OpenAI has released two open-weight models built to apply content policies on demand, not just observe the ones they were trained with.
The models, gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, are post-trained from OpenAI's existing gpt-oss open-weight base models. The defining design choice is that neither model has content rules hard-coded in. Instead, you hand it a policy document and it reasons through that document to classify content against it. A technical report published alongside the release covers baseline safety evaluations, benchmarking the safeguard variants against the underlying gpt-oss models.
That policy-as-input structure matters because most moderation classifiers are static artifacts: trained on a fixed rule set, fragile when the rules change. A model that reads the rulebook on each call could let platforms update their standards without retraining a classifier from scratch. For developers navigating jurisdiction-specific legal requirements or community norms that shift over time, that flexibility is worth something concrete.
OpenAI's open-weight strategy is recent enough that releasing production-oriented safety variants signals the company wants these models in real deployments, not just cited in papers. Whether a 120-billion-parameter moderation layer is actually practical for most platforms is a question the technical report does not answer.