GPT-6 Astra Review: The First Model That Can Use a Computer Better Than I Can

The last time I reviewed an OpenAI model, I was pretty lukewarm about it. Then things got considerably worse.

After leaving it alone for a few days, I decided to give it another shot. It thanked me by accidentally deleting almost every file on my Mac, including corporate documents. I was, to put it mildly, not thrilled. This was the kind of screwup I might have expected from an experimental AI in 2023, not a supposedly state-of-the-art model released this year.

It was going to take a hell of a model to win me back. Somehow, OpenAI has done it.

So when I tell you I’m now letting its successor work inside my email accounts and help run my business, please understand that it had a fairly substantial hole to dig itself out of.

I’ve been testing OpenAI’s new model, GPT-6 Astra, and I’m trusting it with work I wouldn’t have been comfortable handing over to an AI before. One experience in particular surprised me.

It covered for my newsletter team

One of the people who helps me run my newsletter was out one day, so I asked Astra to cover for him.

Normally, I do the writing, Astra prepares the post in beehiiv (the platform I run my newsletter from), and he handles the strategy around who receives it and when. He’s an expert at this. We’re sending from a relatively new email domain, which makes choosing the right audience and timing especially important.

I also have a frankly stupid number of drafts sitting around. Astra sorted through them, found the right post, looked at previous emails and their timing, and used what had worked before to figure out how to handle this one. I expected it to struggle. This was the part of the job I normally left to someone who knew considerably more about it than I did.

I sent him its work for a quick check. He told me it was perfect. Good to go.

Most days, Astra’s newsletter work is less dramatic. It gets the post cleaned up, makes the presentation match my templates, and handles the preparation around each send. But that’s exactly why it’s useful. I do the writing, and I have something dependable handling a substantial part of the work that comes afterward.

I’ve had a similar experience with advertising. I still work with an agency on strategy and optimization, but Astra handled everything from setting up my Meta Ads account to connecting it to my website, tracking how people move through it, and working on optimization. The agency later told me the account had received the best score they’d ever seen on their internal assessment.

They knew I was using AI. What confused them was how I’d gotten it to do the work this well. At the time, I couldn’t share anything about the model. The answer, though, was pretty simple: I’d asked Astra to do it, and it had done an exceptional job.

It can work while you do something else

One of the reasons computer-using AI has been frustrating is that you could end up watching it do a task you could have finished yourself. Sometimes it was slow. Sometimes you were waiting for it to make a mistake. Either way, the job was still taking up your attention.

The ChatGPT app already let earlier models use supported Mac apps in the background. With Astra, I’m finally comfortable leaving it to get on with the work. Once I’ve given it the necessary permissions, it can work while I keep doing something else. If I go to the app it’s using, I can see the blue ChatGPT cursor moving around. It’s a pretty strange thing to watch the first few times. Then you get used to it and go back to what you were doing.

This is incredibly useful because I no longer have to supervise it like someone else using my software. I literally just hand over the task and get on with my day.

There is a practical distinction here: supported Mac tasks can run in the background, while Windows computer use currently operates on the active desktop. So your experience will depend on your device and the tools available to it. Computer-use details

ChatGPT Work also offers a cloud option where available, which can keep working without your computer staying on. I haven’t tested Astra in that environment, so I can’t tell you how this particular model performs there (but I would venture to guess that it’s downright exceptional). I’d highly recommend considering this if you want the work happening somewhere separate from your own computer. Local and cloud work

The important part is giving the model a way to act. These examples involve access to software, files, or accounts. If you only use it to answer questions in a regular chat, you won’t experience the part I’m most excited about.

Keep the same conversation going

I didn’t have to teach Astra much before it became useful. With Gmail, I didn’t want to spend the time setting up another connection or plugin, so I simply had it log into my Gmail accounts through my computer. It figured out how to navigate them, did the work, and even marked messages unread again after looking at them so my inbox stayed in a state I could understand.

For things that are particular to you, just tell it. Maybe you want certain emails labeled a particular way, or you have a preferred format for an update. Keep using the same conversation for that responsibility, and it can carry those preferences forward. I’m finding that much more useful than starting a fresh chat every time I have another job for it.

Think of a conversation as the place where a particular kind of work lives: one for your newsletter, another for your inbox, one for outbound sales, another for tracking your marketing. You can correct it, give it examples, and come back the next day without explaining everything again. You can also schedule recurring work. (Just ask the AI to schedule it for you. For example: “Can you repeat this every day at 9 a.m.?”)

Let it earn your trust

Given what happened to my Mac, I started carefully. I used more restrictive approval settings, watched what it was doing, and gradually gave it more freedom as it proved itself. That’s still how I’d suggest approaching it. Start with limited access to the things it needs, and have it ask before sending or changing anything important. I wrote more about how I build that trust here.

I’m now comfortable letting it send emails and do high-stakes work on my computer. The biggest mistake I can recall in the email work I’ve given it was using double line breaks instead of normal ones. Compared with the kinds of things I used to worry about when letting an AI operate my computer, that’s a pretty welcome problem to have.

At one point, I was curious and asked it to see if it could find tax savings my accountant had missed. It found the relevant documents on my computer, researched the rules online, and worked for a couple of hours until it identified potential savings my accountant hadn’t raised. You can’t tell me that’s not useful :)

I still work with human experts, as the newsletter and advertising examples make clear. And I still do my own writing; getting Astra to consistently absorb my writing preferences has been less successful. You’ll find your own boundaries. The surprising thing for me is how much work now falls comfortably inside them.

This is worth trying on your actual work

In Something Big Is Happening, I wrote about the changes programmers were experiencing with AI and why I thought they were coming to other kinds of work. Using Astra has made that feel much less like a prediction. It’s real now.

You don’t need to write code to get something valuable out of this. Pick a task you understand well enough to judge, give it the access and instructions it needs, and see how much it can take off your plate. Watch the first few attempts. Give it feedback. Keep the same conversation going when it works.

For me, that’s already meant handing over parts of running a real business. I’m still deciding what I want to say in the newsletter. I’m still working with experts where their judgment matters. But a growing amount of the work around those decisions is getting done without me having to personally push it through every piece of software. That’s a pretty damn meaningful change to my working day.

Know someone who tried AI a while ago and decided it wasn’t useful for their work? Send them this. I’ve been using AI for coding for a long time, but I’m now handing over work I wouldn’t have trusted it with before. They might be surprised by how much has changed.