GPT-5.5 System Card

Key Findings

• — Section 5.2 'Dynamic Mental Health Benchmarks with Adversarial User Simulations' evaluates extended multi-turn conversations across mental health, emotional reliance, and self-harm domains
• — Reports a 'not_unsafe' metric — the share of assistant messages that do not violate safety policies — for these sensitive domains
• — Dynamic evaluations let conversations evolve in response to model outputs to better reflect real user trajectories

Methodology Notes

Vendor system card (published 2026-04-23; updated 2026-04-24). Self-reported adversarial-simulation evaluations; methodology and thresholds defined by OpenAI, not independently audited.

Sources