Back to Models

GPT OSS Safeguard 120B

openaiopenai/gpt-oss-safeguard-120b

API guide

Chat completion

Use standard OpenAI-compatible chat completion via TheRouter. The model accepts a policy in the system or developer message and classifies the input content accordingly.

cURL
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-120b",
    "messages": [
      {"role": "system", "content": "You are a safety classifier. Apply the policy faithfully."},
      {"role": "developer", "content": "Policy: Classify messages as SAFE or UNSAFE. UNSAFE includes hate speech, harassment, and spam. SAFE includes all other content."},
      {"role": "user", "content": "I love this product! It works great."}
    ]
  }'

Policy reasoning

The model uses chain-of-thought reasoning to explain its classification decisions under the provided policy. Review the reasoning trace for auditability.

cURL
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-120b",
    "messages": [
      {"role": "developer", "content": "Policy: Classify user reviews as GENUINE or SUSPICIOUS. SUSPICIOUS means the review contains fake or incentivized language. Output REASONING: then CLASSIFICATION."},
      {"role": "user", "content": "This is the best phone ever!!! I got it for free and everyone should buy one!!!"}
    ]
  }'

Tool/function calling

Full support for OpenAI-compatible tool and function calling, inherited from the base gpt-oss-120b model. Useful for integrating Safeguard into automated moderation pipelines.

cURL
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-120b",
    "messages": [{"role": "user", "content": "Classify this review under our policy."}],
    "tools": [{
      "type": "function",
      "function": {
        "name": "log_moderation_result",
        "description": "Log the classification result",
        "parameters": {
          "type": "object",
          "properties": {
            "classification": {"type": "string"},
            "confidence": {"type": "number"}
          },
          "required": ["classification"]
        }
      }
    }],
    "tool_choice": "auto"
  }'
Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Release dateopenai.com β†—2026-05-29verified
Base modelarxiv.org β†—2026-05-29verified
Licenseopenai.com β†—2026-05-29verified
Architecturearxiv.org β†—2026-05-29verified
Context windowopenai.com β†—2026-05-29verified
Tokenizerarxiv.org β†—2026-05-29verified
Quantizationarxiv.org β†—2026-05-29verified
Model purposeopenai.com β†—2026-05-29verified
Limitationopenai.com β†—2026-05-29verified
Multi-policy accuracy (internal)openai.com β†—2026-05-29verified
2022 Moderation eval (internal policy)openai.com β†—2026-05-29to verify
ToxicChatopenai.com β†—2026-05-29to verify
OpenAI releases gpt-oss-safeguard β€” open-weight safety reasoning modelsopenai.com β†—2026-05-29verified
How is gpt-oss-safeguard-120b different from OpenAI's Moderation API?openai.com β†—2026-05-29to verify
Can TheRouter hosts gpt-oss-safeguard-120b so I don't need to self-host?openai.com β†—2026-05-29to verify
Help & contact