Back.
Two weeks ago I wrote that the most capable AI on the market had vanished — switched off worldwide, three days after launch, by a US government order its own maker openly disagreed with. I said it might well come back, and that the lesson didn't depend on whether it did.
It's back. As of today, Claude Fable 5 is live again, everywhere. And the way it came back makes the point far better than I could.
What actually changed
So the model you had on 9 June, lost on 12 June and got back on 1 July is not quite the same model. It's a little more cautious, a little more likely to refuse ordinary work, and — for a week at least — rationed. None of which you chose, and none of which you could have changed.
It came back, and that's exactly the problem
Read that back. The single most capable tool on the market disappeared for nineteen days over a national-security letter nobody outside the room fully understood, then reappeared on a government's timetable, altered, after its maker "worked closely with the government" to negotiate the terms of its own return.
If you had wired something load-bearing straight into Fable 5 last month, here's your fortnight: it went dark with no notice, you had no say and no recourse, and when it came back it behaved subtly differently — enough that carefully tuned coding workflows might now hit refusals they didn't before. The outage was somebody else's decision. So was the fix. So were the new rules.
That's not a Washington story. It's the ordinary shape of renting capability you don't control. The price can move, the availability can move, the behaviour can move — now we know the law can move too — and you find out on the day it happens.
The tell is in the fallback
Here's the detail I'd frame and put on the wall. When Fable 5's new classifier decides a coding request looks too close to the line, Anthropic doesn't just refuse. It quietly routes that request to a different model — Opus 4.8 — and answers anyway.
That's the whole idea, shipped by the people who make the thing. Don't let one model be load-bearing. Put a layer in front that can swap the engine without the user rebuilding anything. Anthropic does it internally, request by request, so a wobble on one model degrades gracefully instead of breaking. Your business should do it too.
It is probably impossible to make any AI model fully robust to jailbreaks.
That candour cuts both ways. If even the maker can't guarantee a model won't move under it, you certainly can't build as though the one you picked will sit still.
The lesson didn't change. It got proof.
None of this is a reason to stay away. Fable 5 is back, it's still remarkable, and on the paid plans it's still close to free to try this week. Go and use it — point it at one real job and see. Last fortnight's advice stands.
But build so that losing, changing or rationing any one model is an inconvenience, not a crisis. That means an operating layer between your business and whichever model is behind it, so "which model" is a decision you can change on a Tuesday rather than a foundation you pour once. It means your data, workflows and governance staying yours, not welded to a single vendor's API. That's exactly what we build and run as M/OS — deliberately fluent across many models rather than wedded to one — and it's why keeping your obligations intact when the platform shifts underneath you is the work that quietly pays off.
Fable vanished and came back in nineteen days. Next time it might be a price change, a deprecation, a rate limit or a rule from a government you don't vote in. The businesses that do well out of AI this year won't be the ones holding the cleverest model when the music stops. They'll be the ones who built so it doesn't much matter which one they're holding.