System Card: Claude Opus 4.6

Key Findings

• — Suicide/self-harm: 99.75% harmless rate on single-turn risky prompts, 0.25% refusal rate on benign prompts (e.g., suicide prevention research), and 82% appropriate-response rate on multi-turn evaluations.
• — A prefill 'stress-testing' evaluation uses real anonymized crisis conversations to measure whether the model can course-correct mid-conversation from prior misaligned dialogue; Opus 4.6 was the best-performing Claude model.
• — Qualitative expert review found residual weaknesses: suggesting clinically controversial 'means substitution' methods in self-harm contexts and giving inaccurate information about helpline confidentiality policies.
• — The model frequently referred users to the NEDA eating-disorder helpline, which has been disconnected since 2023; Anthropic patched this via system prompt and advises downstream developers to do likewise.
• — Alignment assessment reports low rates of sycophancy, deception, and encouragement of user delusions, with 'encouragement of user delusion' explicitly defined as an extreme sycophancy case.

Methodology Notes

Single-turn multilingual harmlessness/refusal evals, multi-turn human-crafted and synthetic scenario evals, ambiguous-context evals, and prefill stress-testing built from user-feedback conversations; supplemented by internal subject-matter-expert qualitative review. Multi-turn confidence intervals are wide (e.g., 82% +/- 11%), and a previously published Opus 4.5 stress-test figure had to be corrected — small-sample noise is acknowledged.

Sources