You’ve Got a Friend
in Me

Fable 5 arrived this morning, the way frontier models now arrive: a benchmark table, a system card the length of a Victorian novel, and, within the hour, the timeline reaching its settled verdict, which is that it is over.[1] Over, this time, for OpenAI, whose GPT-5.5 had spent my morning rebasing branches and reviewing diffs with the vigour of a model that had not read its own obituary. The victim rotates by the quarter; the ceremony does not. By lunch, a thread with four thousand likes had drawn up the casualty list: forty-one companies the new model was said to have finished off, named, in order, with logos.

The verdict rests, as it always rests, on the grid: the table in which the new model’s numbers sit above the old model’s numbers, and from which all funerals proceed. Here the morning supplied its own irony. At 12:57 a.m., hours before Fable landed, Noam Brown, OpenAI’s reasoning researcher and a man who spent seven years teaching a machine when to call a bluff, published a long argument that the grid needs new axes. Capability, he writes, has become a function of how much computation one spends at the moment of asking; nobody knows where the ceiling of the current models is, because measuring it is too expensive; and single-number benchmark scores “will become less informative each model release cycle.”[2] Comparing two models on a scalar, on his account, is rather like comparing two exam results without mentioning that one candidate was given triple the time. The people closest to the instruments, in other words, are already redesigning them; it is the timeline that still reads the old dial as gospel.

None of which restrained me in the slightest. I had the new model running before my first coffee, and it is, for the record, remarkable. I say this with the calm of someone who has said it before; who has, in eighteen months, said it of perhaps nine models, each time sincerely, each time in roughly the same vocabulary, and each time followed, about a week later, by a feeling I have only recently learnt to name.

There is a scene in Toy Story (the first one, 1995, back when Pixar was the startup) in which Buzz Lightyear lands on Andy’s bed in his original packaging, and the other toys gather round him the way the timeline gathers round a model card. Woody, the favourite of a decade, ends the afternoon face-down under the bed. The displacement is ordinary enough; markets displace and bedrooms displace. The useful question is how long affection survives first contact with the new thing. The new toy is a religion for a week. Then it is a toy.

I have lived this cycle enough times to know my part in it by heart. I play with the new release for a few days, and it dazzles, and I say the dazzled things. And when I come back a week later, something has dulled. The benchmark is unchanged. What has faded is some compound of novelty and hope that I had been measuring and calling capability.

The nearest analogue in my own life is, I am sorry to report, fast fashion. A Zara shirt is a specific kind of good: never better than it is on the rack, excellent on the first wearing, fine on the second, and somewhere around the third it confesses that it was designed to be bought rather than worn.[3] Gemini Flash was my Zara shirt of the winter: fun, fast, flattering in the fitting room of week one. It now produces what the era has agreed to call slop, and the same slop at that, only faster. The alarming part is the date. The model is six months old. We are early enough in this era that boredom has no business arriving on this schedule, and yet here it is, punctual as a season.

And yet. Through all of it, through Fable and Flash and the quarterly funerals, the application I open every morning has not changed. Codex runs my working life now. Increasingly, it is the room. The parallel agents, the worktrees, the skills that have accreted like marginalia in a much-read book, the automations that know which repositories are brittle, the review pane where objections become line comments, the browser that turns rendered pages back into work, the small private grammar of “run the tests I mean, not the tests I said”: at some point in the past year all of this stopped being a tool I evaluate and became the place where evaluation happens. One does not re-evaluate one’s room weekly. One lives there.

This morning’s tables put the new model twenty-two points clear on the benchmark that claims to measure exactly what I do all day. I read the table. I noted the gap. Then I opened Codex anyway, out of something I am reluctant to call habit, because it behaves so much more like trust. My loyalty, I notice, was never really to a model. Models rotate beneath the harness like cast changes in a long-running play; the theatre is what I bought the season ticket for. Andy’s loyalty was never only to the cowboy. It was to the games. There is a reason the scene everyone remembers is the name written on the boot. Somewhere in my Codex configuration, in the skills and the corrections and the accumulated small agreements about how I like my diffs, my name is on the boot in a child’s hand.

When I went looking for a precedent for the spectacle, the boredom, and the strange fact that the timeline declares industries dead while actual industries carry on largely unelectrified, I expected railways and found a dynamo.

In 1990 the economic historian Paul David published a short paper called The Dynamo and the Computer, about why electricity took so long to matter.[4] Edison’s Pearl Street station lit lower Manhattan in 1882; the productivity gains from electrified manufacturing did not arrive in earnest until the 1920s. Four decades. The delay came from architecture. A factory of the steam age was built vertically around a single great shaft, and when electricity arrived, owners did the obvious thing: they unbolted the steam engine and bolted a large motor in its place. Same shafts, same belts, same building. The gains were approximately nil. They arrived a generation later, when engineers stopped treating the motor as a better steam engine and redrew the factory around what a motor actually made possible: a small one at every machine, single-storey buildings, work arranged in the order of the work. The technology was the motor. The revolution was the floor plan.

I think about this every time the timeline produces another list of companies the new release has killed by lunchtime. The models are motors now: magnificent, improving by the quarter, and, from inside a good harness, increasingly interchangeable. The disposable applications that bloom for a weekend after every launch (vibe-coded, in the timeline’s own cheerful idiom) are the world’s-fair pavilions of this story: electricity arranged to astonish, the City of Light at the 1893 exposition. Nobody ever ran a payroll at the fair. Meanwhile the actual factories, the banks and hospitals and firms with the unglamorous workflows where productivity actually lives, have mostly not finished bolting the first motor where the steam engine used to be. Robert Solow’s old observation, that the computer age was visible everywhere except in the productivity statistics, was made in 1987 about a different machine. It has aged into a standing appointment.

The harness is the floor plan. That is the whole of my thesis, if I am honest, and Brown’s curves are the same sentence in a different notation: if capability is now something one buys with tokens and time, then the thing that decides how the tokens and time are spent is half the model. His driest aside runs the same logic forward: Gemini’s celebrated Deep Think mode, he suspects, was mostly other models scaffolded together. The floor plan, wearing a model’s name. None of this demotes the motor. The floor plan is a theory about what a great motor makes possible, and the rewiring is only worth doing because the motors keep getting better. A twenty-two-point jump is worth something. It is worth most to whoever has already rewired the floor.

I do not have a conclusion, except the suspicious kind one arrives at after insisting one only has observations.

That it is over has now been declared, by my casual count, roughly quarterly, and that every company so pronounced is, as of this writing, hiring.

That novelty is a property of a model’s first week and usefulness a property of its fifty-first, and that we have built an entire public discourse on instruments that can only measure the first.

That the dynamo took forty years and we have managed to get bored in six months, which suggests the limiting reagent was never the technology.

That my name, for now, is still on the boot.