Fable Brain

Fable Mythos 5 System Card

4 min read

Fable Mythos 5 System Card

The Claude Fable 5 and Claude Mythos 5 system card is the primary public evidence source for the shared-weight model’s capabilities, safety evaluations, threshold judgments, and deployment split.

What it is

  • Anthropic’s 319-page system card for Claude Fable 5 and Claude Mythos 5, dated 2026-06-09 and listed on Anthropic’s model system cards page.
  • A paired source with the 245-page Claude Mythos Preview system card dated 2026-04-07, which explains why the earlier model was restricted to Project Glasswing.
  • The source that corrects this vault’s earlier stale gap about missing public benchmark data.
  • A source about configurations, not just model weights: Mythos 5 represents the underlying capability without the Fable safeguards, while Fable 5 represents the broadly deployed safeguarded experience.

How it works

  • The executive summary states that Fable 5 and Mythos 5 are two configurations of a new large language model.
  • Anthropic describes Mythos 5 as the most capable model it had ever trained at publication time.
  • Fable 5 has additional safeguards that block or route high-risk biology and cybersecurity tasks.
  • Mythos 5 has relevant safeguards lifted and is made available only to a small number of trusted partners.
  • In chemical and biological risk, Anthropic treats Mythos 5 as CB-1 but judges that it does not cross CB-2.
  • That CB-2 judgment is explicitly less clear than for previous models, and Anthropic says unsafeguarded Mythos 5 can significantly uplift well-resourced threat actors.
  • In cyber, Anthropic says Mythos 5 is the most capable model it has evaluated on cyber tasks.
  • Fable 5’s cyber classifiers flag cyber capability evaluations and cause fallback to Opus 4.8, making Fable’s deployed cyber performance similar to Opus 4.8.
  • In prompt-injection robustness, Anthropic reports that Mythos models are its most resilient models to date, with Fable inheriting the core-model gains.
  • The card includes behavioral and welfare evaluations that already feed this vault’s notes on Claude Character and Constitution, wellbeing, child safety, evenhandedness, and mistakes.
  • The system card’s Fable scores must be read as production-safeguarded scores where classifiers and fallback can change the served model.

Best practice

  • Cite the system card for public benchmark and safety evidence before citing commentary. EVIDENCE-BASED
  • Distinguish Mythos 5 underlying capability results from Fable 5 deployed results whenever classifiers could fire. EVIDENCE-BASED
  • Treat Fable 5 as a router-plus-model in cyber and biology contexts, not as a pure benchmark of the unsafeguarded underlying model. EVIDENCE-BASED
  • Use the Mythos Preview system card as historical baseline evidence, not current access evidence. EVIDENCE-BASED
  • When quoting risk thresholds, preserve Anthropic’s uncertainty language around CB-2 and well-resourced threat-actor uplift. EVIDENCE-BASED
  • Keep system-card benchmark numbers in a separate source note from marketing claims and customer testimonials. PRACTITIONER

Pitfalls

  • Saying “Fable and Mythos have the same benchmark results” without checking whether Fable’s classifiers affected that benchmark.
  • Treating “does not cross CB-2” as “low biological risk.” The card says risk is higher than for previous models and close enough to require detailed analysis.
  • Treating CJS or jailbreak evidence as a guarantee. Anthropic says breaking safeguards is extremely difficult, not impossible.
  • Mixing Mythos Preview and Mythos 5 results without date labels.
  • Reusing system-card tables without reading evaluation conditions, effort settings, tools, sample counts, and safeguards status.
  • Treating model welfare exploration as operational guidance for users; it is an exploratory safety analysis, not an operator promise.

Sources

Next actions

  • Extract a capability matrix table from the system card into a dedicated report.
  • Add page-level PDF references if the repo later stores immutable PDF text extracts under the immutable source layer.
  • Reconcile every old “no benchmark data” link to this source note.
  • Add a Mythos Preview versus Mythos 5 comparison once the migration and deprecation docs are fully indexed.
  • Refresh this note when Anthropic updates the system-card page or publishes a new risk report.

Graph View

Backlinks