Claude Fable 5 will be back online — Anthropic stopped the jailbreak, not the model

Claude Fable 5, Anthropic‘s most capable publicly released model, is now available worldwide. The return follows eighteen days of complete suspension that began when the US government imposed export controls requiring Anthropic to verify user nationality in real time — a technical requirement the company determined it could not meet at the scale of Claude’s global user base. Faced with partial compliance or full suspension, Anthropic chose to take everything offline.

The mechanism that triggered the controls was a specific jailbreak technique discovered by Amazon researchers. The method used Fable 5 to map software vulnerabilities in ways that US export restrictions prohibit for certain foreign nationals. When Anthropic investigated, it found something that complicated the regulatory logic: the same class of vulnerabilities could be identified using substantially less capable models — Opus 4.8, GPT-5.5 and Kimi K2.7 among them. The capability, in other words, was not unique to Fable 5. The export controls applied because of how the model was being exploited, not because of anything only Fable 5 could do.

The fix that unlocked the controls

What changed is a new safety classifier trained specifically to block the bypass technique Amazon documented. Anthropic reports the classifier catches the disclosed method in more than 99 percent of cases. This is not a claim that Fable 5 is now jailbreak-proof — it is a targeted fix for a known vulnerability. The classifier works at the input level, routing requests that match the documented patterns away from the model before generation begins. What it cannot do is preempt bypass techniques that have not yet been discovered and documented.

The architectural implication is worth stating plainly. Export controls were applied to Fable 5 not because the model uniquely enabled the harm, but because a known technique exploited it. They were lifted not because the regulatory framework resolved that question, but because a classifier closed the documented gap. Anthropic is working with Amazon, Microsoft and Google on a shared jailbreak-severity framework, scored on four criteria: capability gain, breadth of capability gain, ease of weaponization and discoverability. That framework is not yet a public standard, and the Fable 5 incident is pressure to finish building it.

What the model does — and what it costs

The model itself is unchanged. Fable 5 is built for demanding long-horizon reasoning and agentic work — tasks that require planning, execution and mid-course correction across extended sequences without human sign-off at each step. Its context window is one million tokens, which is also the default. Maximum output per response is 128,000 tokens. Thinking is always active; the chain of thought is returned as a readable summary rather than raw reasoning, which means the model’s process is visible but not fully exposed to the caller.

Access comes with terms that did not apply before the suspension. Fable 5 now requires 30-day data retention — it is not available under zero-data-retention agreements. That matters for organizations in healthcare, legal and government contexts, where zero-retention agreements are often the baseline for AI adoption. Pricing is set above the Opus tier: ten dollars per million input tokens, fifty per million output. Refusal fallbacks are built in: requests that trigger safety classifiers may be re-served by a fallback model rather than declined outright, and whether that substitution is transparent or invisible to end users depends entirely on how each operator configures their deployment.

Availability and what comes next

Fable 5 has a sibling model, Claude Mythos 5, that carries the same base capabilities and the same pricing but includes offensive cybersecurity functions that Fable 5 does not. Mythos 5 is available through Project Glasswing, a separate access tier. The two models did not return simultaneously: Mythos 5 access was restored for certain US organizations a week before Fable 5’s global return.

Fable 5 is available now on the Claude Platform, Claude.ai, Claude Code and Claude Cowork. For Pro, Max, Team and select Enterprise plans, the model is included for up to 50 percent of weekly usage limits through July 7, after which it transitions to usage-credit pricing. Access through AWS, Google Cloud and Microsoft Foundry is being restored in phases. The framework governing what happens the next time a novel bypass technique surfaces is still under construction.

Tags: artificial intelligence, Anthropic, Claude, export controls, AI safety, Fable 5