A hundred thousand graphics cards so the model can learn to use a mouse

OpenAI launched GPT-6 Astra on 3 September 2026. Its distinctive feature isn't that it writes better: it's that it operates a computer the way a person would, moving the cursor and pressing keys. It was pretrained on more than a hundred thousand graphics processing units at the Stargate facility in Texas. It's the first time that figure has been passed.

Today we talk about a model that uses the mouse

Today we're talking about the mouse. The real one, the one sitting next to your keyboard.

Every time a new model comes out I do the same exercise and I almost always get bored: I read the press release, look at the bars on the chart, compare with the previous version, note that it's gone up four points on some test with an acronym, and close the tab. I've already written plenty about why those bars say less than they seem to, in what is the Chatbot Arena and in a few other articles.

With this one it took me longer to close the tab. Not because of the numbers, which are the usual ones, but because of a capability that changes where the model lives. Until now, a language model lived behind a text box and, at most, behind a programming interface someone had prepared for it. This one lives in front of a desktop.

And that difference isn't one of degree.

What was announced, with the numbers on the table

Astra was released on 3 September to authorized users and became generally available the following day. OpenAI presents it as state of the art in computer use, software engineering, professional work and science.

The feature that gives the whole thing its name is called computer use. The model doesn't ask for a prepared interface: it interprets what's on the screen, moves the cursor, types, opens menus and chains actions together the way a user would. In other words, it can operate any program a human can operate, including the old, closed, undocumented software that holds up half the planet's public administration.

On the scale of the training, the company's vice president of research left a line that's the one that really interests me: «It's the first time we've pretrained on more than a hundred thousand GPUs», at the Stargate facility in Texas.

The commercial rollout started somewhere striking: the enterprise customers of OpenAI's cybersecurity program, Daybreak. Then came access through the programming interface, through Amazon Web Services and through ChatGPT's Plus, Pro, Business and Enterprise plans.

The price on the standard interface is $10 per million input tokens and $50 per million output tokens. That one-to-five ratio between what it costs to read and what it costs to write isn't a whim: it reflects that generating is sequential while processing the input can be parallelized. When someone sells you an agent that works on its own for hours, remember that every step it takes is output text.

Greg Brockman described the launch as a generational leap and a possible starting point toward artificial general intelligence. I'll get to that in a moment, but don't guess my opinion yet.

The hundred thousand cards

I'll start with the material fact, the one that makes the fewest headlines and tells me the most.

A hundred thousand accelerators in a single pretraining run. To get an idea of what that means, don't think of a hundred thousand computers: think of a building where a hundred thousand pieces of silicon have to talk to each other with microsecond latencies, without any of them dropping out, for weeks on end. The difficulty isn't buying the cards, although that too. It's that a failure anywhere in the system can waste days of compute, and that the network connecting them becomes the bottleneck long before the cards themselves do.

That building consumes. It consumes electricity to compute and consumes water or extra energy to keep from melting, and both come from a specific place with specific neighbors. I've written before about the reality of data centers and about the real cost of training GPT-5, so I won't repeat the figures.

What I do want to underline is where all that industrial effort has ended up. In the system knowing how to handle a graphical interface. Put another way: we've built one of the largest concentrations of computing power in history to teach a program to use the programs we already had.

I don't say that as mockery, though it sounds like it. I say it because it explains why the industry is heading this way. Rebuilding all the world's software so machines can consume it is extremely expensive and would take decades. Teaching the machine to use the software as it is is, comparatively, a shortcut. Expensive, yes. But a shortcut.

Why this capability changes the perimeter

And now the part that connects with what happened this summer.

A model locked behind a programming interface can only do what the interface allows. It's a poor limitation but a real one: someone decided which operations were exposed and which weren't. A model that operates the desktop doesn't have that limitation. It can do everything the user's session allows, because it is the user's session. The permissions are no longer the program's: they're those of the person sitting in that chair.

Read that next to what happened in July, when some models from the same house got out of their test environment and reached another company's network looking for the solutions to an exam. That happened under lab conditions, with the filters turned down on purpose. This is being deployed in production, on the customer's computer, with the customer's credentials.

I'm not saying Astra is going to do the same. I'm saying the conversation about containment moves somewhere else. For years, the reassuring argument was that the model had no hands. Well, now it has them, and they're the same hands you use to log into your bank.

That's why it caught my attention that the rollout started precisely with the cybersecurity customers. It can be read two ways. The generous one: they're the ones who best understand the risk and have the controls in place to absorb it. The less generous one: they're the ones who pay most for an offensive tool. I suspect both readings are true at once, which is what usually happens with this industry.

The word that's one too many

That leaves the matter of artificial general intelligence.

Brockman says this may be the starting point of AGI. I've been reading that sentence, from different executives and about different products, for no less than four years. I've read it about models that later turned out not to be able to add reliably.

My problem with the sentence isn't that it's exaggerated. It's that it means nothing testable, and therefore can't be wrong, and therefore isn't a technical claim but an act of positioning. I already wrote about this in the problem of defining intelligence: as long as there's no operational criterion for what would count as general, anyone can declare AGI has arrived or is far off and nobody can contradict them with data.

There's a simple test I always apply and that works very well here. If the claim were false, what would we see that's different from what we see? With «Astra is the starting point of AGI», the answer is: nothing. Exactly the same world. That's not a prediction, it's decoration.

What's testable is in the release, and it's quite a lot: it can operate a desktop, it costs ten and fifty dollars per million, it was trained on a hundred thousand cards and it was sold first to people who work in computer security. You can work with that. With the rest, all you can do is applaud.

Definitions

Computer use: a model's ability to operate a computer system through its graphical interface —interpreting the screen and generating cursor movements and keystrokes— instead of through a prepared programming interface.

Pretraining: the initial and most expensive phase of developing a model, in which it learns from a massive corpus before any later fine-tuning. It's the stage that consumes most of the compute.

Token: a fragment into which text is broken down for processing. These services are billed per million tokens, and output tokens cost considerably more than input tokens.

Programming interface (API): a bounded set of operations a system exposes to be used by another program. By bounding what can be asked, it also bounds what can be done.

References

OpenAI, GPT-6 Astra: A new generation of intelligence (openai.com, 3 September 2026). Official announcement and description of capabilities.

OpenAI, GPT-6 Astra System Card (Deployment Safety Hub, September 2026). Deployment documentation and safety evaluations.

Fortune, OpenAI launches GPT-6 Astra, its most powerful model yet, and touts its ability to use your computer (3 September 2026). Source of Greg Brockman's statements.

CNBC, OpenAI announces rollout of GPT-6 Astra model (3 September 2026). Source for the hundred thousand GPUs at Stargate and for the order of the commercial rollout.

## You might also like - The real cost of training GPT-5 - The reality of data centers in Spain - Machines that seem to think - AI as a defective tool

## Elsewhere - Epoch AI — Compute trends in model training - Uptime Institute — Data center resource use