OpenAI published a research preview of gpt-oss-safeguard, a family of open-reasoning models designed to classify content according to policies you define. They come in two sizes, gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, and are released under the Apache 2.0 license so anyone can use, modify, and deploy them.
What is gpt-oss-safeguard
It's a different approach from traditional classifiers. Instead of training a model to infer a policy from examples, gpt-oss-safeguard receives the policy at inference time and reasons about it to decide whether a message, a response, or a chat complies with that policy.
Sounds odd? Think of it as handing the model the rules of the game just before asking for a verdict. That makes the tool flexible: you can tweak the rules without retraining a whole classifier.
How it works in simple terms
The model takes two inputs: the policy (the definition of what you consider harmful or not) and the to evaluate.
content
It returns a conclusion about where the content falls and explains its reasoning (chain-of-thought), so you can review how it reached that decision.
The policy isn't embedded in the model; it's provided on each call. That lets you iterate quickly.
Advantages and use cases
Flexibility: ideal when risk changes fast or rules need frequent tweaking.
Nuanced judgments: it works well in complex domains where a small classifier might fail.
Low-data scenarios: useful if you don't have thousands of examples to train a robust classifier.
Concrete examples: a gaming community that wants to detect conversations about cheating, or a review site trying to flag potentially fake reviews using its own policy.
Evaluation and performance
OpenAI tested the models on internal checks and public benchmarks. On multi-policy metrics (evaluating an input under multiple policies at once), gpt-oss-safeguard and OpenAI's internal Safety Reasoner outperformed other models, even larger ones in some months of testing.
On datasets like the 2022 moderation evaluation and ToxicChat, results were solid, with variation depending on the dataset and the model size.
Important limitations
Classifiers trained with tens of thousands of high-quality examples can still beat gpt-oss-safeguard on certain complex risks. If you have lots of data and need top accuracy, a dedicated classifier may be better.
It requires more compute and can be slower. It's not the best choice to review all content in real time at scale if low latency is critical.
OpenAI mitigates this internally by using fast classifiers as a pre-filter or running the reasoning evaluation asynchronously when needed.
Why open-weight and Apache 2.0 matter
That the models are open-weight and under Apache 2.0 means you can download, inspect, adapt, and deploy the technology without heavy legal barriers. OpenAI published them to gather feedback and for the community—researchers, safety teams, and developers—to collaborate on open tools for protecting online spaces.
How developers can get started
The models are available for download, for example from Hugging Face. OpenAI worked with organizations like ROOST, SafetyKit, and Discord to test them and create documentation. There's also a community called ROOST Model Community to share practices and results.
If you want to try it: define a clear policy, test with real examples from your platform, and use gpt-oss-safeguard as a reasoning layer that complements fast classifiers.
Reflection
It's not a silver bullet, but it is a paradigm shift: moving from static classifiers to models that reason about rules you provide. That opens real possibilities for teams that need flexibility, explainability, and speed to adjust their safety policies.
