GPT-6 Astra was lately launched. I received entry to it on Friday night, European time, and have been utilizing it extensively ever since. On this article, I will share my first impressions of the mannequin and the methods that I make the most of to get as a lot as potential out of the mannequin to maximise its effectiveness.
I will additionally share a few of the downsides I’ve skilled with the mannequin and the way I am making an attempt to take care of them within the coming weeks, after I’ll be utilizing the mannequin increasingly, and we’ll be attempting to maximise how productive I might be with the mannequin.
This infographic highlights the principle contents of this text. I will talk about the right way to get essentially the most out of GPT-6 Astra and my first experiences with the mannequin. Picture by ChatGPT.
Why use GPT-6 Astra
Firstly, I all the time prefer to cowl why you must care in regards to the matter of an article. On this case, it is as a result of GPT-6 is the newest launch from OpenAI, which is among the frontier labs. This naturally makes it one of many LMs that you simply instantly need to check out to see how properly it performs on my workflows. My workflows largely encompass coding duties, although I do have another duties comparable to utilizing my pc, transferring round within the browser, and doing analysis, so measuring how properly it really works for deep analysis and so forth.
I exploit coding brokers primarily every single day, and thus I consider I’ve an excellent baseline to match new mannequin releases in opposition to. GPT-6 can also be a really hyped launch, and it is a new main launch from a frontier lab, which is after all extraordinarily thrilling.
My first impressions of GPT-6 Astra
To start with, let’s cowl my first impressions of the mannequin. Instantly, as soon as I began utilizing the mannequin, I began testing a number of issues:
-
Operating it on duties that I’ve performed earlier than
-
Operating it on some new duties (characteristic implementations and bug fixes)
-
Beginning to search for refactoring alternatives and, on the whole, alternatives to enhance my code repositories.
My first impressions will thus be based mostly on my expertise when working these three duties. I wish to observe, nevertheless, that first impressions won’t give one of the best image of how good the coding mannequin is. The explanation I say that is that I bear in mind when GPT 5.6 Sol was launched, I received extraordinarily good first impressions of the mannequin, and I nonetheless would say it is an excellent mannequin. Nevertheless, over time, I began to note some quirks with the mannequin the place it did not work as properly, and these quirks have been one thing I did not uncover for every week or two by actively utilizing the mannequin. Thus, I would not have found such quirks but on the subject of GPT 6.
Nonetheless, I will give my assessment. So on the subject of working duties that I’ve performed earlier than and verified, or that I verified that each Claude Code and GPT-5.6 Sol have been in a position to do, I’d say GPT-6 was, after all, in a position to do all of them. However one factor I truly observed in favor of GPT-6 is that it achieved it manner sooner, and I do not assume this has something to do with inference pace. In my expertise, this simply looks as if the mannequin is simpler at using its tokens and is extra in a position to full duties shortly.
This impression was additional verified as soon as I began to run the mannequin on new duties as properly, each discovering bug fixes and have implementations. It simply looks as if the mannequin was in a position to full duties tremendous shortly whereas additionally doing it appropriately, at the very least when evaluating it to each the earlier technology of OpenAI fashions and when evaluating it in opposition to Claude Fable 5, which is the opposite frontier mannequin obtainable in the marketplace.
I discovered this very helpful and thus instantly switched to GPT-6 as my principal driver to carry out coding duties. Presently, I do not see any space wherein Fable 5 is superior, aside from the truth that Fable is simpler at spinning up sub-agents. I do discover that Fable is best at spinning up sub-agents. So if I wanna full a bunch of smaller duties, I do have a tendency in direction of utilizing Fable 5.1 nonetheless.
···
Now the final take a look at that I additionally did was beginning to search for refactoring alternatives with the mannequin. That is general a really enjoyable train, and I try this with each new mannequin launch. For instance, when Fable launched, when Opus 5 launched, when GPT 5.6 Sol launched, and so forth. And my normal immediate is one thing like:
I then have the agent work for nevertheless lengthy it takes, and it comes again to me with a full HTML report. For those who run this with fashions comparable to Opus 4.8, Opus 5, or GPT 5.6 Sol, it can give you some okay suggestions, but it surely will not be actually significantly good suggestions that may assist your code repository lots. Sure, you positively need to do the refactoring frequently, however I’d say that I observed a really stark distinction when doing that with these earlier generations of fashions versus doing it with both Fable 5 or GPT-6. And in my expertise, GPT-6 can also be superior to each Fable variations on the subject of detecting alternatives to reinforce my code repository.
I merely discovered that GPT-6 was higher capable of finding points that I had in my code. Not solely did it discover some stuff in my CI-CD pipeline that I might prune to make it sooner and a few parallelization alternatives, but it surely additionally found some limitations I had all through my codebase that I wasn’t conscious of. For instance, how I used to be utilizing sure packages, some UI points I wasn’t conscious of, and on the whole, I simply discovered it stronger at detecting such points. Thus GPT-6 would positively be my principal driver to carry out refactoring going ahead.
I wish to observe, nevertheless, that although I do discover GPT-6 superior to Fable, it isn’t far superior to Fable; it is one step above, however I’d say the distinction between GPT-6 and Fable versus the earlier generational fashions (Opus and GPT-5.6) is an even bigger distinction than the distinction between Fable and GPT-6.
···
Lastly, I simply need to point out one draw back that I’ve observed with GPT-6 when utilizing it now, and that’s that it does appear to ask a bit too typically for permission to do stuff, when in actuality I simply need it to proceed working till it is fully performed with the duty. Now, I’m conscious that this may very well be a problem with the prompts that I am utilizing and the markdown recordsdata in my repository, so I’m working to optimize this, however this is a matter that I’ve observed with GPT-6, the place it isn’t as robust as I’d have hoped for.
The most effective coding agent, for my part, is the one which merely takes the duty you give it, instantly asks any clarifying questions or issues that it must know earlier than beginning to implement, after which it solely comes again to you as soon as it is fully completed with the implementation. Now, after all, I do understand that in some instances the coding agent has to ask you clarifying questions due to some issues that come up throughout implementation. However I do assume that GPT-6 tends to do that a bit an excessive amount of, which may very well be a problem with the mannequin not performing due diligence sufficient earlier than beginning an implementation or being too uncertain about itself when truly performing the implementation. That is positively one thing I will take into account going ahead as I attempt to optimize, and I will come again at a later level with my up to date opinion on this difficulty with GPT-6.
General, my first impressions are superb, and I’ve turned to GPT-6 for my principal driver for coding, solely returning to Fable on the subject of spinning up sub-agent-heavy classes or, after all, after I need to assessment code with a separate coding agent the place it’s helpful to run it with a distinct coding agent than the coding agent that was implementing the code.
Easy methods to get essentially the most out of GPT-6 Astra
Now, let’s begin speaking about the right way to get essentially the most out of GPT-6 Astra. The very first thing I’d instantly do if I have been you is to begin performing refactoring in your repository. There’s seemingly a bunch of issues that you are able to do along with your code to make it simpler for coding brokers to carry out fixes there sooner or later, whether or not it is GPT-6 Astra working in your repository or a earlier technology of fashions. Thus, I urge you to begin doing refactoring instantly, for instance, utilizing the immediate I listed above.
This may most likely enhance coding iteration pace, scale back the possibility of bugs in your repository, and general enhance your coding productiveness.
Subsequent, I’d most likely transfer to beginning to take a look at GPT-6 for brand new coding implementations. In my expertise, it is superb at getting stuff performed, although, after all, it does have the one draw back I discussed earlier, the place it does ask for permissions a bit an excessive amount of. I urge you to attempt to make clear as many questions as potential beforehand and make it as clear as potential to the agent what it has to implement, the issues it has to contemplate in the course of the implementation, and make it clear what permissions you are giving to the agent so it simply does not need to ask for too many permissions.
One factor to contemplate when utilizing GPT-6 Astra is that, after all, it has restricted utilization. Particularly, as a result of they do not have 5-hour utilization, you might be spending a bunch of your weekly allowed utilization in someday, and I’ve discovered that I have been in a position to spend at the very least 1.5 weekly usages in someday by using resets.
That is as a result of Codex gives utilization resets. They have been giving out lots within the final week at the very least, so I had 3 obtainable on one among my subscriptions.
Nevertheless, I am not essentially positive that switching to GPT-5.6 Sol will make this difficulty any higher. The reason being that I discover that GPT-6 is extra token-efficient, and the precise price for implementing a job won’t be that totally different for GPT-5.6 versus GPT-6, which is unquestionably value holding in thoughts if you wish to attempt to save tokens. As a substitute, if you wish to save tokens, I urge you to consider the enter tokens you feed the model- for instance, wanting by means of your MD recordsdata, any MCP instruments that you simply’re loading into reminiscence, and so on- and attempt to restrict these as a lot as potential so you do not take up too most of the enter tokens the mannequin has obtainable.
I additionally assume it is value being conscious that the mannequin has round a 260,000-token window, which is smaller than the token window you are allowed to make use of on Fable or Opus with Claude Code, the place you are allowed 1 million tokens. This has each execs and cons. The draw back of a smaller configured context window is that the agent might have to compact extra typically throughout long-running duties. The advantage of this method, nevertheless, is that the mannequin will seemingly be sooner than for those who allowed a 1 million-token context window as a result of extra enter tokens imply the mannequin is slower to reply. And for my part, high quality goes up in case you have fewer enter tokens.
Codex gives a 1 million-token context window setting you can allow, although I don’t advocate utilizing it as a result of it isn’t advisable by the folks creating GPT-6, and thus you seemingly should not be utilizing it your self both. Each due to efficiency and due to the way it will affect your utilization limits.
General, when using GPT-6, I urge you to consider the utilization that you’ve with it. Begin by refactoring some repositories; it will make the agent simpler when working within the repository, and it’ll make some other coding agent that you simply use sooner or later extra environment friendly when performing implementations. I then urge you to only attempt it out on duties, offering it with as a lot data beforehand as potential as a result of the mannequin does are likely to ask you for permission a bit an excessive amount of somewhat than a bit too little, based mostly on my first expertise.
Conclusion
On this article, I mentioned my first expertise with GPT-6 Astra, which is the newest mannequin launched from a frontier lab and the brand new mannequin household mainly of OpenAI. I used to be extraordinarily impressed after my first expertise with GPT-6 Astra, which I have been utilizing all through the entire weekend. The mannequin is extraordinarily efficient at finishing duties, and it looks as if it is simply extra in a position to full duties appropriately in a shorter period of time than any earlier coding agent that I’ve used, whether or not it’s OpenAI’s code brokers or the frontier fashions supplied by Anthropic with Claude Code.
To get essentially the most out of the mannequin, you must begin by reviewing your repository for any refactoring alternatives. After which, earlier than you progress on to begin implementations comparable to bug fixing or new options, you must make sure that the duty and permissions the mannequin is given are very clear beforehand to keep away from any pointless interruptions by the mannequin.
👉 My free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day E-mail Course)
📚 Get my free Imaginative and prescient Language Fashions book
💻 My webinar on Imaginative and prescient Language Fashions
👉 Discover me on socials:
💌 Substack
