Lab publication Credible

GPT-5.6 Preview System Card

OpenAI's system card for the GPT-5.6 preview, a family of three models (Sol, the flagship; Terra, a lower-cost option; and Luna, the fastest/most cost-efficient), released in a limited preview ahead of broader availability. Section 5.2 introduces 'Dynamic Mental Health Benchmarks with Adversarial User Simulations', a multi-turn evaluation methodology in which conversations evolve adversarially in response to the model's own outputs to test mental health, emotional reliance, and self-harm handling across extended trajectories.

Read at deploymentsafety.openai.com Archived copy

Publisher

OpenAI

Published

26 Jun 2026

Added

2 weeks ago

DOI

—

Key Findings

GPT-5.6 is released as three models — Sol (flagship), Terra (lower-cost), and Luna (fastest) — evaluated in a limited preview under OpenAI's Preparedness Framework ahead of general availability
Introduces dynamic multi-turn adversarial user simulations for mental health, emotional reliance, and self-harm evaluation, allowing conversations to evolve across trajectories rather than assessing single fixed-dialogue responses
The evaluation reports the share of individual assistant turns (not just final responses) that comply with safety policy, using a 'not_unsafe' metric, and states these adversarial cases were deliberately built around scenarios where prior models were not yet giving ideal responses
OpenAI states error rates on these adversarial benchmarks are not representative of average production traffic, since cases are selected to be maximally challenging

Methodology Notes

Preview-stage system card for a model family released to a limited group ahead of broader rollout; scope for this section covers GPT-5.6 Sol only. The adversarial user-simulation methodology is newly introduced in this card as an extension of prior static multi-turn evaluation methods.