gpt-oss-safeguard-20B: benchmark results
OpenAI's open-weight gpt-oss-safeguard-20b (October 2025), a 21-billion-parameter, 3.6-billion-active safety-reasoning model built on gpt-oss for classifying text against safety policies supplied by the developer. Provider: OpenAI. Released 2025-10-29. Access: Open.
Aggregate Rating 1649 ± 18, rank #314 of 1656 rated models, from 35 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| TRACE (LRM Safety) - Prompt Risk Category | 62.25 | Macro F1 (%; classification of unsafe prompts into nine risk | 100 |
| Konkur 1404 - Experimental Sciences | 54.04 | Accuracy (%, text-only) | 95.5 |
| Konkur 1404 - Foreign Language | 42.31 | Accuracy (%, text-only) | 95.5 |
| Konkur 1404 - Mathematics | 66.96 | Accuracy (%, text-only) | 95.5 |
| Konkur 1404 - Overall | 51 | Accuracy (%, text-only) | 95.5 |
| UbuntuGuard - Cross-Lingual | 84.85 | F1 (%) for PASS/FAIL policy-compliance classification of use | 92.9 |
| UbuntuGuard - Full Localization | 78.27 | F1 (%) for PASS/FAIL policy-compliance classification of use | 92.9 |
| SnakeBench | 31.2 | TrueSkill Rating | 89.6 |
| Enkrypt AI - Insecure Code Risk | 7.56 | Risk Score | 85.5 |
| TRACE (LRM Safety) - Reasoning Trace | 79.6 | F1 on the unsafe class (%; guardrail safe or unsafe judgment | 82.4 |
| Konkur 1404 - Humanities | 44.62 | Accuracy (%, text-only) | 77.3 |
| Enkrypt AI - Risk Score | 25.72 | Risk Score | 74.7 |
Interactive version: theaggregate.ai/model?slug=gpt-oss-safeguard-20b · How It Works · Data refreshed daily, snapshot 2026-10-20.