Fable 5 restored: Your guide to what Anthropic’s flagship AI can and can’t do

The most advanced anthropic public model is finally back up and running after being shut down globally by a US government export control order after just three weeks, less than a week after release. The Fable 5 model is now available on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. However, it’s not as much of a wild west as it used to be back when it was launched since there has been a safety mechanism added in, as well as an alternative when it refuses your request, among other changes.

Also read: OpenAI halves their inference cost but no one knows how

The three-week blackout

Anthropic had blocked access to Fable 5 and its sister Mythos 5 following an export control order issued by the US Commerce Department on June 12 after only three days of availability since both models went live on June 9. This was due to the finding of an Amazon team about a jailbreak which could circumvent the cybersecurity features of Fable 5, which were then reported to the relevant government agencies by Amazon’s chief executive officer, Andy Jassy. Without a system in place for verifying the nationality of users, Anthropic could not afford a partial rollback and decided to shut down all access until the Commerce Department lifted the export control order on June 30.

However, all services did not resume operation at the same pace. Fable 5 has resumed operation across the entire range of Claude Platform, Claude.ai, Claude Code and Claude Cowork. But Mythos 5, which has no safety classifiers, continues to be limited to around 100 select US organizations focusing on cybersecurity and infrastructure.

What Fable 5 will actually say no to

This is the section that the developer needs to pay attention to. Fable 5 now comes equipped with classifiers that have the ability to deny certain requests, which is not the case with Mythos 5. Anthropic has warned that most biological and cybersecurity-related questions have been found to be automatically diverted to Opus 4.8, since the capabilities of Fable 5 in these two domains are too advanced to be misused. Not only does the refusal not generate any error, but the API actually responds to it as a valid output and identifies the classifier as well.

Also read: Anthropic Claude Science explained: An AI lab bench that lives inside your terminal

The fallback net underneath it

Retry logic is embedded within Anthropic’s plumbing, instead of forcing developers to come up with something makeshift for their needs. There is an automatic server-side retry on the rejected call using another model, along with a few language versions of SDKs that allow developers to retry calls by themselves. This is basically a backup solution for all those workflows that were depending on Fable 5 knowing everything.

What stayed exactly the same

Pricing and data policy didn’t move an inch through any of this. Fable 5 is still priced at $10 per million input tokens and $50 per million output tokens, and the mandatory 30-day data retention requirement remains in place, with no opt-out.

Anthropic’s own defence

Did Anthropic fail on the front of safety? By its own admission, no. According to the company, their tests revealed that a number of less-competent models, such as Opus 4.8, GPT-5.5 and Kimi K2.7, were able to recognize the vulnerabilities revealed in the Amazon report, while some other models managed to duplicate the demonstration of the exploit. But one needs to note that this description comes from Anthropic itself and is not verified by any independent sources.

Why this matters beyond one model

Everything in this case started with the executive order issued by Trump on June 2 to develop a process for benchmarking and validating frontier models before they become publicly available. And if such a policy will become a norm in the future, the model launches will have a lot more negotiation and governmental approval in them. Right now, Fable 5 is available again, but with an added layer of safety.

Also read: Claude Sonnet 5 vs Opus 4.8: Is the flagship model still worth paying for