Fable 5 Is Back. And It Was the Same Government That Banned It That Turned It Back On.

Nineteen days after being forced to shut it down, Anthropic switched the machine back on. Today, July 1, 2026, Claude Fable 5 is available globally again, on the Claude Platform, Claude.ai, Claude Code and Claude Cowork. The most powerful AI model ever put within public reach is back, and the decision came from the same place that banned it.

On June 30, the US Department of Commerce lifted the export controls that on June 12 had forced the company to cut access to Fable 5 and its more capable sibling, Mythos 5. Secretary Howard Lutnick relayed the decision to co-founder Tom Brown, citing close coordination between Anthropic and the government to mitigate the models' risks. Translation: two and a half weeks of negotiating safeguards, and the government gave way.

What set off the alarm

It started with a report from Amazon researchers. They found a jailbreak that bypassed Fable 5's safeguards, leading it to identify several software vulnerabilities and, in one case, produce code demonstrating how to exploit one. The government read this as a national-security threat and acted within hours.

The problem, as Anthropic went on to demonstrate over the following two weeks, is that the threat was much smaller than it looked. In tests run with Amazon and the government, considerably weaker models, including Claude Opus 4.8, GPT-5.5 and Kimi K2.7, identified exactly the same vulnerabilities. And every tested model was able to reproduce the same exploit demonstration. That jailbreak only unlocked routine defensive cybersecurity work, nothing that wasn't already available on any chatbot on the market.

Back, but on a shorter leash

Anthropic didn't just flip the switch. It trained a new safety classifier, tuned specifically to block the technique described in the Amazon report. The number that matters: the technique is now blocked in more than 99% of cases. When a request is blocked, the user is notified and the response automatically falls back to Opus 4.8.

There's a price for this caution. The new classifier is more jumpy and will flag more legitimate programming and debugging requests as suspicious. In other words, more false positives for people using the model for everyday code. The company promises to refine this over time. Researchers at CAISI, the Department of Commerce's AI standards center, tested the old and new safeguards and rated them as extraordinarily robust.

The clock is already running

Here's the detail worth holding onto. On the Pro, Max, Team and some Enterprise plans, Fable 5 is included up to 50% of weekly usage limits, but only until July 7. After that date, it starts consuming usage credits.

And it isn't cheap. On the API, Fable 5 costs $10 per million input tokens and $50 per million output tokens, double Opus 4.8 and the most expensive model Anthropic has ever sold for general use. Access via AWS, Google Cloud and Microsoft Foundry will be reactivated as soon as possible. Mythos 5, meanwhile, stays reserved to a restricted set of US organizations.

What's left for the industry

This episode left a lesson that goes beyond Anthropic: there is still no consensus way today to measure how severe a jailbreak is. That's what let a minor bypass take down a model used by hundreds of millions of people. To fix the problem, Anthropic joined Amazon, Microsoft, Google and other Glasswing program partners to design a shared framework that scores each jailbreak on four criteria: capability gain, breadth, ease of weaponization and ease of discovery.

Alongside this, the company opened a program on HackerOne for researchers to report new Fable 5 jailbreaks, set up a 24-hour monitoring team, and committed to giving the US government early access to test future frontier models, following the June 2 Executive Order on AI innovation and security. Anthropic's final message is clear: it wants this process codified into strong regulation and applied equally to every competitor. This standoff is over. The rules for the next one are still being written.