I Used Fable 5 Until It Went Dark — Ahead of Schedule

A line popped up at the top of Claude Code the other day.

The gist: a new model, built for complex work that runs for a long time; until the 22nd it costs nothing extra — it's on my plan — and after the 22nd, if I want to keep using it, I pay per use through the API.

I got stuck on the phrase "complex, long-running work." That's all I do. The long tasks — the ones that run forever, need many steps, can't be interrupted halfway — are exactly the hardest kind to babysit.

So I made a call: while it's still on-plan, use the hell out of it.

I normally can't spend it

Some background. I'm on the highest tier there is.

How roomy is that quota? In a normal week of heavy use, the "7d" weekly bar barely reaches fifty-something percent. I've never hit the ceiling — I don't even bother looking at it.

So I felt safe: even going all-in, a week's quota should be plenty.

I was wrong.

What I actually fed it

I wasn't chatting in one window. I was doing this:

Six sessions running side by side, each minding its own business; on top of that, multi-agent workflows — one command splits into four or five agents researching and working at once. I fed it the heaviest, longest, most token-hungry tasks I had. All day, no stopping.

This kind of usage burns tokens by the multiple. A normal conversation burns one unit; six sessions plus multi-agent burns a dozen, dozens.

The bar starts climbing

Then I watched that weekly quota — the one I never look at — make itself known for the first time.

Day three, evening: 88%. I did a double-take — a normal week doesn't reach 60, and here it was at 88 in three days.

"Eh, keep going, it's worth it." A few hours later:

94%. Shortly after, empty. A quota I usually find half-spent excessive — drained clean in just over three days.

Then it left early

Per the banner, everything was on-plan until the 22nd, pay-as-you-go after. I had it figured out: no panic if the weekly quota ran dry — free until the 22nd, pay through the API after if I really wanted, either way more time with it.

It didn't make it to the 22nd.

Not "I ran out of quota" gone — the model itself went offline, and the API wouldn't reach it either. It left earlier than the date it printed on its own banner.

A model I'd spent three or four days with — just getting a feel for it, learning how to talk to it — gone, just like that. And not coming back.

A model isn't software. It's fresh produce.

This changed a default assumption I'd never noticed I was making.

I used to assume tools are permanent. You install something and it's just there — opens today, opens tomorrow, unless I delete it myself.

But a model isn't software. A model is more like fresh produce — it has a shelf life. And the date printed on the label is the optimistic estimate; it can come early.

Once that clicked, I stopped regretting those three days of going all-in.

Picture the other version: I pace myself, ration the quota, tell myself "no rush, it lasts till the 22nd." What would I be holding today? Nothing. The quota resets to zero at week's end — nothing saved. The model went offline early — not one extra second used. Empty both ways.

Saved quota isn't yours. A retired model isn't yours. The only thing that's truly yours is the part you've already spent — already turned into output.

So

If you're the type who hoards the good stuff, saving it for later, take this as a nudge:

When something is good, finite, and could vanish at any moment, the biggest waste isn't using it too hard — it's hesitating.

I used it hard for four days, and it still left first. But what those four days produced — that stayed.