Fable 5 came back — but not the way you left it

Two weeks ago I wrote that Claude Fable 5 had been pulled, and I ended on the question that actually mattered: would it come back?

It has. Fable 5 returned globally on July 1, across Claude.ai, Claude Code, and Claude Cowork. The US Commerce Department lifted the export ban a day earlier, on June 30, and Anthropic's strongest model was live again within hours. Mythos 5, their cybersecurity model, returned at the same time.

If you missed the first installment: the government had frozen Fable 5 in mid-June over a security concern tied to a jailbreak technique, a framing Anthropic disputed at the time. Now it's cleared to operate again.

That sounds like the end of it, and mostly it is. But there's a detail in the fine print worth your time. Fable 5 didn't come back exactly as it left. One condition of its return was a new safety classifier, built to block the disputed jailbreak technique that started the whole episode. Anthropic says it stops that technique more than 99% of the time. The trouble with any filter tuned that tightly is that it also catches things that were never a threat.

The quiet downgrade

When a request trips the new classifier, it isn't blocked. It's rerouted to a weaker model, Claude Opus 4.8, instead of Fable 5.

Ordinary coding and debugging work is exactly what tends to set off false positives here. A security question, an unusual snippet, an exploit you're trying to patch: none of it malicious, but any of it can quietly leave you with a lesser answer than the one you asked for.

Whether you're told when this happens isn't clear. Some reports say there's a notice, others say there isn't, and I'd rather not claim more than the sources actually support. So take this as the practical version: don't count on noticing by feel alone.

There is a way to check. If a debugging reply seems thinner than usual, more generic, or less willing to explain its reasoning, ask the model directly which version answered. Then run the request again. If the second answer is clearly better, the first one was probably Opus 4.8.

The July 7 change worth checking

This next one is a separate issue on a tighter clock.

Several reports describe a pricing and quota change arriving around July 7, in which roughly half of your weekly plan allowance gives way to a pay-as-you-go credit model. Anthropic hasn't confirmed this directly, so I won't tell you your bill is about to change.

What I will say is that it's worth opening your plan settings and reading them before the 7th rather than after. Look for any new "usage credits" or "pay-as-you-go" wording. If it's there, you're better off knowing your limits going into a heavy week than finding them partway through one.

What it means for you

1. If a coding or security answer feels off, ask which model produced it and run the request again. A silent downgrade to Opus 4.8 is easy to undo once you've spotted it.
2. Open your billing settings before July 7 and check for usage-credit language, so nothing about your quota catches you out later.

The short version: Fable 5 is back, but "back" now includes a filter that can decide, on your behalf, when you don't get it. That's a fair trade for the security it buys. The only thing worth insisting on is that you make the trade knowingly instead of discovering it by accident.