System cards and model cards

A system card is published alongside a model release and reports what testing found: capability evaluations, safety testing, identified risks and known limitations.

Why they are worth reading even if you are not a safety researcher

Three practical uses:

They date the release. This reference’s release date for Claude Opus 5 — 24 July 2026 — comes from its system card, because announcement posts are not always dated unambiguously.

They state limitations plainly. Marketing describes what a model does well. System cards describe where it fails, which is the half you need for a risk assessment.

They explain availability decisions. Why a model is narrowly released, or why a capability is restricted, is usually answered here rather than in the product pages.

Related but different

Model cards describe a model’s characteristics and intended use. System cards cover the deployed system, including safeguards around the model. Anthropic publishes both; the terms are sometimes used loosely in secondary coverage.

Knowledge cutoffs are documented in two forms too — reliable knowledge cutoff (the date through which knowledge is most extensive and reliable) and training data cutoff (the broader range). Those differ per model: Opus 5 is May 2026 for both, while Haiku 4.5 is February 2025 reliable against July 2025 training.

Where to find them

The Transparency Hub is the index. Individual cards are linked from model announcements.