Fable 5 Is Back After 19 Days. Here's Whether It Gets My Default Slot Again.

My June 13 post ended with a model I liked and couldn't use. A US Commerce Department export-control directive had forced Anthropic to disable Claude Fable 5 and Mythos 5 globally — four days after I'd made Fable 5 the default in my Claude Code. On June 30, Anthropic announced the directive was lifted. On July 1, Fable 5 was back for everyone — Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. Nineteen days, end to end.

The obvious sequel question is "so, switching back?" The honest answer took me a week of actually using it to write down — because the model that came back isn't quite the one that left, and the lineup it came back to isn't the one it left, either.

What actually happened

Quick recap for anyone who skipped June. The directive was triggered by Amazon researchers who found a jailbreak that bypassed Fable 5's cybersecurity safety classifiers — prompting the model to identify software vulnerabilities and, in one demonstration, write exploit code. The government's read was that this crossed a national-security line, and on June 12 access was restricted worldwide, mid-Friday-evening, with my session open.

On June 30, Anthropic posted that Commerce had lifted the controls, and restoration started the next day. The price of readmission is public: Anthropic trained and deployed a new cybersecurity safety classifier targeting the specific jailbreak — which it says blocks the reported exploit in over 99% of cases — and agreed to proactively detect and address security risks, report malicious activity to the government, and help develop standards for future frontier releases.

The model that came back isn't the model that left

Weights-wise, it presumably is. But a new classifier now sits in the request path, and a classifier is a filter with opinions. For the median coding session — refactors, features, tests — you will never see it. The population that should care is exactly mine: people who ask the model to read code and find what's wrong with it. That was the "narrow jailbreak" all along — vulnerability-finding is both Fable 5's headline capability and the reason it disappeared. The same capability question has since shown up on intelligence-agency letterhead; I wrote about the Mythos reference buried in the Five Eyes statement two weeks ago.

So I spent the week running my normal security-adjacent work through it: auditing my agent's tool permissions, reviewing an auth flow, the find-the-injection-risk passes I do before anything touches my server. The classifier tax on my workloads so far: one refusal-shaped answer on a prompt that boiled down to "find the injection risk in this handler," which a single rephrase turned into a normal audit. That's acceptable for me. But that's a sample size of one developer — if security review is your daily bread, test the classifier against your own workflows before you re-default. Don't inherit my conclusion.

The complication named Sonnet 5

While Fable was away, Anthropic shipped Claude Sonnet 5 and quietly made it everyone's default — at $2/$10 intro pricing through August, edging Opus 4.8 on knowledge work. I did the day-two math on it here. Which means the June question — "Fable 5 or back to Opus 4.8?" — is stale. The July question is: what does Fable 5 add over a Sonnet 5 default, per dollar, on your actual workloads?

For me the answer split cleanly. Sonnet 5 keeps the daily slot: it's fast, cheap, and clears the bar for routine building with room to spare. Fable 5 takes the escalation tier — architecture reviews, multi-file debugging where the bug is a relationship between files rather than a line, and the read-everything-and-tell-me-what's-wrong work that made it my day-one switch in the first place.

What the nineteen days actually taught me

The fallback wiring stays. In June I rebuilt my Claude Code config around model aliases so a one-line change swaps the fleet. That was supposed to be incident response; it's now permanent architecture. Model availability is a dependency like any other — and this one has a regulator.

Both halves of the precedent count. A government switched off a frontier model in an afternoon — and switched it back on in under three weeks. Plan for the first without pricing in permanent doom from the second: build fallbacks, don't flee platforms.

Capability and liability are now the same axis. The thing Fable 5 is best at is the thing that got it suspended. Any model that leads on security-relevant capability inherits regulatory tail risk. That's a routing input now, same as price and latency.

The verdict

Fable 5 is back in my stack — one tier up from where it was: the specialist, not the default. Sonnet 5 holds the daily slot, the fallback aliases stay forever, and the June post-mortem doesn't get retracted — it gets a sequel with a better ending than I expected and exactly the lesson I already paid for. The happy ending doesn't refund the tuition.

Sources

Anthropic's announcement is on X. CNBC and Axios covered the lift and the July 1 restoration; The Hacker News has the best technical summary of the new classifier and the 99% figure; Forbes and 9to5Mac cover the conditions Anthropic agreed to in exchange.