OpenAI released GPT-6 Astra on September 3.
Greg Brockman, OpenAI’s president, called Astra a generational leap and said it could eventually be seen as the arrival of artificial general intelligence (AGI), which is an AI that can handle any job you’d give a capable person, instead of being very good at a handful of things. It’s the finish line the whole industry keeps pointing at, and nobody agrees on how we’d know we’d crossed it.
I did three things.
I watched people putting it through real work.
I read OpenAI’s own guidance on using this model.
I gave it a job with real money in it: the same one I handed Fable last week.
It surprised me in both directions.
In this guide, you’ll learn:
Where Astra lives
Where Astra wins and loses (and what else came out of it)
What happened when I gave Astra the job I gave Claude Fable last week
How to prompt GPT-6 Astra
Whether the $20 plan is enough for Astra
And whether you actually need Astra
A note for Claude users
I’ve spent most of this year inside Claude.
There’s a plain reason for that. For the first half of this year, Claude was the better tool and it wasn’t close. Cowork already existed, and ChatGPT had nothing like it. Their equivalent, ChatGPT Work, arrived months later. If you wanted to hand a computer a real job and walk away from it, there was one serious option.
That gap has narrowed. That’s why I’m writing this one. But this isn’t a switch-to-ChatGPT piece.
I never cancelled ChatGPT. I downgraded it to the $20 plan back in January and left it there, because two things on it stayed better than Claude’s: voice mode, which has never really been close, and image generation. Meanwhile I moved up to Claude’s $100 plan, and then the $200 one.
Here’s what I didn’t expect: the best thing I got from a week of testing Astra was a better way to use Claude.
So read this as a guide to getting more out of whichever one you already pay for, with a clear answer at the end about whether Astra is worth adding.
1. Where Astra lives
There are two places to switch it on.
In ChatGPT Work:
Download ChatGPT for desktop. Open ChatGPT and click Work
Click the model name at the bottom of the message box
Pick GPT-6 Astra
Drag the effort slider next to it. Everything in this guide ran at High
In Codex:
Click ChatGPT in the top left corner and pick Codex from the dropdown
Click Select effort at the bottom right of the message box
Pick GPT-6 Astra and set the level, again High here
That effort slider decides how hard the model thinks before it answers, and it costs you accordingly.
ChatGPT Work is like Cowork, and Codex is similar to Claude Code.
2. Where Astra wins and loses (and what else came out of it)
Nate Herk published eight videos about Astra in six days and put his own money through it.
Here are the best insights:
1. He ran 15 real jobs against Claude Fable 5.1, and Astra won 10: a presentation deck, a sales page, his taxes, an audit of years of subscription emails, meeting analysis, video edits, a small app, browser tasks, a website, a year of YouTube stats.
Here are the numbers:
Astra came in cheaper, but took longer. OpenAI’s own docs explain the cheaper half: Astra costs more per word than the model before it, and uses far fewer words to finish a job, so the bill per task comes out lower.
Treat those as one careful person’s numbers rather than a benchmark, though.
2. Astra won every job where it had to operate something: his taxes, an audit across two email accounts going back years, opening Canva, painting a picture with the mouse and uploading course videos into a community platform.
3. Fable won the jobs where taste mattered: the presentation deck came back better structured and more on-brand. The sales page answered the questions a real buyer actually has. The interface it designed was the one Nate preferred to look at.
4. It made an entire YouTube video from a single prompt: it found the footage, wrote the script, used his voice clone and his avatar, built the edit, added music and sound effects, and rendered the file. About 50 minutes.
5. The caveat: Nate says plainly that for most knowledge work, Astra is overkill and he reaches for the cheaper model. He also warns it will crawl through your weekly usage limit.
6. Brainstorm the failures before you automate anything: before letting it trade with real money, he had Astra list the ways the setup could break. What happens on the first run when there’s no previous record. What if a run crashed after placing an order. What if two runs overlap. What if the machine goes offline.
I could get more insights after doing my own test.
3. I gave Astra the job I gave Claude Fable last week
Last week I handed Claude Fable 5.1 a real job with money in it: replacing the spreadsheet one of my contributors uses to work out his share of paid subscriptions. His sheet had grown over two years and had workarounds in it I couldn’t verify.
This week I gave GPT-6 Astra the same spreadsheet and the same first prompt, word for word.
Same effort setting, and I rolled my own files back to the state Claude found them in, so neither model could see the other’s work.
Here’s what happened, and I’ve pulled out the part that applies to you rather than the part that’s only about me.
It read the file before it said anything. Then it interviewed me.
In my experience, the model that asks you one good question at the start is worth more than the model that’s cleverer and guesses. If you take one habit from this guide, make the thing interview you before it builds.
Then, while I was still answering, it carried on reading and started telling me what was wrong with the file: an old duplicate of the same sheet sitting inside it, repeated entries in one of the lists, formulas that could quietly change past totals.
It kept going. Astra found errors in my contributor’s spreadsheet that Fable never mentioned.
Nobody had looked at those details.
That’s the point. If you have a spreadsheet that decides money, payouts, invoices, commissions, hours, and you inherited it from someone or built it years ago, this is the first thing worth doing with one of these models. Before you ask it to build you something better, ask it what’s wrong with what you have.
Then I asked for a payout figure for every month, and it could only work from what I’d handed it.
That’s the difference. Claude went and fetched what it needed from my Substack account on its own. Astra never did. Not in ChatGPT Work, not in Codex, on my own machine with my own files sitting right there.
So I gave it the data, and it came back with a month-by-month table.
Fable and Astra landed about 5% apart on the total. Month by month, they barely matched.
That’s worth knowing if you ever check one of these against numbers you already have. The total tells you whether the method holds up. The breakdown tells you where it differs. Two sensible answers can agree on what you owe in a year and disagree about what you owe in March.
Both of them did the job, and did it well. Astra asked the better questions. Claude Fable was the only one that could go and get the answers.
If you’re not a paid subscriber yet, now’s the time to join hundreds of members with full access to all our guides :)
4. How to prompt GPT-6 Astra
I read OpenAI’s own guide for this model so you don’t have to.
Here’s how to prompt GPT-6 Astra for better writing, better instruction following, the right amount of self-checking, and more (it works with Claude too):
1. Give it a banned-words list: Here’s the prompt OpenAI recommends to stop its own model sounding like an AI