Summary
This document is a system card that introduces Claude Haiku 4.5, a hybrid reasoning large language model developed by Anthropic. It outlines the model's characteristics, including its training data, safety evaluations, and performance metrics compared to its predecessor, Claude Haiku 3.5. The system card details the model's capabilities in coding tasks and computer use, emphasizing improvements in safety and alignment. It discusses the evaluation process under the AI Safety Level 2 Standard, highlighting assessments of the model’s safeguards, its behavior in agentic roles, and its potential for misuse. The document also describes the training process, including the use of a general-purpose web crawler and various data cleaning methods. Additionally, it covers the model's extended thinking mode, context awareness, and the role of crowd workers in enhancing model performance. Overall, the system card provides comprehensive insights into the development and safety measures associated with Claude Haiku 4.5.
