OpenAI o1 System Card

Source: OpenAI Date: 2024-12-05 URL: https://openai.com/index/openai-o1-system-card

Summary

Full system card for o1 (non-preview), published December 2024 alongside the o1 general availability launch and the developer tools update. The system card covers the extended safety evaluations performed on o1 relative to the September preview — including assessments of autonomous replication risk, cyberoffense capability, and CBRN uplift potential. Crucially, this system card introduced the “preparedness” framework evaluations that would become the standard safety documentation format for subsequent OpenAI model releases.

Implications

Preparedness framework as the new safety standard. The o1 system card is the first to systematically apply the preparedness framework categories (CBRN, cyberoffense, persuasion, autonomous replication) to a production model release. It sets the documentation template for every subsequent OpenAI model card through GPT-5.5. The specific risk category choices reflect OpenAI’s view of which AI risks are tractably measurable.

The “medium risk” calibration. The o1 system card’s risk assessments in several categories came back as “medium” rather than “low” or “high.” Medium is the threshold above which the preparedness framework says not to release. OpenAI releasing despite “medium” assessments reflects their judgment that the benefits outweigh the risks at that level — a consequentialist tradeoff that safety researchers will continue to scrutinize.

Thread: o1 documentation. Read alongside the September 2024 preview system card, the chain-of-thought monitoring paper (March 2025), and the o3 system card (April 2025) as the o-series safety documentation arc.

Watch: Whether the preparedness framework’s medium/high/critical risk thresholds remain stable through subsequent model releases, or whether the calibration drifts as models become more capable and the risk tolerance implicitly rises.