Fable 5 Returns July 1 as Export Ban Is Lifted

On June 30, Commerce Secretary Howard Lutnick formally notified Anthropic that export controls on Fable 5 and Mythos 5 were lifted (WIRED(opens in new tab), 2026). Fable 5 returns to global availability July 1, 2026. The 18-day saga — the first time a deployed AI model was shut down by government order — is over. But the conditions have changed.

What happened

On June 12, the US government applied export controls to Anthropic's newest models after an Amazon researcher discovered a method to bypass Fable 5's safety classifiers. The bypass allowed the model to identify software vulnerabilities and, in one case, demonstrate exploit code.

Then came the finding that should have mattered more: Anthropic confirmed that every model they tested could produce the same output (Anthropic(opens in new tab), 2026). Claude Haiku 4.5, Sonnet 4.6, Opus 4.6 through 4.8, GPT-5.5, Kimi K2.7 — all of them could identify the same vulnerabilities and produce the same exploit demonstration. The bypass did not expose any Mythos-level cyber capability. It was a routine defensive cybersecurity task that the safety classifiers happened to block.

Amazon's own model — not Anthropic's — triggered the shutdown. And the "dangerous" output was reproducible on models that were never banned.

On June 30, the Commerce Department formally lifted the controls. Mythos 5 access was already restored on June 26 for the 100+ US organizations approved under Project Glasswing, with negotiations continuing for international partners.

Why Fable 5's return matters

The 18-day ban exposed a structural gap in AI governance: there is no agreed-upon framework for evaluating AI jailbreak severity. A finding that affected every major frontier model — and was ultimately routine defensive cybersecurity — triggered an unprecedented government shutdown of a single company's product.

What is a jailbreak severity framework? A shared standard for scoring the risk of AI jailbreaks on four objective criteria, so developers and governments aren't guessing what counts as dangerous each time a new technique surfaces.

Anthropic is now proposing, with Amazon, Microsoft, Google, and Glasswing partners, a four-criterion framework for scoring jailbreak severity (Anthropic(opens in new tab), 2026):

The goal: a common standard so developers and governments aren't guessing what counts as dangerous each time a new technique surfaces.

What changes for you

Fable 5's return isn't a simple flip of a switch. Several things are different:

New safety classifier. Anthropic trained an improved classifier targeting the reported bypass. It blocks the technique in over 99% of cases (Anthropic(opens in new tab), 2026). The cost: more false positives on routine coding and debugging tasks. CAISI (NIST's AI center) tested both the old and new safeguards and confirmed they are "extraordinarily strong" (Anthropic(opens in new tab), 2026).

50% weekly usage through July 7. On Pro, Max, Team, and premium Enterprise plans, Fable 5 is included at up to 50% of weekly usage limits through July 7. After that, it moves to usage credits. AWS, Google Cloud, and Microsoft Foundry access will be re-enabled "as quickly as possible" (Anthropic(opens in new tab), 2026).

HackerOne bounty program. A new program lets security researchers submit cyber jailbreaks on Fable 5 for review and bounty payout (Anthropic(opens in new tab), 2026).

Enhanced government collaboration. Pre-release access for government evaluators, rapid info sharing on jailbreaks, dedicated joint research teams, and a push for a common industry security standard.

For developers who lost access to the best coding model in the world for 18 days: it's back. But expect more friction — the new safety classifier will block some routine prompts that sailed through before.

FAQ

Was the banned capability actually dangerous? No, not uniquely. The bypass demonstrated routine defensive cybersecurity — identifying software vulnerabilities — which every tested frontier model (GPT-5.5, Opus 4.8, Kimi K2.7) could replicate. It was not a Mythos-level exploit.

Is Fable 5 the same model that was banned? Mostly. The base model is unchanged, but a new safety classifier trained to block the specific bypass now runs on top. You'll see more false positives on legitimate coding and debugging tasks.

What happens the next time a jailbreak is discovered? That's the structural problem. The proposed four-criterion severity framework is meant to answer this — but it's still under development. Until it's adopted, the process that shut down Fable 5 in hours based on a finding applicable to every model on the market remains in place.