OpenAI on Wednesday announced two new artificial intelligence reasoning models — gpt-oss-safeguard-120b and gpt-oss-safeguard-20b — designed to help developers detect and classify online safety harms on their platforms.
The models, introduced as open-weight systems, mark one of OpenAI’s most significant steps yet in improving transparency and safety across digital ecosystems increasingly shaped by generative AI.
According to the company, the models are fine-tuned versions of its gpt-oss family, first announced in August. Their numerical names reflect their size, with the 120-billion and 20-billion parameter variants offering different levels of computational power and precision. Both are built to enable developers to implement custom moderation frameworks or adapt the models to specific content policies.
Register for Tekedia Mini-MBA edition 19 (Feb 9 – May 2, 2026): big discounts for early bird.
Tekedia AI in Business Masterclass opens registrations.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab: From Technical Design to Deployment (next edition begins Jan 24 2026).
Unlike open-source systems, where the full code is available for public modification, open-weight models make only their trained parameters accessible. This approach, OpenAI said, allows developers to study and deploy the systems while maintaining integrity over their design and operation.
These models can be configured to reflect an organization’s unique policy needs, because they are reasoning models that show their work, developers can gain clearer insight into how an output was produced — and why.
Practical Applications Across Platforms
The company illustrated several possible use cases for the new safeguard models. A product review site, for instance, could use them to automatically detect fake or manipulated reviews, improving the credibility of its listings. Likewise, a video game forum could deploy the models to classify or flag posts discussing cheating or game exploits.
The models are designed to reason through nuanced policy definitions rather than rely solely on keyword matching or predefined rule sets — a capability that could make moderation more context-aware and less error-prone.
The models were co-developed with Robust Open Online Safety Tools (ROOST), an organization that builds open frameworks for digital safety research. OpenAI said the collaboration with ROOST reflects its renewed focus on making AI safety tools accessible to the broader community, not just to major corporations or regulators.
ROOST was joined in testing by partners including Discord and SafetyKit, both of which provided feedback during the models’ pilot phase. The models are currently being released in research preview, and OpenAI said it will gather insights from safety experts and academic researchers before expanding access.
“As AI becomes more powerful, safety tools and fundamental safety research must evolve just as fast — and they must be accessible to everyone,” said Camille François, President of ROOST.
In tandem with the model release, ROOST is launching a model community for researchers and practitioners who work on using AI to secure online spaces. The group will focus on sharing best practices, developing open benchmarks, and identifying weaknesses in emerging AI systems.
The rollout of the safeguard models comes amid heightened scrutiny of OpenAI’s commercial growth and safety posture. The company, now valued at $500 billion, has faced criticism from some members of the research community who argue that it has expanded too rapidly without fully addressing ethical or security risks tied to powerful AI systems.
OpenAI is thus signaling a willingness to prioritize transparency and invite external oversight by making these safety models open-weight — a move analysts see as an attempt to rebuild trust among regulators and researchers.
OpenAI’s flagship product, ChatGPT, has become the most widely used AI chatbot globally, with more than 800 million weekly active users. The company’s scale has made its safety architecture a focal point of debate among policymakers and technology ethicists alike.
A New Chapter After Recapitalization
The announcement also follows OpenAI’s completion of a recapitalization process on Tuesday, which officially cemented its structure as a nonprofit-controlled organization with a majority stake in its for-profit arm.
The company, originally founded as a nonprofit research lab in 2015, has since transformed into the most valuable U.S. startup, driven largely by the success of ChatGPT and enterprise adoption of its GPT models.
OpenAI confirmed that eligible researchers and developers can download the model weights on Hugging Face, a leading AI repository and collaboration platform. The research preview phase will focus on gathering feedback from the online safety community, particularly around model interpretability, bias detection, and effectiveness across different languages and cultural contexts.
With the new safeguard models, OpenAI is effectively extending its AI portfolio into the domain of digital safety infrastructure, an area increasingly seen as critical to the responsible use of artificial intelligence.