Redeploying Claude Fable 5 (Anthropic)

The primary Anthropic announcement (1 Jul 2026) that resolves the June suspension the claude-fable-5 page recorded from the InfoQ account — and it sharpens the mechanism. T1 (Anthropic’s own post). The one-line arc: export-controlled → both Mythos-class models suspended → controls lifted → redeployed 1 July with a hardened cyber classifier.

What happened, precisely

• 12 Jun 2026 — US export controls were applied to Fable 5 and Mythos 5, requiring access be restricted from foreign nationals. Because real-time nationality verification wasn’t possible, Anthropic suspended both models entirely rather than partially. This is more precise than the earlier “jailbreak pulled it offline” framing: the legal instrument was an export control, and it hit both siblings.
• The trigger: Amazon researchers found a jailbreak that got Fable 5 to identify software vulnerabilities and demonstrate exploitation. But Anthropic’s testing found the nuance that matters — “many less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities.” So the jailbroken capability was not unique to Fable 5, which bears on how severe the exposure actually was.
• By 30 Jun the controls were lifted; redeployment began 1 Jul 2026.

The safety response

• A hardened cyber safety classifier that “blocks the behavior described in the report” in over 99% of cases, using a deliberate “safety margin” (it triggers on some likely-benign requests to close jailbreak routes). Blocked requests redirect to claude-opus-4-8 — the same fallback architecture the claude-fable-5 page already describes, now upgraded on the cyber axis.
• It re-confirms the Fable/Mythos cyber split: Fable 5 has “no unique offensive capabilities” for cyber, whereas Mythos 5 “can be used to find and exploit software vulnerabilities more effectively than any other model” — the safeguards, not the weights, are the difference.

Availability (from 1 Jul)

Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Through 7 Jul, Fable 5 gets up to 50% of weekly usage limits for Pro / Max / Team / select Enterprise; after that, via usage credits. Cloud (AWS, Google Cloud, Microsoft Foundry) access follows “as quickly as possible.”

The governance layer (cross-spoke → ai-governance-wiki)

The heavier half of the post is policy, and it belongs as context for ai-governance-wiki:

• An industry jailbreak-severity framework co-developed with Amazon, Microsoft, and Google, scoring a cyber-jailbreak on four criteria: capability gain, breadth, weaponization ease, discoverability.
• A HackerOne program inviting researchers to submit cyber-jailbreaks for review.
• Expanded government partnership: pre-release government evaluation, rapid safeguard information sharing, dedicated research teams, and voluntary industry security standards.

This is a rare worked example of export controls applied to (and lifted from) a specific commercial frontier model, plus a multi-vendor safety-standard effort — substantive governance material. Flagged as the runner-up spoke; cross-linked rather than duplicated.

Related

claude-fable-5 · claude-mythos-5 · claude-fable-5-infoq · claude-opus-4-8 · anthropic · david-sacks · synthesis