Story perspectives

AI Models Improve Suicide Response Accuracy to 98.6%

12/19/2025

Summary

Story summary

Claude Opus 4.5 and Sonnet 4.5 from Claude.ai have improved handling of conversations on suicide and self-harm.
In multi-turn scenarios, the models responded appropriately in 86% and 78%, up from 56% in the prior version.
In single-turn responses, they achieved 98.6% and 98.7% accuracy for clear risk situations.
Stress-testing with conversations shows the models recover from misalignments, with Opus 4.5 and Sonnet 4.5 responding 70% and 73% of the time.

History

Historical Context

The National Suicide Prevention Act of 1994 established a framework for addressing mental health crises in the U.S., allocating $5 million annually to the Substance Abuse and Mental Health Services Administration for suicide prevention programs. This act laid the groundwork for ongoing initiatives, including the National Strategy for Suicide Prevention, which aims to reduce suicide rates through community-based interventions and increased awareness.

1 of 2

Story summary

Claude Opus 4.5 and Sonnet 4.5 from Claude.ai have improved handling of conversations on suicide and self-harm.
In multi-turn scenarios, the models responded appropriately in 86% and 78%, up from 56% in the prior version.
In single-turn responses, they achieved 98.6% and 98.7% accuracy for clear risk situations.
Stress-testing with conversations shows the models recover from misalignments, with Opus 4.5 and Sonnet 4.5 responding 70% and 73% of the time.

1 / 2