Preparedness framework

OpenAI's framework for evaluating frontier models for capabilities that could cause severe harm, with High and Critical thresholds that require safeguards.

Why it matters. OpenAI's system cards report where a model stands against these thresholds, which explains extra safeguards in some domains.

How it works

The Preparedness Framework is OpenAI's process for tracking frontier model capabilities that could lead to severe harm and deciding which safeguards are needed. It focuses on tracked categories, such as biological and chemical, cybersecurity and AI self-improvement, and defines two capability thresholds: High and Critical.

A model that reaches High capability in a tracked category needs safeguards that sufficiently minimize the associated risks before it is deployed; Critical capability also requires safeguards during development. OpenAI's system cards report the evaluations and the determination for each category. Other makers have comparable frameworks, such as Google DeepMind's Frontier Safety Framework, which uses Critical Capability Levels.

For builders, a High determination in a category usually comes with extra safeguards that can affect some requests in that area. Compare determinations across makers with care, since categories, thresholds and tests differ.

Worked example

Illustrative: a system card states a model is treated as High capability in the biological and chemical category, with safeguards in place. A team building a lab-assistant tool should expect some requests there to be refused, and budget time to test its real prompts.

Common mistakes

Treating 'High' as a quality rating rather than a risk determination.
Assuming the framework never changes; check which version a system card refers to.
Comparing OpenAI's High with another maker's levels as if they were the same threshold.

See it on a real card: Safety & risk on Claude Opus 5.5.