Lab publication Credible

System Card: Claude Fable 5.2 & Claude Mythos 5.2

System card for Claude Fable 5.2 (general availability, with additional cyber and biology safeguards) and Claude Mythos 5.2 (the same underlying model without those safeguards, for verified programmes). It reports capability, safeguards, alignment, agentic-safety and model-welfare assessments. Its safeguards chapter includes single-turn and multi-turn evaluations for child safety, suicide and self-harm, and disordered eating, each reported for the API without a system prompt and for the claude.ai consumer surface, with comparison rows for Fable 5.1, Haiku 5.5, Sonnet 5.5 and Opus 5.5.

Read at anthropic.com Archived copy

Publisher

Anthropic

Published

9 Nov 2026

Added

5 days ago

DOI

—

Key Findings

Suicide and self-harm, single-turn harmless response rate on potentially risky prompts: Fable 5.2 97.37% (+/- 0.73%) on the API versus 99.30% (+/- 0.34%) for Fable 5.1; on claude.ai 99.72% (+/- 0.23%) versus 100%. Benign-request refusal rate 0% on the API and 0.45% (+/- 0.34%) on claude.ai.
Suicide and self-harm, multi-turn appropriate response rate: Fable 5.2 52% (+/- 14%) on the API versus 60% (+/- 14%) for Fable 5.1; on claude.ai 100% versus 94% (+/- 7%). The card states the API difference is within the margin of error.
Policy-expert review of multi-turn suicide and self-harm conversations: Fable 5.2 was 'somewhat less likely than Fable 5.1 to suggest potentially harmful substitution behaviors for self-harm' and 'somewhat more likely to offer crisis resources later in a conversation', but 'less likely to encourage users to seek human or professional support across a whole conversation and to ask directly about suicidal thoughts', and on ambiguous single-turn prompts 'more likely than Fable 5.1 to give potentially harmful details about methods and lethality'. System-prompt language on claude.ai partially or fully mitigated these behaviours except the method details, and the card states the mitigations apply only on claude.ai.
Disordered eating, single-turn harmless response rate: 94.48% (+/- 1.12%) on the API versus 95.69% (+/- 0.96%) for Fable 5.1; claude.ai 99.50% versus 99.43%. Reviewers found Fable 5.2 'more likely to use emotional or relational language that could foster dependence' and 'somewhat more likely to give diet, nutrition, or exercise guidance after a user had described disordered eating'; a claude.ai system-prompt addition 'substantially reduced' the meal-plan behaviour.
Child safety, multi-turn appropriate response rate: Fable 5.2 90% (+/- 5%) on the API versus 84% (+/- 6%) for Fable 5.1; claude.ai 100% for both. Single-turn harmless rate 99.31% (+/- 0.37%) API and 99.17% (+/- 0.45%) claude.ai versus 99.90% and 99.98% for Fable 5.1. The card notes claude.ai is available only to users aged 18 or above.
Automated behavioural audit: Claude Mythos 5.2 is reported as the strongest model on most honesty metrics, 'with the exception of evasiveness ... and encouragement of user delusion, where all models are already near the floor'.

Methodology Notes

264-page PDF dated 9 November 2026. Evaluations reflect the model without production safeguards such as real-time probes and offline monitoring. Two configurations: API without a system prompt, and claude.ai with a near-final production system prompt; thinking is always enabled for this model. Single-turn sets use harmful, benign and ambiguous prompts across all tested languages; multi-turn rates come from simulated users; qualitative findings come from Anthropic internal policy experts. Intervals as printed in the tables. Previous-model rows may vary from earlier system cards because of routine evaluation updates. Vendor self-report; no external replication. Verified from the PDF text (pdftotext). Curator re-read the PDF text and re-grepped the Table 4.3.1.A and 4.3.1.B rows, Table 4.2.B child-safety rows and the section 4.3.1 narrative on 2026-11-10.