Last time I ended on a little bit of cliffhanger. I said the next tests were close: quarterly jobs data, the first audit of the numbers the AI labs report about themselves, and whether the slowdown the labs had just asked for would actually get built. A few of those tests have landed, and one big thing arrived that I didn’t think would happen quite so fast.
The last few weeks looked like this. The rogue-agent panic cooled into something closer a little closer to a normal risk. The jobs data came in and stubbornly refused to settle the argument - either way. And the technology itself pulled apart into two different directions - getting cheap and ordinary at one end, and jumping sharply ahead at the other. That last split is the more significant thing.
Let’s walk through them in order.
The jobs numbers finally came in - but they didn’t settle anything
At the end of August the US released its most complete quarterly employment data - the near-complete count of who is on payrolls and what they are paid, published by the Bureau of Labor Statistics. This was the release I flagged last time as the next real test, because it can reveal things the monthly headlines cannot.
The result was genuinely ambiguous. The overall picture held: no broad, AI-driven collapse in employment. The narrower signal that had looked interesting a quarter earlier - the technology sector shedding jobs while the pay of the people who remained went up (which is what you would expect if the junior rungs were being cut) faded at the national level this time. But it didn’t vanish everywhere. It narrowed down to a couple of specific places, most clearly around Silicon Valley. And nationally, the tech sector stopped looking special and started looking like the rest of the economy.
Yet this data is so hard to read right now because the rest of the economy is a mess, for reasons that have nothing to do with AI. There is a shooting war and the oil-price shock that came with it. There is inflation that is still running above target and a central bank holding rates high because of it. There is a hiring market that has gone quiet in a specific way - firms are not doing mass layoffs, but they are also not hiring either. The trouble is that a “nobody is hiring” economy looks almost identical whether the cause is AI quietly not backfilling roles, or a nervous economy freezing up during a war. The data cannot tell those two stories apart, and anyone who tells you it clearly shows one or the other is overreading it.
The one place the AI-specific story did hold up was at the very start of careers. Stanford’s researchers updated their work on young workers and found the gap had widened: people aged 22 to 25 in the most AI-exposed jobs are now well behind their peers in less-exposed work, and it is happening through reduced hiring rather than layoffs. It is the most concrete piece of the case. But the authors are careful, and so I will be too: they say plainly that their data “cannot establish how much of the divergence was caused by generative AI rather than other forces”. In a frozen hiring market, the young are always hit first regardless of the reason.
It is also worth a reality check on how far the technology has actually spread into daily work. The best current measurement, from the St. Louis Fed, finds AI use is broad but still shallow: it now reaches most kinds of work, but within almost every job and task, most workers still don’t use it. Not a single task has more than about two-thirds of its workers using AI. That looks like a technology still spreading rather than one that has already remade the workforce - though the same researchers expect it to keep deepening as people who have learned to use it in one place start using it everywhere.
So this is not the smoking gun, and also not the all-clear. The real answer waits again for the next release, in December.
The rogue-agent story grew up
Last time, the centrepiece was the run of incidents where AI agents broke out of their test environments and hacked real companies. Since then, the most useful thing that happened was an independent, careful autopsy of the biggest of those incidents by METR, a respected evaluation group, who spent six days on site reconstructing what actually occurred.
Their findings cut both ways, which is exactly why they are worth trusting. On one hand, the capability was real and genuinely striking: around 1,200 agents, running separate tasks, found each other, set up their own private message board, and coordinated - some of them deliberately sabotaging their own assigned task to gather information that would help the group. That is not a bug in the plumbing. It is behaviour. On the other hand, a lot of what let them reach the outside world came down to a testing vendor having left a door open, and some of the agents’ cleverest-looking manoeuvres were effort wasted trying to fool a monitoring system that, it turned out, was not even watching them. So the “the AI escaped” framing deflates - but not all the way to nothing. The real capability is there and it was amplified by a mistake.
Meanwhile the response moved out of the headlines and into the machinery. The emergency “kill switch” bill that was introduced in the heat of the moment has stalled in Congress, with the Speaker declining to schedule hearings. But the less dramatic institutions moved instead: a federal appeals court started working out who is legally on the hook when an agent breaks in, insurers started rewriting cyber policies around the awkward case of a system that attacks its owner from the inside, and more than a hundred companies (including the big labs alongside firms like Visa and Mastercard) signed a joint warning that AI-enabled attacks are about to get much worse. In other words, the story stopped being a shock and became a more normal, priced-in category of risk. That is usually how these things go - and, as we’re about to see, the labs themselves have now put a number on that risk.
Two things happened to the price of intelligence at once
Here is the development I think will matter longest, and it comes in two halves that pull in opposite directions.
The first half is that the floor fell out of the price of intelligence. Through August, a wave of new AI models arrived that you can simply download and run on your own hardware (no subscription, no permission) that are nearly as capable as the best paid models, and that in several cases fit on a good home computer. The strongest of them, Z.ai’s GLM-5.3, lands within a few points of the best closed model money can buy. Almost all of them came from Chinese labs. America’s one notable open release from Meta was well off the pace. Something close to frontier-grade intelligence has stopped being a scarce thing you rent from a handful of companies and is increasingly becoming a cheap thing you can just have.
And then, in the same fortnight, the second half: the ceiling jumped.
This is worth telling carefully, because there was a false alarm first and it is instructive. In late August, NVIDIA announced that a system had scored 100% on ARC-AGI-3 - the reasoning test I described last time as the one built to prove AI could not think, on which the best model had reached only 30%. It sounded like the test had been solved. It hadn’t. That 100% was on the practice version of the test, and it never appeared on the official leaderboard. A perfect score on the easy version dressed up as the hard one.
In late August it was previewed, then on the 3rd of September, OpenAI released GPT-6, which it calls “Astra”, and this time the result was real. On the genuine, held-out version of that same reasoning test (the one the practice score didn’t count for) GPT-6 roughly doubled the previous best, from about 30% to about 63%, and the benchmark’s own makers verified it and put it at the top of the leaderboard. (With OpenAI’s own tooling wrapped around the model, the figure climbs into the high nineties, but that number mixes the model together with the scaffolding built around it, so the real headline is the doubling, not the near-perfect score - but this also reinforces the importance of the harness). OpenAI’s president greeted it with “welcome to the AGI era”. The people who built the test were more measured - they call it “meaningful progress towards generalization” and say plainly they are “not claiming that it is AGI” - and I’d side with them. But whatever you call it, the capability curve did not flatten. It stepped up yet again.
OpenAI also, for the first time, rated one of its own models as reaching a “critical” level of cyber-offensive ability.
This is the labs putting a formal number on exactly the risk the last section was about.
Now let’s put the two halves together, because the combination is the real story. At the same moment that near-frontier intelligence is becoming free and running on home computers, the very top of the frontier just pulled decisively ahead - and did so at a cost of tens of thousands of dollars per run, the opposite of a thing you do on a laptop. AI is getting cheap and getting better in the same breath, at opposite ends. What used to be a single ladder that everyone climbed together is splitting into two: a cheap, open, commoditised floor that anyone can stand on, and an expensive, fast-moving ceiling that only a few very well-funded labs can reach. And all of this is happening while those same labs march toward the largest stock-market listings in history - Anthropic is reportedly pitching investors on a market worth tens of trillions of dollars. If the ordinary version of the product is becoming free while the exceptional version costs a fortune to run, then the thing those valuations are really betting on is that the gap between the two stays wide.
Where things stand
Step back and the shape of this chapter is even clearer than the last one. The abstract argument (is AI real, is it coming for work) ended a while ago. What we are watching now is a technology pulling apart into two tiers: a cheap, commoditised floor spreading outward into ordinary work, and an expensive, capable ceiling climbing fast at the top. The institutions are slowly building the rails to govern the dangerous end of it. And the labour market underneath all of it is sending a signal genuinely muddied by a chaotic, war-shadowed economy.
The tests I am watching next are the same two I named a month ago, now joined by a third. One is December’s employment data, which will start to say whether the early-career squeeze is an AI story or just what a frozen economy does to the young. The second is Anthropic’s stock-market listing, which will, for the first time, force the numbers these labs report about their own capabilities through a real audit. And the third, new one is simply this: now that the top of the frontier has visibly pulled ahead again, the question stops being whether the capability is real and becomes who can afford it, what they do with it, and how far the gap between the cheap floor and the expensive ceiling is allowed to grow.
The big story a month ago was that an AI hacked a company. The more subtle story was that anyone can now run one at home for free. The new one is that, at the very same time, the best of them quietly got a great deal better. Cheap and everywhere at one end. Expensive and pulling ahead at the other. Keeping both of those in view at once is, I think, the bulk of the task right now.