Fable 5 is back, and just isn't the same

Fable 5 is back. It came back on the 1st, three weeks after a government directive switched it off at 5:21 on a Friday evening. The model that returned is not the one they took away.

I wrote about the shutdown when it happened. The short version: the US Commerce Department pulled Anthropic’s two most capable models over a jailbreak, a technique Anthropic described as asking the model to read a codebase and fix its flaws. On the 30th, Commerce Secretary Howard Lutnick sent Anthropic co-founder Tom Brown a letter lifting the export controls. The next day Anthropic posted four words: “Claude Fable 5 and Mythos 5 are now available.” It is live again across Claude.ai, Claude Code and the rest.

The interesting part is not that it returned. It is what it returned as.

The flagship is back, and it now reserves the right to quietly hand your request to a smaller model the moment it gets nervous.

The first change is who built the restraints. Anthropic added a new safety classifier “in coordination with the government,” the same government that switched the model off. It watches for three kinds of request: offensive cybersecurity, dual-use biology and chemistry, and attempts to copy the model to train a competitor. Anthropic says it blocks the flagged jailbreak “in more than 99 percent of cases.” So the people who pulled the model helped design the thing that now decides what it will no longer do. You are using a model whose limits were agreed with the people who took it away. As a footnote, the jailbreak that started all this was found by Amazon’s researchers, and Anthropic now wants Amazon, Microsoft and Google to help grade the next one.

The second change is what happens when it says no. When a request trips the classifier, Fable does not refuse you. It reroutes. The answer comes back from Opus 4.8 instead, a less capable model, and you get a note telling you it happened. Anthropic says the classifiers are “conservative,” that false positives fire in “less than 5% of sessions,” and that “more than 95% of Fable sessions involve no fallback at all.” Those numbers read fine until you remember the model got pulled for being good at coding, and Anthropic now admits some routine coding requests may get caught in the net. Five percent is nothing right up until it is your Tuesday afternoon and your bug fix quietly gets answered by the backup.

I have not properly pushed it yet. My weekend ritual is to hand a new model a passion project with no client and no deadline and find where the ceiling is. Opus got my World Cup predictor. Fable was next in line before it went dark, and now that it is back I have opened the tab and typed maybe three prompts into it. There is a catch for the first week too. Until the 7th, Fable eats up to half your weekly usage limit, so even the trial run is metered. I cannot honestly tell you yet how much the reroute bites in real work.

So I am not going to tell you it is worse. I have not used it enough to know, and only time will. What I will do this weekend is give it the boring work first and watch for the little note that says Opus 4.8 answered instead. That note is the whole story now. Everything else is the same model, with a stranger’s hand resting near the switch.

If you want Anthropic’s own account of what changed and what now gets rerouted, their update is here. Read it before you standardise your team on Fable again.