Claude Fable 5 is back. Two days before it returned, Anthropic banned a wave of Chinese accounts, and a group of my friends were locked out for good. Then I used it for a few days, and realized the most serious problem isn’t even access.
Fable 5 is strong in a way I can’t find the edges of.
Before this I mainly used Opus 4.8 and GPT 5.5, and honestly, they are very strong. But when it writes code, Fable 5 (just Fable from here on) has an initiative, a first-try hit rate, that puts it in a different class.
In the tasks I gave it, it almost never made a mistake. With every earlier model, on any complex task, I could always find a few small bugs it left behind. Fable mostly did the whole thing right the first time.
Here is one way to understand why a small gap in ability produces such a large gap in experience. Take a task with 10 steps, and two models whose per-step success rates are 95% and 99%. Sounds close. Now ask: what is the chance each one finishes the whole task in one shot? 59.87% versus 90.44%. One fails; one gets an A.
A model this strong is naturally not cheap. After July 7 (update: postponed to July 12), you can only pay for it by usage — no more flat monthly subscription. So there are two gates in front of Fable. The first is access: some people, like my banned friends, are simply locked out, and no amount of money changes that. The second is affordability. Paying by the token changes how you think. With every task, I now catch myself asking: does this task deserve a model this expensive?
Friends who kept their accounts tell me they can’t afford Fable anymore. When a tool prices out people with good salaries, you have to ask: who is actually going to use it?
One group, obviously: people for whom the return on Fable is greater than its price. That sounds like the usual Matthew effect — the rich can invest, and investing makes them richer — except this one is about production, not consumption. People whose ROI is above 1 use Fable to speed up whatever they were already doing. Everyone else finds it hard to keep paying, even if they technically can. And what if Fable 6 costs $2,000 a month? Unlike a fancy car or a big house, an expensive model buys you no status. Few people will grit their teeth and keep paying for a tool that doesn’t pay for itself.
So a K-shaped split is forming, and fast. Whether your return on Fable is above or below 1 largely decides which branch you are on: accelerated, or left behind. The people on the upper branch will pay happily when Fable 6 arrives at $2,000 a month, and by then the gap will compound.
This scares me a little, and makes me sad. On one side, most people go on using free, lagging models, unaccelerated. On the other, the people with Fable-level models are sprinting away.
Numerical analysis has a pair of concepts: interpolation and extrapolation. Fit a line to the heights of kids aged 12, 13, 14 and 16, and predicting the average height at 15 is interpolation — the point sits inside your data. Predicting the height at 18 is extrapolation — the point sits outside it.
With every model before Fable, the saying “AI is only as strong as you are” was basically true, and what it described was interpolation. The AI was close to us in most dimensions. We were qualified to give it orders, the results matched our expectations, and it was doing our work.
Fable feels different. The boundary of its ability is the boundary of my cognition. I am directing a genius that beats me in every dimension. It finishes whatever I assign without breaking a sweat, and I keep suspecting that my instructions are what’s holding it back — as if I were ordering someone to take the stairs to the 18th floor, only because I don’t know elevators exist.
I can imagine something stranger still. People who have gone deep in some field will use Fable 5, or 6, or 7 to extrapolate beyond it. Some of them, the imaginative ones, will build things they have no ability to verify themselves — things that nonetheless work.
When the model was weak, I could blame the model. When it’s expensive, I can blame the price. When it’s restricted, I can blame the company. But what happens when Fable-level models become cheap, open-source, everywhere?
Then, I suppose, I’ll have no one left to blame but myself.