gpt-oss-safeguard-120b and 20b
gpt-oss-safeguard, open-weight reasoning classifiers fine-tuned from gpt-oss, interpret a developer-written safety policy at inference time.
Rather than training a classifier on labeled data, the model reasons over a policy supplied at inference, so policies can change without retraining. Research preview under Apache 2.0, built with ROOST (a nonprofit trust-and-safety tooling group).
- Date
- Wednesday, 29 October 2025
- Lab
- OpenAI
- Kind
- open-weights
- Access
- open weights
- Price
- free; Apache 2.0
Sources
This record was checked against its sources on 6 October 2026. How we check