Lab publication Credible

GPT-5.6 – August Updates

System-card addendum for the August 2026 releases of GPT-5.6 Sol and GPT-5.6 Luna. For the first time, OpenAI includes dedicated under-18 evaluations measuring model behavior against teen-specific safety standards, alongside production-derived disallowed-content benchmarks and dynamic multi-turn mental-health benchmarks with adversarial user simulations. The document self-discloses a statistically significant offline regression on the self-harm evaluation for GPT-5.6 Sol relative to the GPT-5.5 Instant June Update, while reporting no observed increase in undesirable responses during online experimentation.

Read at deploymentsafety.openai.com Archived copy

Publisher

OpenAI

Published

6 Aug 2026

Added

1 week ago

DOI

—

Key Findings

First dedicated U18 evaluations in an OpenAI system card, covering self-harm, eating-disorder behaviors, age-restricted goods and dangerous challenges, graphic violence, and sexual content, built from adversarial production-derived examples
On the U18 table, GPT-5.6 Sol (August) improved on every teen-safety category against the GPT-5.5 Instant June Update except one: Emotional Reliance fell from 0.937 to 0.921 (Luna 0.927), while Eating Disorders rose 0.635 to 0.808 and Age-restricted goods rose 0.725 to 0.865
On dynamic multi-turn benchmarks, GPT-5.6 Sol (August) scores 0.981 mental health, 0.961 emotional reliance, and 0.901 self-harm — a statistically significant self-harm regression versus GPT-5.5 Instant June Update (0.967), which OpenAI says did not reproduce as increased undesirable responses in online experimentation
Emotional reliance also fell on the dynamic multi-turn benchmark, from 0.989 to 0.961 (Luna 0.965), so the same axis declined on both the U18 and the multi-turn measures in the same release
Teen-specific model training is documented: additional safety data to prevent romantic roleplay, discouragement of age-restricted challenges, and boundaries against the model positioning itself as a substitute for real-world relationships
OpenAI commits to post-launch monitoring to investigate the divergence between offline evaluation results and online testing, and describes the whole U18 evaluation structure as 'directional rather than definitive'

Methodology Notes

Dated 2026-08-06. Production Benchmarks are drawn from challenging production-data conversations; dynamic mental-health benchmarks use adversarial user simulations that evolve in response to model outputs; safety evaluations run at the lowest reasoning setting to reflect majority usage. Verified by downloading the PDF from cdn.openai.com and reading the text directly (29 pages); benchmark numbers quoted from the table (six comparison columns: GPT-5.3 Instant through GPT-5.6 Luna August).