总的来说，这是一个非常有趣的练习，我会在每次新模型发布时都这样做。例如，当 Fable、Opus 5、GPT 5.6 Sol 等发布时。我通常使用的提示词如下：

Scan through the repo and look for refactoring opportunities anywhere where I’m not following good software engineering principles that will lead to higher chances of bugs or coding agents spending longer implementing code in my repository. This can be something like not following the principles of don’t repeat yourself, bad separation of concerns, or similar. Furthermore, I wanted to analyze the code base and see how it can increase code shipping speeds, i.e., I want to be able to ship code to dev faster and I want to see what we can do to optimize that. That could be optimizing how the repository is structured and the code in the repository, or simplifying it, or it could be optimizing the CI/CD pipeline. Provide me a full HTML report of your findings prioritized. 扫描代码库，寻找任何未遵循良好软件工程原则的重构机会，这些问题可能导致更高的 Bug 风险，或导致编码智能体在我的代码库中实现代码时花费更长时间。这可能包括未遵循“不要重复自己”（DRY）原则、关注点分离不当等。此外，我想分析代码库，看看如何提高代码交付速度，即我希望能够更快地将代码交付到开发环境，并希望看看我们能做些什么来优化这一点。这可能包括优化代码库的结构和代码、简化代码，或者优化 CI/CD 流水线。请提供一份按优先级排序的完整 HTML 报告。

I then have the agent work for however long it takes, and it comes back to me with a full HTML report. If you run this with models such as Opus 4.8, Opus 5, or GPT 5.6 Sol, it will come up with some okay feedback, but it won’t be really particularly good feedback that will help your code repository a lot. Yes, you definitely have to do the refactoring on a regular basis, but I would say that I noticed a very stark difference when doing that with these previous generations of models versus doing it with either Fable 5 or GPT-6. And in my experience, GPT-6 is also superior to both Fable versions when it comes to detecting opportunities to enhance my code repository. 然后让智能体工作，无论需要多久，它最终会给我一份完整的 HTML 报告。如果你用 Opus 4.8、Opus 5 或 GPT 5.6 Sol 等模型运行此任务，它会给出一些尚可的反馈，但不会是那种能对你的代码库有很大帮助的特别好的反馈。是的，你确实需要定期进行重构，但我注意到，使用这些前几代模型与使用 Fable 5 或 GPT-6 进行重构时，存在非常明显的差异。根据我的经验，在检测增强代码库的机会方面，GPT-6 也优于两个版本的 Fable。

I simply found that GPT-6 was better able to find issues that I had in my code. Not only did it find some stuff in my CI-CD pipeline that I could prune to make it faster and some parallelization opportunities, but it also discovered some limitations I had throughout my codebase that I wasn’t aware of. For example, how I was using certa… 我发现 GPT-6 能够更好地发现我代码中的问题。它不仅在我的 CI/CD 流水线中发现了一些可以精简以提高速度的内容以及一些并行化机会，还发现了我整个代码库中一些我未曾意识到的局限性。例如，我如何使用 certa…