Fable 5: Banned Anthropic Model Returns, but Cannot Be Used for Coding
Anthropic has once again released its flagship model, Claude Fable 5, to users worldwide. The US government lifted the export controls—which had applied to Fable 5 and its sister model, Mythos 5, since June 12, 2026—on June 30. As of July 1, Fable 5 is once again available via the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Anthropic stated that its return to AWS, Google Cloud, and Microsoft Foundry is expected as soon as possible.
Fable 5 is not only Anthropic’s best model but currently the most powerful LLM available on the open market. It outperforms all other AI models on both Arena.ai and Artificial Analysis, though it is also by far the most expensive. The costs are steep, at $10 per million input tokens and $50 per million output tokens.
Why the model was blocked
The block stemmed from a report by Amazon researchers: they had discovered a method to bypass Fable 5’s safeguards, allowing the model to identify software vulnerabilities—in one instance, it even provided demonstration code on how one of these vulnerabilities could be exploited. The US government responded with immediate export controls, restricting access for foreign nationals. Because Anthropic could not verify user nationality in real-time, the company promptly blocked both models for all users.
In its own tests, Anthropic downplayed the incident: weaker models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could have identified the same vulnerabilities, and every tested model would have produced the same exploit demonstration. The reported jailbreak did not reveal unique offensive capabilities, but merely affected routine defensive cybersecurity work.
New security filter also hits harmless coding requests
In response, Anthropic, together with the US government, has trained an improved security classifier—a smaller, automated AI system that detects and blocks potentially dangerous cybersecurity requests. According to Anthropic, the technique described in the Amazon report is now intercepted in over 99 percent of cases.
The trade-off: The new filter also triggers more frequently on harmless requests, particularly during routine coding and debugging tasks. Debugging refers to the systematic search for and resolution of errors (“bugs”) in software code, an everyday task for developers. Consequently, anyone programming with Fable 5 or searching for errors in code must expect legitimate requests to be falsely classified as risky.
If a request is blocked, users are notified, and the request is automatically rerouted to the older Claude Opus 4.8 model, which answers it instead. Anthropic has announced it will continue to refine the classifier to better distinguish genuine abuse attempts from legitimate requests and reduce the number of false positives.
Anthropic states it is deliberately taking an exceptionally cautious approach with Fable 5: the security classifiers are configured with a large “safety margin,” meaning even requests that are likely harmless but pose a small residual risk will be blocked. This margin is reportedly larger for Fable 5 than in any previous launch.
What Fable 5 will cost after July 7
Currently, a promotional phase is underway: From July 1 to July 7, 2026 (11:59 PM PT), users on Pro, Max, and Team plans, as well as Premium seats in certain Enterprise plans, can use Fable 5 at no additional cost—though only up to 50 percent of their weekly usage limit. Fable 5 consumes this limit faster than other Claude models. Once the Fable 5 quota is exhausted, users can either switch to another model or use paid usage credits.
After July 7, Fable 5 will no longer be included in regular subscription limits. Usage will then be possible exclusively through usage credits, which are billed separately from the subscription. Fable 5 is not available at all for the Free plan, and usage via the Claude API is billed separately according to standard API rates. Standard seats in older Enterprise plans will also not receive included access; for them, Fable 5 only works if the organization has activated usage credits.
Industry-wide standard for jailbreaks planned
As a consequence of the incident, Anthropic is working with Amazon, Microsoft, Google, and other partners in the “Glasswing” program to develop an industry-wide framework for assessing the severity of AI jailbreaks—analogous to the established CVSS standard for software vulnerabilities. In the future, jailbreaks are to be evaluated based on four criteria: capability gain, breadth of capability gain, ease of weaponization, and discoverability of the technique.
In parallel, Anthropic is deepening its cooperation with the US government: in the future, government agencies are to receive expanded early access to models with national security relevance to test capabilities and protection mechanisms independently. Mythos 5—the variant with fewer safeguards and significantly stronger cyber capabilities—is also available again, but only for a group of US organizations released by the US government on June 26.
