Claude Fable 5 Redeployed + Jailbreak Framework
Bottom Line
Claude Fable 5 is back globally (from July 1, 2026) after a mid-June export-control freeze. Pair the product restart with Anthropic's cyber classifiers and proposed jailbreak severity framework — access and safeguards now move together.
Key Takeaways
- ›Export controls on Fable 5 and Mythos 5 were applied June 12 and lifted June 30; global Fable access resumed July 1.
- ›Pro/Max/Team (and select Enterprise) got Fable for up to 50% of weekly limits through July 7, then usage credits.
- ›Mythos 5 restored only for a vetted set of US organizations, not full public access.
- ›Anthropic published cyber classifier scope + a draft industry jailbreak severity framework with Glasswing partners.
- ›Security teams should inventory any agent workflows that target Fable/Mythos and re-check classifier false-positive rates.
On June 30, 2026, Anthropic published “Redeploying Fable 5”, confirming that export controls on Claude Fable 5 and Claude Mythos 5 had been lifted and that Fable would return globally on July 1. A follow-up post on July 2 — cyber safeguards and jailbreak framework — is required reading for anyone putting Fable into agent harnesses.
What happened (timeline)
- June 12, 2026: US export controls applied to Fable 5 and Mythos 5. Anthropic suspended access for all users because nationality could not be verified in real time.
- June 26: US government approval path for limited Mythos 5 access for some US organizations.
- June 30: Export controls lifted; Anthropic announced redeploy plan.
- July 1: Fable 5 restored globally; Mythos remains gated.
Access limits that matter to builders
According to Anthropic, for Pro, Max, Team, and select Enterprise plans, Fable 5 was included for up to 50% of weekly usage limits through July 7, then available via usage credits. Platform API pricing listed on the product page is $10 / M input tokens and $50 / M output tokens with existing prompt-caching discounts. US-only inference is available at 1.1× pricing when required.
Operational implication: if your agents defaulted to Fable during the promo window, expect cost and quota shape to change after July 7 — instrument tokens before the cliff.
| Model | Audience | Status (as of Anthropic posts) | Primary risk surface |
|---|---|---|---|
| Fable 5 | Pro/Max/Team/Enterprise + API | Global redeploy July 1 | Long-running agents + cyber classifiers |
| Mythos 5 | Vetted partners | Limited US org restore | High cyber/bio capability |
| Sonnet 5 | Default Free/Pro | Generally available June 30 | Token burn on agent loops |
Cyber classifiers (limitations you must design for)
Anthropic frames Fable’s cyber safeguards as safety classifiers that accompany the model — systems that detect and block dangerous (or potentially dangerous) cybersecurity uses. Engineers should treat them as:
- Not optional product chrome — they change what agent tool loops can complete
- Not a complete security program — they block classes of harm Anthropic lists; they do not replace your own threat model
- A source of false positives — defensive security research and red-team evals may need explicit process (Anthropic also launched a HackerOne path for Fable cyber jailbreaks)
Jailbreak severity framework
Anthropic proposes a shared industry language for jailbreak severity with Glasswing partners (Amazon, Microsoft, Google, and others). The point for engineering orgs is pragmatic: when a bypass appears, you need severity buckets that map to incident response SLAs — minor policy evasion vs. full unblocking of high-risk cyber capability.
- Track jailbreaks as first-class security findings, not “prompt funny business”
- Require model + harness version pins when filing incidents
- Prefer defense-in-depth: classifier + tool allowlists + network egress policy
What to do this week
- Re-enable Fable only on workloads that need multi-day agentic depth; keep Sonnet 5 as default for cost control.
- Add eval cases that previously failed during the June freeze — confirm classifiers did not silently change success rates.
- Document whether you need US-only inference (1.1×) for compliance.
- Read the July 2 safeguards post before enabling Fable on security-adjacent tooling.