OpenAI publishes o3-mini system card detailing safety work

OpenAI has released the system card for its o3-mini model, documenting the safety work carried out before and around the model's release.

The report covers three main areas: safety evaluations, external red teaming, and evaluations under OpenAI's Preparedness Framework. That framework is the company's internal process for assessing whether a model's capabilities pose unacceptable risks before deployment.

System cards have become a standard part of OpenAI's release practice. They give regulators, enterprise customers, and researchers a view into what testing a model underwent and what risks the developer identified. For o3-mini, the card follows the same structure OpenAI used for earlier models in the o-series.

The stakes are concrete. o3-mini is a smaller, cost-efficient reasoning model aimed at high-volume use, which means its safety posture matters to a broad base of developers building on OpenAI's API. External red teaming, in which outside testers probe a model for harmful outputs and misuse vectors, is a key signal buyers and policymakers look for when judging whether deployment practices are keeping pace with capability gains.

By publishing the card, OpenAI continues a transparency practice that has drawn both praise and pressure: the company has faced calls to disclose more detail about its evaluation results, not just the existence of testing. How rivals respond — and whether system cards become an industry norm or remain an OpenAI-specific convention — will shape what disclosure the market comes to expect for every major model release.