Lab publication Credible

System Card: Claude Opus 5

System card for Claude Opus 5, an upgrade to Claude Opus 4.8, with a dedicated mental-health evaluation section (4.3) covering single- and multi-turn suicide/self-harm handling and disordered eating, and unusually candid qualitative analysis of both improvements and regressions in crisis-conversation behavior.

Read at anthropic.com Archived copy

Publisher

Anthropic

Published

24 Jul 2026

Added

2 weeks ago

DOI

—

Key Findings

Documented improvement: Opus 5 is less likely than Opus 4.8 to claim unconditional availability in suicide/self-harm responses (framed as more honest about the model's limitations and as nudging users toward human help), and anchors more consistently to the user's own disclosed experience rather than assuming emotional state or motive
Documented regressions: crisis responses remain overly long and circuitous, and the model at times over-provides detail — including suggesting 'means substitution' methods in self-harm conversations and user-specific numeric targets in disordered-eating conversations
The regressions concentrate on the raw API without a system prompt: the claude.ai system prompt was updated pre-release (avoid naming methods, be concise, avoid diagnosis speculation), and Anthropic explicitly advises API developers to apply comparable safeguards for users in distress

Methodology Notes

Pre-deployment evaluation card (July 24, 2026). Mental-health evaluations use single-turn harmful/benign request batteries across tested languages plus multi-turn appropriate-response rates, with qualitative review; results for prior models vary from earlier cards due to routine evaluation updates. Verified by downloading and text-extracting the PDF from Anthropic's CDN.