OpenAI: gpt-oss-safeguard-20b

OpenAI gpt-oss-safeguard-20b is a safety reasoning model from OpenAI, built on top of gpt-oss-20b. It is an open-weight model with 21B parameters in a Mixture-of-Experts (MoE) architecture, and OpenAI positions it for safety related work such as content classification, LLM filtering and trust oriented tasks. It was released by the provider on 2025-10-29 and sits in the text category on api14.

The model takes text as input and returns text as output. It has a context window of 131072 tokens and a maximum output of 65536 tokens, which leaves plenty of room for long inputs alongside a substantial generated response. Because it is a reasoning model, it can expose its reasoning behaviour, and the supported parameters include include_reasoning and reasoning for that purpose. It is built on a Mixture-of-Experts design, which OpenAI describes as offering lower latency for safety tasks.

Those characteristics point to a fairly specific set of uses. Content classification is the clearest one: sending text and asking for a category or label. LLM filtering follows from the same idea, where the model judges whether another model's input or output should pass through. Trust and safety workflows, including moderation style checks and policy oriented review, fit the description OpenAI gives. The long context window means you can pass large documents, transcripts or batches of content in a single request rather than chunking them first, and the large output ceiling supports detailed labels, explanations or structured results.

You call the model through api14's OpenAI-compatible API using the model id "openai/gpt-oss-safeguard-20b". Because the interface is OpenAI-compatible, existing OpenAI client libraries and request shapes work with only the base URL and model id changed. The supported request parameters are include_reasoning, max_tokens, reasoning, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools and top_p. In practice that means you can control sampling with temperature and top_p, cap output with max_tokens, use stop sequences, make runs reproducible with seed, request structured or JSON style output through response_format and structured_outputs, and give the model tools with tools and tool_choice. The reasoning and include_reasoning parameters let you decide whether reasoning content is surfaced in the response.

There are no stated limitations in the facts beyond the modality and context bounds, so the main thing to weigh is fit. Input and output are text only, so if you need image or audio handling this model is not the right choice. If your task is general conversation or generation rather than safety oriented classification and filtering, a general purpose model may suit you better. For classification, filtering and trust work where you want reasoning exposed and a long context, this model is a reasonable candidate to test.