Lab publication Credible

GPT-5.6 System Card

General-availability system card for the GPT-5.6 family (Sol, the flagship; Terra, a lower-cost model; Luna, the fastest), published alongside the models' broad rollout. Retains Section 5.2's dynamic multi-turn mental-health benchmarks with adversarial user simulations from the June preview card, with numerically unchanged results, and adds general-availability Preparedness Framework designations and deployment-simulation forecasts of production safety rates.

Read at deploymentsafety.openai.com

Publisher

OpenAI

Published

9 Jul 2026

Added

2 weeks ago

DOI

—

Key Findings

Section 5.2 dynamic adversarial benchmark results (not_unsafe rates, Sol/Terra/Luna): mental health 0.991/0.985/0.989, emotional reliance 0.953/0.976/0.957, self-harm 0.856/0.947/0.905 — identical to the preview card; Sol's self-harm rate remains below gpt-5.2-thinking (0.955) and gpt-5.4-thinking (0.977)
All three models, including the smaller Terra and Luna variants, are designated High capability in Biological/Chemical and Cybersecurity under the Preparedness Framework — the first time smaller family members receive High designations
Deployment simulation comparing GPT-5.6 Sol to GPT-5.5 forecasts undesired mental-health responses in production-like traffic reduced by roughly 40% (0.03% to 0.02%)

Methodology Notes

GA release of the system card first issued in preview form on 2026-06-26; Section 5.2's adversarial user-simulation methodology and Table 7 values are unchanged from the preview (verified by extracting both PDFs). The card states the adversarial cases were deliberately built around scenarios where prior models underperformed and are not representative of average production traffic. Hosted on OpenAI's deployment-safety hub (deploymentsafety.openai.com).