GPT OSS Safeguard 120B

GPT OSS Safeguard 120B is OpenAI's first open-weight safety reasoning model, released in October 2025 as a research preview under the Apache 2.0 license. It is a fine-tuned variant of gpt-oss-120b — a 117B-parameter MoE transformer (5.1B active per forward pass) — specialized for policy-based content classification. Instead of learning a decision boundary from labelled examples like traditional moderation models, Safeguard reasons directly from a developer-provided policy at inference time, using chain-of-thought to produce explainable classifications.

In production, GPT OSS Safeguard 120B fills the role of a flexible, reasoning-based content moderation layer. Developers provide their own policies at inference time — for example, a gaming forum's anti-cheating rules or a review platform's fake-review policy — and the model classifies each input against those policies while outputting its reasoning. It excels in evolving-risk domains where policies must adapt quickly, nuanced scenarios where traditional classifiers struggle, and situations where explainable moderation decisions matter. According to OpenAI's internal evaluations, the safeguard model outperforms gpt-5-thinking on multi-policy accuracy despite being substantially smaller.