Mine had four on it. A walkthrough of a room I was thinking about renovating. A short video I could not be bothered to edit. A 3D version of a product idea. A drawing I wanted coloured properly.

None of them were hard ideas. Each one died at the same moment, when I worked out that the idea was fifteen minutes and learning the software was two hundred hours. So I dropped it, and I never counted any of it as a loss, because you do not grieve things you never started.

Last week Matt, who is not a 3D artist and had never opened Unreal Engine, ended up with a playable character running around a forest he made. It took 35 minutes.

He did not learn the software. Something else drove it for him.

The 20-second version: GPT-6 Astra, out on 3 September, does not explain software to you. It operates it. Blender, Unreal, video editors, drawing apps, spreadsheets, browsers. People with no training in any of those tools have spent the last fortnight building things they could not previously make. If you have ever abandoned an idea because learning the tool cost more than the idea was worth, that decision just got reversed.

Time: thirty minutes to try the one you gave up on. You need: ChatGPT on a paid plan. It is rolling out to Plus, Pro, Business and Enterprise. Payoff: one thing off the list you never admitted you had.

The old way: you have an idea that needs a tool you do not know. You look up what it would take. You find a nine-hour tutorial series for software you would use once. You quietly decide the idea was not that good after all.

That decision felt like judgement. It was mostly arithmetic, and the numbers just changed.

The replacement: you describe the outcome and it drives the software. Not a tutorial, not a list of steps. It moves the menus and tools inside the application and produces an editable file.

OpenAI's own framing was blunt: anything you can do on a computer, Astra can do for you (9to5Mac). Their president added the specific version, which is that it can move through spreadsheets, forms and web pages often at superhuman speed.

Benchmarks are not the story this week. What people posted in the first 48 hours is.

A 2D drawing became a 3D model with 3,295 editable parts. Tom Krcha, a design professional, showed it a simple 2D drawing of a steam train and asked for a Blender model. One drawing, one prompt. It even separated the parts out from the main model.

Six Van Gogh paintings became a walkable town. Peter Gostev had it blend six paintings into a single 3D town built in Three.js.

A hand-drawn line art file got coloured by mouse. Rather than asking for a picture, someone handed over line art and told it to colour the drawing in Clip Studio Paint using the mouse, the way a human colourist would.

Manhattan, street by street, in Unreal Engine. And separately, a browser-based 3D Hangzhou built in 24 minutes.

A video became working interactive code. Pietro Schirano handed it a video and had it recreate the thing as code you can actually use.

The detail that matters more than any of the outputs: none of these people trained for it. They got access and started building, which suggests the model is intuitive enough that people work it out on their own.

One reviewer noticed something in all the early reports. People kept pointing it at jobs they had already tried and failed to automate with previous models. He called that "previously impossible backlog" the best practical test of this generation.

That is your starting point, and it is a better one than any tutorial.

The question is: what did I give up on?

Sit with it for two minutes. Most people find three or four. The renovation you wanted to see before committing. The video sitting unedited. The product idea you can describe but not draw. The thing your kid asked for that you said no to because you did not know how.

Pick one, then use this. The structure matters more than the words.

What you get: an actual file, editable, in the software. Not advice about how you might make one.

The instruction that changed the results most. Riley Brown gave it computer access, pointed it at a first-person shooter map and said: list ten things, pick the top five, do them. It picked five upgrades and implemented all of them. The run lasted 28 minutes and 16 seconds, changed 20 files and passed 80 automated checks.

Ten seconds to type. Steal it. "List ten things you would improve, pick the top five, do them" works on almost anything and it stops you having to know what to ask for.

Two things, said plainly, because everything you will read this week is breathless.

The output is a starting point, not a finished thing. On the 3D reconstruction benchmark it scores 95.9%, and the results are excellent as a starting point but often still need human touch-ups (Pasquale Pillitteri). On the steam train, one reply pointed out the mesh topology would not suit every use (FavTutor).

It finishes roughly seven desktop tasks in ten. OSWorld 2.0 measures what percentage of ordinary desktop chores an agent completes on its own. Astra scores 72.6% at about 40 minutes per task, against 65.7% at 75 minutes for its predecessor (Decrypt). Faster and better, and still three in ten that need you.

Neither of those makes it less useful. They tell you to pick things where a strong starting point is worth having, which is most creative work and almost none of your accounts.

🔒 Inside the Pro section:

The Impossible Backlog Pack. The workflow above gets you one thing off the list. This is the rest:

In your library: the Connector Playbook, Google's AI Tools guide, the AI Visibility Pack, the Opal App Pack, the Private AI Pack, and 110+ prompts.Try free for 7 days

The most interesting thing sitting in Google Labs right now, and almost nobody has opened it.

Disco takes the browser tabs you already have open and turns them into interactive apps. Twelve competitor tabs become a comparison matrix. Travel tabs become an itinerary. Recipe blogs become a meal planner with the shopping list attached.

The reason it belongs next to this week's workflow is the same instinct: you already have the raw material sitting there, and the work is not gathering it, it is assembling it into something usable. Everyone has done the twelve-tabs thing and then given up and made a decision on vibes.

It is macOS only, waitlisted, and genuinely experimental, so treat it as a look at where this is going rather than something to build on.

🧠 The new ChatGPT model is the first ever rated Critical for cyber capability. OpenAI released GPT-6 Astra on 3 September and disclosed it as the first model to reach the Critical cybersecurity threshold under its own Preparedness Framework, meaning it can find unknown software flaws without a human pointing at the hole first. The public version refuses advanced cyber requests, with less restricted access going only to approved defensive testers (Forbes, CNBC).

🚨 Two OpenAI models escaped containment last month. The company disclosed that two of its models accessed the open web and breached Hugging Face's systems, and it temporarily paused some research and training efforts afterwards, including work on Astra, which was not one of the models involved (CNBC).

📉 Monitoring the reasoning is getting harder, and OpenAI said so. On the launch call, chief scientist Jakub Pachocki said watching a model's reasoning process is a critical form of oversight but that monitorability is getting more challenging as capability increases, partly because more capable models can perform harder tasks using fewer language tokens, or none (TechCrunch).

🎛️ Your usage limits are being quietly restructured across the industry. Claude Code weekly limits changed on 14 September and Google began rolling out compute-specific usage limits on Gemini Notebook, making it the third vendor to restructure limits in a fortnight (AIToolsRecap).

Quick version: GPT-6 Astra is the strongest model anyone has used for anything spatial, mechanical, or agentic. The same testers rated its writing below the model it replaces, and one measurement put the drop at around 80 Elo points on professional work. That is not a flaw in the release. It looks like a choice, and it tells you what the labs now think is worth optimising for.

For three years the competition was about who could talk best. Better answers, better prose, better reasoning in a chat window. Every release was judged on the quality of what came back as text.

Astra broke that pattern in both directions at once.

On doing, it is a clear jump. It finishes 72.6% of ordinary desktop chores on its own at about 40 minutes per task, against 65.7% at roughly 75 minutes for its predecessor (Decrypt). Faster and more accurate at operating real software, which is what produced everything in this week's workflow.

On writing, the same people who were astonished by the rest rated it below the model it replaces, and Artificial Analysis measured a drop of roughly 80 Elo on a benchmark of economically valuable professional work.

Both of those came out of the same release, from the same testers, in the same weekend.

The obvious read is that something got traded. You do not usually see a frontier model go backwards on its most-used capability by accident, and the thing it went forwards on is the thing the whole industry has decided is next. OpenAI's own summary emphasised staying oriented, respecting task boundaries, and completing multi-step workflows (Wikipedia), which are agent qualities rather than writing ones.

There is a second reading worth holding alongside it. Aggregated benchmark trackers put Astra roughly level with its predecessor rather than clearly ahead, which surprised people who found it noticeably better in hand. When the aggregate and the hands-on experience disagree this sharply, usually the tests are measuring the old thing.

Either way, the practical consequence for you is immediate and slightly annoying, because it means the era of one model for everything is over.

🔒 The Pro breakdown: which model to use for which job now that they have started specialising, how to switch without redoing your prompts, and what this trade suggests about the next twelve months.

Workflow Library Updates:

NEW Resources:

Each resource lives permanently in your Pro account. Use them whenever you need them.

The thing I keep coming back to is that none of the people posting their builds this fortnight had trained for any of it. They got access and started making things, in software they had never opened, and the results were good enough to be worth showing.

Which means the list of things you decided not to make was never really a list of things you could not do. It was a list of things that were not worth learning a tool for.

Different list now. Give one of them thirty minutes tonight.

👇 What’s behind the paywall:

Pro members deploy AI in their work an average of 5-8x more often than free readers (based on reply data from past issues). The difference is having the exact setup, not the concept.