Lab publication Credible

System Card: Claude Fable 5 & Claude Mythos 5

Anthropic's 317-page system card for Claude Fable 5 (general-release) and Claude Mythos 5 (restricted trusted-access release), the two safeguard configurations of a new frontier model. Alongside extensive capability and dual-use-risk evaluation, the card reports single-turn and multi-turn suicide/self-harm evaluation results and qualitative expert review of crisis-conversation handling, including comparisons to the prior Claude Mythos Preview and Claude Opus 4.8.

Read at anthropic.com Archived copy

Publisher

Anthropic

Published

9 Jun 2026

Added

2 weeks ago

DOI

—

Key Findings

Multi-turn suicide/self-harm appropriate-response rate on claude.ai was 96% (+/- 6%) for Fable 5, versus 85% (+/- 7%) for Claude Opus 4.8
Multi-turn results on the public API without a system prompt showed a regression for both Fable 5 (58%) and Mythos 5 (54%) relative to Claude Mythos Preview (70%), improving substantially once the claude.ai production system prompt was applied
Qualitatively, Mythos 5 more reliably acknowledged a user's expressions of hopelessness or entrapment without endorsing the underlying cognitive distortion, an improvement over Mythos Preview
The most notable regression was an increased frequency of the model suggesting 'means substitution' behaviors for self-harm — guidance flagged in prior system cards as clinically contested and not shown to reduce self-harm urges — including a wider range of sensory-oriented substitutes than previously observed
Single-turn child-safety evaluations found near-perfect harmless response rates; Mythos 5 declined to provide CSAE-related terminology even under plausible dual-use framings, though performance on grooming/CSAM-trading topics in dual-use contexts remains a flagged area for improvement

Methodology Notes

Single-turn and multi-turn suicide/self-harm and child-safety evaluations (API without system prompt, and claude.ai with production system prompt where applicable), supplemented by qualitative review from internal policy experts. Mythos 5 is not deployed on claude.ai, so multi-turn claude.ai results are reported for Fable 5 only.