OpenAI o3 and o4-mini System Card

Source: OpenAI Date: 2025-04-16 URL: https://openai.com/index/o3-o4-mini-system-card

Summary

Summary

The system card for o3 and o4-mini documents safety evaluations, capability assessments, and risk mitigations for OpenAI’s April 2025 reasoning model releases. O3 represented a significant capability jump over o1/o3-mini in complex reasoning tasks; o4-mini was a smaller model optimized for cost-efficient chain-of-thought reasoning. The system card covered evaluations on dangerous capabilities (biological, chemical, radiological, nuclear), autonomous replication, and persuasion.

Implications

Safety/alignment thread. O3 and o4-mini system cards are significant because they represent the first o-series generation where OpenAI’s evaluations found meaningful evidence of uplift in dangerous-capability domains — shifting the language in the cards from “low risk” to careful quantified assessments. The system card process itself is under scrutiny at this point (April 2025): OpenAI’s Safety Advisory Group had flagged concerns about the adequacy of pre-deployment evaluations, and the cards are both a transparency mechanism and a liability management document. The capability-safety race dynamic is visible in the text: the models are more capable and the safety evaluations are more extensive, but the gap between what can be assessed and what might emerge in deployment persists.