OpenAI o3-mini System Card

Source: OpenAI Date: 2025-01-31 URL: https://openai.com/index/o3-mini-system-card

Summary

The safety and capability documentation accompanying the o3-mini launch, covering red-teaming results, refusal behavior, dangerous capability evaluations (bio uplift, CBRN, cyberoffense), and the reasoning model-specific risk surface. Published alongside the o3-mini model announcement, the system card reflects OpenAI’s now-standard practice of shipping safety documentation with every major model release — a practice that began with GPT-4’s system card in March 2023.

Implications

The system card as safety norm. By January 2025, every major OpenAI model release ships with a system card. This practice, which OpenAI pioneered, has become a de facto industry standard — Anthropic publishes model cards, Google publishes Gemini safety reports. The o3-mini card is notable because reasoning models have a qualitatively different risk profile: extended chain-of-thought can reason toward harmful outputs in ways that single-shot generation doesn’t.

Reasoning model evaluation gap. The system card for a reasoning model must grapple with the fact that the model’s intermediate reasoning steps are not easily audited. This is a known evaluation challenge — evaluators test outputs, not chains of thought. The o3-mini system card’s approach to this gap sets precedent for how the industry evaluates reasoning models going forward, including DeepSeek R1 and later o3-level systems.