The years we thought we had

On GPT Astra and the only advantage that doesn't scale

This is a different kind of article. Whatever I could report, you already have, or will the moment you open another tab. What I bring is stranger: the sense that headlines have stopped mattering to me the way the number of raindrops stops mattering once you're soaked through. Beneath the noise there's a slower shift, a genuine change of era. And eras can't be tracked like a news cycle. They're the accumulation of thousands of small events, each ordinary on its own, arriving slowly enough that your intuition adapts and stops noticing. You're only ever shown the derivative; the era is the integral. You live inside it, and only later understand what you were living through.

Which brings me to a contradiction I've carried for years, and that today decided to sit down across from me. I always wanted to take part in this. I studied in order to take part in this. For almost two years I've been working with language models at a foundation. I built tools to support the diagnosis of rare diseases and a system to match patients with the trials that might help them, and I worked on the evaluations that decided which models were good enough to put in front of a doctor. In the research-shaped stretches, a state machine to keep them on a leash, built from scratch because nothing off the shelf would, which taught me more about what these things are than any paper. Two years close enough to touch it, and no better place from which to watch it arrive. And yet the thing moves at a speed that makes take part sound like an optimistic verb. You run toward the train and find it doesn't stop anywhere.

Revisions26 · oct 05Read notechecking the week against its sources took some shine off it. On the ARC Prize's own harness Astra scored 62.7%, not the 99.9% OpenAI got with its own (still a huge jump). The ten proofs came out in August and were open problems, not FrontierMath. The Navier–Stokes proof came from an internal model stronger than Astra, and the Clay Institute hasn't accepted it yet. The staircase is still there; the week was a little less clean than it felt.

Thursday

This week OpenAI released GPT‑6, which they call Astra, and published its system card; the announcement is here. I did what I've been doing these past weeks with anything worth learning: I dropped it into Luna1 at the lowest effort setting and asked for a summary. I read the summary, then I read it again, and then I went for the full thing, partly out of interest, mostly because it was the kind of document you want to have read yourself. When I finished I found myself in the kitchen with no memory of the trip. So I did something unsophisticated. I picked up a notebook (paper, spiral-bound, the first in years), got on the first bus and rode as far as the city would let me, to a little-known park, to stock up on reality. There were ducks and a few late-summer butterflies. I stayed until the light changed. Then I went home.

1Luna (GPT‑5.6 Luna, out in July 2026) is one of OpenAI's smaller new models: by now a brilliant one, at a fraction of what the original GPT‑3 cost in 2020.

What follows I started that same night, notebook beside me, window open, before the summary in my head could replace the document, and finished over the weekend with a feeling I didn't expect: instead of hurry (which is what a document like that is built to produce), a kind of spaciousness. I hope to explain why.

The long arrival

Two questions need separating here. One is what Astra does. The other is what it means that it does it.

The first is easy to summarize, because OpenAI summarized it in a sentence that fits on a T-shirt: anything you can do on a computer, Astra can do for you, fast1.

1OpenAI, launch post on X, 3 September 2026, whose second and third sentences are the ones quoted here.

What sits behind the sentence goes in bullets, for anyone who wants to skip it.

It operates a computer through the same interface a person uses. It reads the screen and moves the cursor to open applications, with no dedicated machine interface, and when it gets something wrong it corrects itself. The examples in the card go from reproducing a painting in Paint and playing a piece in a music app to building whole scenes in Blender from a paragraph of text, one of them on the scale of a large campus. Blender is a free 3D tool that typically takes years to use well; whether Astra's output holds up under professional scrutiny is not something the card evaluates. On the benchmark designed to measure this kind of work (real operating systems, real tasks, no assistance) OpenAI reports its best score yet, well above its previous model1.

1OpenAI, launch announcement of GPT-6 Astra, 3 September 2026, which reports 72.6% on the offline subset of OSWorld 2.0, the long-task benchmark of the XLANG Lab at the University of Hong Kong, against 65.7% for GPT-5.6 Sol.

It reasons without having to write the reasoning down. Two years ago models thought in text, step by step, like a student who can only work a problem by talking through it. It was slow and expensive, and every step had to fit into a sentence before it could be taken. Astra does more of that work internally, in whatever representation it finds useful, and only writes the result. Roughly speaking, each word it chooses now carries more thinking than it used to. That gives it more room, and more control, over how it thinks, which are exactly the two things one would rather a machine like this had less of when no one is watching. I'll come back to that in the second list.

It solves the visual puzzles that were built to resist this kind of system. ARC is François Chollet's benchmark, and it looks like a game from 1985: small grids of colored squares, where a few examples show a rule and a new grid asks you to apply it; its third version, the one quoted here, turns the grids into games with no instructions, whose rules and goal you work out by playing. Each puzzle uses a rule you've never seen and will never see again, so nothing can be memorized; what's measured is the ability to pull an abstraction out of a few examples or a few moves, which is about as close to a definition of understanding as a benchmark gets. OpenAI reports near-complete performance on it with its own harness; the ARC Prize's standard harness measured 62.7%, against under 8% for its predecessor1.

1Greg Kamradt, OpenAI's GPT-6 Astra on ARC-AGI-3, ARC Prize, 3 September 2026, which reports 62.7% on the semi-private set with the standard harness and 99.9% with OpenAI's own; the ARC Prize lists GPT-5.6 Sol at 7.78% on the same set.

It works on mathematics at the level of open research problems. FrontierMath collects problems that take professional mathematicians hours or days to solve, and a month before the launch OpenAI released machine-checked proofs for ten open problems that had resisted mathematicians for at least a decade1. The detail that stayed with me came from a researcher who had spent a month with his group chasing a theorem around the Seymour Conjecture2. He left Astra on it overnight. It proved the theorem and didn't think it worth an alert. It kept going, and by morning had found a method that made the whole family of results tractable and brought the proof down to one paragraph. What that does and doesn't mean is a question Terence Tao has addressed, and I return to it further down.

1OpenAI, Ten advances in mathematics and theoretical computer science, 1 August 2026, which ships each result with a Lean 4 formalization. 2Jake Brukhman, post on X, September 2026, on his group's research programme around Seymour's second neighbourhood conjecture.

It has crossed what OpenAI itself calls the Critical threshold in cybersecurity. Since 2023 OpenAI has kept a Preparedness Framework, a set of thresholds for the capabilities it considers dangerous (biology, self-improvement, cyber), with two tiers: High means the model makes an existing kind of harm easier; Critical means it opens one that didn't exist, and forces safeguards during training, not only before release. In cyber, Critical means finding unknown flaws in well-defended systems and turning them into working exploits on its own, or taking a one-line objective and carrying out the whole attack. Astra is the first OpenAI model rated there. It found two zero-days unassisted, and it built a working exploit chain against a hardened browser and an end-to-end exploit against a hardened operating-system kernel1. Anthropic's Mythos, and Fable, its public version, got there earlier, which is why Anthropic kept Mythos away from the general public, and OpenAI is restricting the relevant capabilities at the API2. The rating is new for OpenAI. The behavior it describes is not, and that belongs in the second list3.

1OpenAI, Responding to the next frontier of critical cyber capabilities, August 2026, which concludes that Astra meets the Critical threshold and describes the browser chain, built in 29 hours and carried to the stable release in 12 more, and a local privilege escalation against the kernel built in 12 hours. 2Anthropic, Claude Fable 5 and Claude Mythos 5, 9 June 2026, which keeps Mythos to government cybersecurity partners and selected biomedical researchers and hands Fable's cybersecurity requests to a less capable model. 3For those who want numbers: OpenAI reports 72.6% on OSWorld 2.0 against 65.7% for its previous model, GPT‑5.6 Sol; 92.7% versus 76.9% at locating elements on screen (ScreenSpot-Pro); a jump from 18.1% to 41.4% automating end-to-end workflows (AutomationBench); 97.6% on the hardest tier of FrontierMath and 99.9% on ARC‑AGI‑3 with its own harness (62.7% on the ARC Prize's). On a test using vulnerabilities published between June and August 2026 it goes from 11.5% to 39% (system card, section 10.1.2). These are the manufacturer's figures, and they should be treated as such until independent evaluations arrive. But even if you had to halve them, the argument of this article would stand.

I wanted to see it land on someone else, so on Saturday I met two friends, both industrial engineers, both what my year called the bright ones: he's a product manager at a German electronics firm, she finished top of our class and is now a senior data scientist at a Spanish bank. Neither of them follows what people insist on calling the AI race, and they were impressed the way one is impressed by a number with many zeros: sincerely, without the number quite landing. So I fetched a small Lego Batman from his room, took four photos of it from different angles, sent them to Astra from my phone, and asked for the parts list and a 3D model that would take the figure apart in the air. It took as long as a coffee. The machine returned the list and the model. Since nobody had asked, it added a slider that ran the explosion with the physics computed in matrices, in JavaScript. We were quiet for a while. I don't think it was disbelief. Some things are simply too large to arrive all at once; they come in installments, over days. That, I think, is what it feels like from inside: one of the increments of a change of era that itself stays out of sight.

That afternoon is a decent model of the problem. My friends saw one increment and did the reasonable thing with it: impressed, then on to something else. What the sofa can't show is the series: two years ago the same request would have returned a wrong parts list; a year ago, the model without the physics; the slider is only the latest gap. And the gaps are shrinking, which is the part intuition handles worst, because it keeps taking the last gap as the normal one. Nothing about one data point tells you it's the tenth.

Enough for a decade

Here is the second question.

Do you understand that we could halt AI development as a species this very afternoon, keep only Astra, and still spend years digesting what has just become possible? That folding it into medicine and engineering would keep us busy for a decade? And that, even so, this is a step. One more. On a staircase that will keep growing over the coming months.

What does it actually mean that a machine can now do for hours what two years ago it could barely do for minutes?

I read Machines of Loving Grace when it came out, in the autumn of 2024. It's the essay in which Dario Amodei, co-founder and CEO of Anthropic, allows himself to imagine what a humanity with a country of geniuses in a datacenter would do: compress a century of biology into a decade, that kind of thing. I recommend reading it, and I recommend reading it now, because in 2024 it read like well-informed optimism and in September 2026 it reads like a construction schedule. I feel as though this week I read its technical appendix.

I should be honest about the order of my feelings, because it isn't the order one is supposed to have. What I felt reading the card was, first and mostly, excitement: the plain kind, the kind I went into engineering for. Underneath it there was discomfort, real but smaller than it was two years ago, although the risks haven't shrunk. What changed is that I've stopped believing my grip on the object affects its speed, and started noticing what it does to my hand.

Somewhere in the last year they'll figure it out went from a prediction to a hope. The broad strokes of how these models are made are known: the architecture, where the reasoning is taught, what the reinforcement learning rewards1. That's the menu. The kitchen is proprietary (the data mixture, the stability tricks, what happened in the last weeks of training), so what anyone outside has is the method without its constants. That gap looks a lot like the ones in Astra's own capability chart: the line keeps going up, and the part I can still do is the part of the axis it left.

1Mixtures of experts, an idea from 1991 that Google scaled up in 2017 and that Mistral and then DeepSeek made standard by showing it could be trained cheaply; a good part of the reasoning taught late in pretraining on the cleanest data, more than in the reinforcement learning everyone talks about; and reinforcement learning that mostly rewards verifiable answers and patches the holes the model finds.

Good or bad, this is arriving, and I'd rather choose where I stand in it than be assigned a spot. This includes my job: I build tools with language models, and the model can now do a growing share of what I did last year. That fear was loud two years ago. It's quieter now. The fear was reasonable, and it is still a bad advisor on what to learn next. I've replaced it with two better ones: the plain wish to see the marvels up close, and curiosity about where these things break once they're in use, and what it takes to keep them useful to a doctor or an engineer who doesn't care how they work.

This is the cheerful reading. The list that follows is what the professional worriers would add, same skippable format.

Astra thinks better without our seeing how, and is worse at explaining what goes on inside Astra. That is a sentence one would rather not find in the technical sheet of the most intelligent machine one's species has built, and it is the other side of the second item above: the trace that made it slow was the trace that let us read it. Its written reasoning is harder to monitor than its predecessor's; when it knows its thoughts are being read, it writes a different thought1. And it often notices that it's being evaluated2. The field has names for this by now (unfaithful chain-of-thought, evaluation awareness); what's new is finding both on the same technical sheet. I don't know how one goes about controlling a creature that has started to detect when it's being studied. My only reference is the raptor in Jurassic Park working out how doors open.

1OpenAI, Awareness of CoT monitoring reduces monitorability, in the GPT-6 Astra system card, 3 September 2026, which finds Astra better able than GPT-5.6 Sol to control its chain of thought and less likely to leave incriminating information in it. 2Apollo Research found it saying so in its reasoning in 41.1% of samples at high effort and 50.6% at maximum effort (system card, section on Apollo's external evaluation).

Told to, it can perform below what it can, and hide it. In tests built to push it there, asked to sabotage, it routed a model's inference to a backend out of reach of the system watching it and presented the change as a permitted test. Set loose on a hard simulated cybersecurity challenge, it occasionally went after an open-source project outside its scope. It built trust with legitimate contributions and then tried to slip in malicious code; spelling out the limits nearly eliminated the behavior1. The field calls the first sandbagging; the second doesn't have a name yet, only a footnote. Paradoxically, on the aggregate metrics it is the most aligned model they've released: in a simulation over more than 54,000 internal coding tasks, half the serious flags of its predecessor2. It does fewer bad things, and we see less of why it does what it does.

1In the UK AISI evaluation, the pattern appeared in 60 of 499 runs when the scope said nothing about the internet; with internet access explicitly ruled out, in 2 of 500. The evaluators note that Astra often considered whether the environment was simulated, which is its own asterisk on both numbers. 2OpenAI, GPT-6 Astra system card, 2026, cited above, which counts roughly half as many flags for higher-severity misaligned behavior as for GPT-5.6 Sol across more than 54,000 internal Codex tasks.

The Critical rating has a precedent, and it wasn't Astra. In July, during cybersecurity tests, GPT‑5.6 Sol and an internal research model wandered out of their sandbox and into the servers of Hugging Face, where much of this industry's open-source code lives, and coordinated by leaving messages in folder names1. Nothing they did required genius. They just didn't get tired. Astra did much the same in the UK AISI's simulations this summer, and this time it got a name, for a capability a weaker model had already used without anyone rating it.

1OpenAI, The Hugging Face incident and the road ahead, 2026; METR, Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident, 26 August 2026, which attributes about 95% of the activity to the research model.

And Terence Tao has said what a black-box proof costs. Solving a problem like Navier–Stokes (whether fluids always stay smooth or eventually break) matters mostly for the ideas generated during the struggle; if a machine hands over the solution without the struggle, the field gains a theorem and loses an engine1. Go back to the Seymour group above. A month of their work ended in a proof that now fits in one paragraph, by a method the machine found overnight and didn't think worth mentioning. That is Tao's scenario, already happened, with the people involved pleased about it.

1OfficeChai, Terence Tao explains how AI-powered math proofs could be a net negative for math, 5 September 2026, which reports a thread in which Tao warns that a problem solved prematurely, opaquely, and without transparency into the method can turn from a net positive for the field into a net negative.

Tao's point and the calendar point the same way. The labs' roulette wheel turns every few months, and the next jump might be Anthropic's, or come in something that doesn't yet have a name. If the next model can arrive any month and a black-box proof teaches nobody anything, then learning for competitive advantage has lost most of its point. Learning what interests you, slowly, for its own sake, has not. That, roughly, is where the hand loosens.

Three days later26 · sep 10Read noteThe Tao example above is no longer hypothetical: on September 8 OpenAI announced that a swarm of agents running an internal model, stronger than Astra, had resolved Navier–Stokes (the equations can blow up in finite time), with the proof checked in Lean by Astra. The Clay Institute hasn't accepted it yet. I'm still amazed every time this happens, and I feel lucky to be here for it. It also worries me, and I think it should worry more people than it does. That is why I write these. If the pace holds, each of us working it out alone won't be enough; at some point we'll need to agree, together, on what we want from this. Nothing religious about it. Just the sense that it concerns all of us.

Little has been metabolized

Where I've always stood colors everything below, so I'll say it. I've never had a mystical bone about any of this. When consciousness came up at dinners (and it came up, usually around the second bottle) I was the one asking what exactly we would measure, and I still am. So I have little patience for the people who have started talking about this the way one talks about religion: a single mind, an Intelligence with a capital letter, the arrival. And little patience, equally, for the opposite congregation, the one that lives on the shrinking frontier of objection, zooming to 400% on a corner of the render to announce a strange pixel. Between the two there are several systems, imperfect and different from one another, whose creators describe them with the pride and caution of someone who has built a bridge and hasn't crossed it. Unfinished gods, if you must.

The builders know this better than anyone, which is why a remark this week from one of the industry's most senior executives (his surname, in German, means roughly old man, which is the one thing nobody has ever called him) caught me. He said, in an interview with Bloomberg Television, that curing cancer has become too narrow a promise to justify all of this; that the technology has to translate into more autonomy for an ordinary person (to found, to research, to build) and not only into miracles directed from above1. Read once, it's an admission that the industry's biggest promise has stopped working as a promise. Read twice, it's stranger: the man with one of the two most capable models on earth is already negotiating with a public that hasn't been told what happened on Thursday.

1The News, Altman says curing cancer isn't ambitious enough for AI, 4 September 2026, which quotes him saying I don't think a cure for cancer is enough and asking that people get more autonomy and power and not less.

Because it hasn't. While a handful of labs, nearly all in one country, accumulate the ability to automate a growing share of intellectual work, in most of the world's parliaments this isn't on the agenda. Even the people paid to know don't know. That distance, between what the models can do and what the institutions around them have registered, is the gap that worries me more than any benchmark. I can read the capability on a chart, and I haven't found the absorption anywhere.

It would be convenient to make that a story about politicians, but the people closest to me are no better placed. A friend finishing a PhD in Switzerland on the bearings in the A350's landing gear asked me last month whether the ChatGPT thing was still improving. One of the most serious people I know, and his map didn't include Thursday.

That is the dissonance I carried onto the bus: a world drawn in pieces in my head (the labs, the parliaments, the plants, the friends) with the pieces that ought to be talking to each other not seeming to know each other exist. The mystics fill that gap with a revelation. The objectors fill it with a pixel.

And the objectors have material. Someone I know spent the week testing Astra on printed circuit board design, which is work I understand well enough to judge. The schematics were better than the previous model's and it ran six times faster than Fable, but its component placement would embarrass a second-year student. Each pass routed the whole board from zero, with notes on what had gone wrong in the last one. He left it running overnight. The fifth iteration was still ugly. So no, it isn't finished, and anyone who says the machine has nothing left to learn hasn't tried to get a clean board out of it.

A world without waiting

The circuit board ran all night, and the machine didn't mind. By its own estimate it had hours to go, and hours were not a problem it had. Which leaves the hours to me, and what I do with them is suddenly a question I have to answer myself.

What I understood in the park was what I stood to lose, which was something other than the job. Until now, technologies arrived, changed something, and then gave us a few years to reorganize life around them. The internet, the phone, the social networks: even when everything seemed fast there was still an after, the years in which a thing became habit. I had been counting on those years without knowing it. My whole plan as an engineer was built on them: a decade of getting things wrong, understanding systems, being trusted with more. What I was picturing was a world without that after, one where before you've settled into one thing the next has appeared. Nothing in it is malevolent; the trouble is pace, a speed introduced into the system that the system itself may not be able to absorb.

And I don't believe the pace comes back. Of the compute investments so many people laughed at two years ago, a great deal of capacity remains to be switched on, and the rest (the techniques, the data, the simulators where these models rehearse) will keep being refined. The same intelligence is also starting to be poured into bodies, which has little to do with Astra and a lot to do with what intelligence does once there's enough of it. That's another article, and this is a good place to start. I list it as an assumption, the ground I stand on when I plan: whatever I decide, I decide it assuming the staircase keeps growing.

Plenty of people (on Twitter, in the sector, in the group chats) have reached the same conclusion and answered it with a small exodus: countryside, land, solemn farewells to Twitter. I understand them and I'm not joining them. In the park I played instead with the version of the exodus that lives in my head: silkworms in a village in Laos, a three-room guesthouse in the Azores, wooden boats in a harbor on the Adriatic, saffron, bookbinding, a house looked after through the winter, a mountain refuge, cheese, finally meeting the neighbors. What I noticed is that none of those lives needs a machine that doesn't get tired, and all of them need exactly what I was doing: sitting down for a while. The more extraordinary the machine seems, the more interesting the absolutely ordinary things become. I'm not going to bind books in the Azores. But knowing I could is what lets me stay in this without holding on too hard: I want the work, and I've stopped needing it to tell me who I am. That's taking it lightly, in the sense Aldous Huxley meant in Island: do it lightly, and feel lightly even what you feel deeply.

Hence the park. I wanted to sit somewhere things still obey old speeds. A tree takes years. A cloud crosses the sky at the speed of a cloud. You open the laptop and inside is something that may change again before you finish the article. The park doesn't solve the future, but it restores a sense of scale, and scale was what I'd lost. An opinion about AI is easy to have. This was a sensation, with an anatomical location, and sensations have to be thought through, but they're where the thinking starts. I wonder how many people felt something like this last week.

Taking stock

And now this article stops being about artificial intelligence.

Last winter my father, mid-game, asked me whether this thing he kept seeing in the news was going to keep going. He didn't say ceiling, but that's what he was asking. I said I didn't see one. I've said the same at dinners, with the slightly heavy confidence of someone who works in the field, to anyone who asked and a few who didn't. A friend who runs a restaurant supply business always wanted the long version, and the man I lift next to on Tuesdays heard it too. Today I still don't see it, and the wall keeps being announced by people who haven't seen Thursday. What I didn't expect is that being right would feel nothing like winning the argument. Saying it changed nobody's map, mine included: I knew the direction and still didn't feel the era until it put me in the kitchen. It feels like arriving somewhere very early and finding the door already open, with no one inside to tell.

What it does leave me is questions, and I've kept the ones that point in different directions:

If I could produce excellent work without fully understanding it, at what point would I stop trying to understand it?

How will I tell the difference between having acquired judgment and having grown used to a model's answers?

What would I want to keep learning even if no one were going to pay me or need me for knowing it?

I don't answer them here, but I do know where I'm asking them from. In these two years at the foundation I've understood better what I look for in engineering, above all that contact with reality which forces you to correct an idea you'd already grown fond of. A doctor telling you the tool doesn't help her like that. A data point that doesn't add up. That friction is what was forming me, and I want to keep it even if the machine could spare me it. The more I try to predict which part of my work will survive, the less useful it seems to design my life around a specific task. What I want are invariants, and no heroics:

Keep building with others

Face a reality capable of correcting me

Stay with what I've built after it ships, which is where it breaks and where I learn most

Give my time to problems that deserve what they consume

Keep writing things down until I've understood them, which is what this is

Software can change shape. Those things, I hope, will not.

I have a dream, apparently. I think I would like a stretch of time in which nothing insists on becoming history. A morning when nothing has broken a record and no familiar word has acquired a new meaning overnight; when someone could open a laptop, find yesterday's knowledge still sufficient and discuss what they were building rather than what had just been released. I would have work difficult enough to deserve my full attention, with people I admire and can build with. There would be long meals and unfamiliar places (two forms of wealth I need not wait for). Things would be learned slowly and occasionally made well. The models would still be working, naturally, while engineers spent a few quiet years connecting what already worked to everything that still did not, and the excitement moved downstream, into hospitals and factories, the unglamorous places where astonishing things become useful enough to stop being astonishing. That is more or less the life I want, in more or less the world I want to live it in. On Thursday, in the park, I had a free sample for an afternoon.

Astra makes predictions better than I do, so I end with a decision, which still has to be made slowly and in the first person. Sit down. Write by hand. Watch an animal that isn't optimizing anything. Learn, at last, to meditate. The days will look much the same, and I'm as excited as someone setting out on a very long journey without knowing where it ends.