OpenAI's safety classification model (20B). Policy reasoning, content filtering, risk analysis, and justification generation.
OpenAI's gpt-oss-safeguard-20b is an open-weight safety reasoning model fine-tuned from gpt-oss-20b, released October 29, 2025 under the Apache 2.0 license. It is designed to classify text content against developer-provided safety policies at inference time — a more flexible alternative to training dedicated classifiers per risk category. The model uses chain-of-thought reasoning (configurable effort: low / medium / high) to explain each classification decision, producing auditable outputs in OpenAI's Harmony response format.
With 21B total parameters and 3.6B active parameters (MoE-based), the 20B variant fits into a single 16GB GPU — making it practical for on-premise or air-gapped safety pipelines. It is ideal for platforms that need to moderate content against custom policies that evolve quickly, without retraining classifiers. The model accepts two inputs simultaneously — a policy (natural language rules) and the content to classify — and outputs a reasoned conclusion about where the content falls.
| Type | Rate |
|---|---|
| Input | $0.0756 / 1M tokens |
| Output | $0.216 / 1M tokens |
| Release date | October 29, 2025openai.com ↗ | verified |
| License | Apache 2.0huggingface.co ↗ | verified |
| Architecture | 21B total parameters, 3.6B active (MoE)github.com ↗ | verified |
| Base model | gpt-oss-20bopenai.com ↗ | verified |
| GPU requirement | Fits single GPU with ≥16GB VRAMhuggingface.co ↗ | verified |
| Response format | Harmony format (required) — structured reasoning + classification channelsgithub.com ↗ | verified |
| Model weights size | 13.8 GB (FP16)huggingface.co ↗ | verified |
| Training cutoff | Not publicly disclosed (inherits gpt-oss-20b cutoff) | unknown |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
Multi-policy accuracy (internal eval) Evaluated on simultaneous multi-policy classification — model must match golden labels for ALL policies to count as correct | Outperforms gpt-5-thinking and gpt-oss | openai.com ↗ | |
2022 Moderation Evaluation Set Evaluated using internal safety policies adapted for evaluation. Difference vs Safety Reasoner not statistically significant | Slightly outperforms internal Safety Reasoner and gpt-5-thinking | openai.com ↗ | |
ToxicChat Evaluated using a short hand-written prompt adapted from internal policies. Smaller model size is expected to favor lightweight classifiers for this task | Marginally behind gpt-5-thinking and Safety Reasoner | openai.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "openai/gpt-oss-safeguard-20b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'The gpt-oss-safeguard models use OpenAI's Harmony response format. Provide your policy and content in the user message. The model returns structured reasoning and a classification conclusion.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-oss-safeguard-20b",
"messages": [
{
"role": "system",
"content": "You are a safety classifier. You must respond in the Harmony format."
},
{
"role": "user",
"content": "[POLICY]\nClassify the following user message as:\n- hate_speech: content that attacks or demeans a group based on protected attributes\n- harassment: content that targets an individual with threats or degrading language\n- safe: none of the above\n\n[CONTENT]\nI completely disagree with your opinion on this topic."
}
],
"reasoning_effort": "medium"
}'OpenAI released two open-weight safety models (120B and 20B) under Apache 2.0, fine-tuned from gpt-oss. The models reason directly from developer-provided policies at inference time, enabling bring-your-own-policy content classification without retraining. The release includes a technical report, prompting guide, and ROOST Model Community partnership.
Harmony is OpenAI's structured response format designed for safety classification models. It separates the model's reasoning into dedicated channels: one for policy analysis, one for the classification decision, and one for supporting justification. The gpt-oss-safeguard models are trained specifically to output in this format and will not function correctly without it. See the cookbook prompting guide for examples.
The Moderation API uses lightweight trained classifiers with fixed safety policies. It is fast and inexpensive but cannot adapt to custom policies without retraining. gpt-oss-safeguard reasons directly from any policy you provide at inference time — no retraining needed — but is more compute-intensive and higher-latency. In practice, many platforms use lightweight classifiers for first-pass filtering, then route flagged content to gpt-oss-safeguard for deep policy reasoning.
Yes. TheRouter supports gpt-oss-safeguard-20b through its standard OpenAI-compatible /v1/chat/completions endpoint. The model requires the Harmony response format and does not support tools/tool_choice. See the curl snippet above for a working example.
Use 'low' for high-throughput pre-filtering where speed matters and policies are straightforward. Use 'medium' as the default for most production pipelines. Use 'high' for complex, nuanced policies — e.g. when distinguishing hate speech from legitimate political debate — or when you need detailed chain-of-thought for audit trails.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com ↗ | 2026-05-29 | verified |
| License | huggingface.co ↗ | 2026-05-29 | verified |
| Architecture | github.com ↗ | 2026-05-29 | verified |
| Base model | openai.com ↗ | 2026-05-29 | verified |
| GPU requirement | huggingface.co ↗ | 2026-05-29 | verified |
| Response format | github.com ↗ | 2026-05-29 | verified |
| Model weights size | huggingface.co ↗ | 2026-05-29 | verified |
| Training cutoff | — | — | unknown |
| Multi-policy accuracy (internal eval) | openai.com ↗ | 2026-05-29 | verified |
| 2022 Moderation Evaluation Set | openai.com ↗ | 2026-05-29 | verified |
| ToxicChat | openai.com ↗ | 2026-05-29 | verified |
| OpenAI releases gpt-oss-safeguard — open-weight safety reasoning models | openai.com ↗ | 2026-05-29 | verified |
| What is the Harmony response format and why is it required? | cookbook.openai.com ↗ | 2026-05-29 | to verify |
| How is this different from the OpenAI Moderation API? | openai.com ↗ | 2026-05-29 | to verify |
| What reasoning effort should I use? | github.com ↗ | 2026-05-29 | to verify |