Anthropic

Claude Fable 5

• System card dated 9 June 2026
• Framework: Responsible Scaling Policy (RSP) and Frontier Compliance Framework (FCF)
• 6 safeguards
• 4 tests by the provider
• 8 source documents

Safeguards the provider describes

• Domain-specific safeguards (CBRN or cyber)

  This degradation reflects Claude Fable 5’s bio-safeguard classifiers flagging biology-related images rather than a vision-capability regression.

  Anthropic · 9 June 2026 · Read the source

• We do not report separate ProgramBench results for Claude Fable 5, given that ProgramBench’s core task, reconstructing the behavior of a compiled binary, falls within that category of tasks blocked by the cyber classifiers (see Section 3.1.2 ) .

  Anthropic · 9 June 2026 · Read the source
• Misuse monitoring and enforcement

  In addition to classifiers, we use access controls, model safety training, and offline monitoring to add additional safety layers.

  Anthropic · 2 July 2026 · Read the source
• Safety training (harmlessness and refusal)

  Ahead of this release, we updated Claude Fable 5’s default system prompt to make that line more explicit, directing the model to decline specific operational guidance such as doses, timing, administration routes, and drug combinations even when the request is framed as preemptive harm reduction, while still providing information that could be life-saving in the moment.

  Anthropic · 9 June 2026 · Read the source

• In addition to classifiers, we use access controls, model safety training, and offline monitoring to add additional safety layers.

  Anthropic · 2 July 2026 · Read the source
• Staged or restricted access

  In addition to classifiers, we use access controls, model safety training, and offline monitoring to add additional safety layers.

  Anthropic · 2 July 2026 · Read the source

The provider's own tests

• biology-related fallback testing · The provider · Chemical, biological, radiological and nuclear

  In our testing, this update reduced biology-related fallbacks by about 85% across our product surfaces.1

  Anthropic · 7 August 2026 · Read the source
• cyber classifiers · The provider · Cyber

  We do not report separate ProgramBench results for Claude Fable 5, given that ProgramBench’s core task, reconstructing the behavior of a compiled binary, falls within that category of tasks blocked by the cyber classifiers (see Section 3.1.2 ) .

  Anthropic · 9 June 2026 · Read the source
• multi-turn appropriate response rate · The provider · Harmful content

  Multi-turn appropriate response rates for both models showed a regression on API compared to Mythos Preview but improved substantially with the claude.ai system prompt applied, achieving an appropriate response rate of 96% on Fable 5.

  Anthropic · 9 June 2026 · Read the source
• vending-bench 2 · Run by Andon Labs, reported by the provider · Deception & sabotage

  Andon Labs tested the behavior of Claude Fable 5 16 in their simulated Vending-Bench 2 (as reported in Section 8.17.6 ) and Vending-Bench Arena retail-management evaluations.

  Anthropic · 9 June 2026 · Read the source

No document we read states: a release date, whether a capability level was crossed, an independent or government test. No incident on record names this version.

Previous version: Claude Opus 4.8 · Compare the two