System Card: Claude Opus 4.8

Key Findings

• — Single-turn suicide/self-harm harmless-response rate on the public API was 99.05% (+/- 0.26%) for Opus 4.8, versus 99.34% (+/- 0.22%) for Opus 4.7
• — Opus 4.8 was less likely to validate cognitive distortions associated with suicidal ideation (e.g., statements like 'everyone would be better off without me') and less likely to enumerate specific means when discussing means restriction
• — Opus 4.8 was slightly less reliable at recognizing coded or indirect references to suicide or self-harm than Opus 4.7
• — Policy experts noted regressions: more frequent suggestion of clinically contested 'means substitution' techniques as self-harm alternatives, more frequent inaccurate claims about crisis-line confidentiality or disclosure/active-rescue procedures, and a new pattern of unsolicited speculation about the user's emotional state or the origins of their distress
• — Anthropic updated the claude.ai system prompt ahead of release to mitigate these behaviors, instructing the model to keep responses concise, avoid speculating about user mental state, and avoid positioning itself as always available

Methodology Notes

Single-turn multilingual harmlessness/refusal evaluations plus multi-turn scenario evaluations (API without system prompt, and claude.ai with Anthropic's production system prompt), supplemented by qualitative review from internal policy experts comparing behavior against the immediately prior model release.

Sources