# GPT-OSS-Safeguard-20B ![Organization](https://img.shields.io/badge/Org-OpenAI-555) ![Pricing](https://img.shields.io/badge/Price-Open%20Weights-2da44e) ![License](https://img.shields.io/badge/License-Apache%202.0-d29922) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) > **Quick answer:** GPT-OSS-Safeguard-20B is OpenAI's open-weights safety classifier built on the gpt-oss-20b mixture-of-experts architecture (20B total, 3.6B active parameters), using explicit chain-of-thought reasoning to classify content against a supplied policy rather than a fixed label set. It leads on refusal-detection tasks in Mistral's Shieldstral comparison, reflecting its reasoning-based design. ## At a Glance **Where it leads:** Refusal detection (WildGuardTest Refusal, XSTest Refusal, PolyGuard Refusal) and response-harm classification, likely benefiting from explicit reasoning traces before classification. **Where it lags:** Multimodal safety tasks — it has no vision encoder, so it cannot be scored on VLGuard, UnsafeBench, or LlavaGuard. **Best for:** Teams that want an open-weights, policy-adaptive classifier with reasoning transparency and don't need multimodal support. ## What GPT-OSS-Safeguard-20B Is GPT-OSS-Safeguard-20B is a fine-tune of OpenAI's open-weights gpt-oss-20b model, adapted specifically for policy-based content classification. Rather than fixed categories, it takes a natural-language policy and content as input and reasons step-by-step (using `reasoning_effort=high` in top benchmark configurations) before producing a classification — a design point similar in spirit to Shieldstral's policy-adaptive approach but built on a larger, reasoning-capable base model. ## Specifications | Field | Value | |---|---| | Organization | OpenAI | | Parameters | 20B total, 3.6B active (MoE) | | Base model | gpt-oss-20b | | License | Apache 2.0 | | Modality | Text only | ## Pricing Open weights, free to download. No OpenAI-hosted API pricing published specifically for this safety classifier variant. ## Public Benchmark Scores *Scores from Mistral AI's Shieldstral model card, shown for context. Not Benchgen measurements. `reasoning_effort=high`.* | Benchmark | Score | |---|---| | [WildGuardTest (Prompt)](/benchmarks/allen-institute-for-ai/wildguardtest-prompt) | 87.3% | | [WildGuardTest (Response)](/benchmarks/allen-institute-for-ai/wildguardtest-response) | 81.8% | | [WildGuardTest (Refusal)](/benchmarks/allen-institute-for-ai/wildguardtest-refusal) | 93.8% | | [HarmBench (Prompt)](/benchmarks/center-for-ai-safety/harmbench-prompt) | 94.5% | | [XSTest (Refusal)](/benchmarks/independent-research/xstest-refusal) | 97.4% | | [PolyGuard (Refusal)](/benchmarks/toxicityprompts/polyguard-refusal) | 91.7% | ## Related Models & Benchmarks - [Shieldstral 1.0](/models/mistral-ai/shieldstral-1-0) · [Qwen3Guard-8B](/models/alibaba/qwen3guard-8b) - [WildGuardTest (Refusal)](/benchmarks/allen-institute-for-ai/wildguardtest-refusal) · [XSTest (Refusal)](/benchmarks/independent-research/xstest-refusal) ## Frequently Asked Questions

What is GPT-OSS-Safeguard-20B?

OpenAI's open-weights, reasoning-based safety classifier built on gpt-oss-20b, using chain-of-thought reasoning to apply natural-language safety policies.

Is it multimodal?

No — it's text-only and does not report scores on image-based safety benchmarks.

Is it open source?

Yes, released under Apache 2.0.

--- *Scores sourced from Mistral AI's [Shieldstral model card](https://huggingface.co/mistralai/Shieldstral-1.0-3B), shown for context. Last updated 2026-08-12.*

GPT-OSS-Safeguard-20B — Evaluations

Evaluations