Claude Sonnet 4.6 System Card (www.anthropic.com)

Anthropic has announced the release of Claude Sonnet 4.6, the latest iteration of its large language model, which improves upon its predecessor in several key areas, including coding, reasoning, and multimodal capabilities. Evaluations indicate that Sonnet 4.6 approaches or matches the skill levels of the more advanced Claude Opus 4.6, marking a significant step forward in its coding and mathematical abilities, along with specialized skills in finance, cybersecurity, and life sciences. Notably, the model exhibits improved alignment assessments, showing lower levels of misaligned behavior compared to earlier models, which is vital for ensuring responsible AI deployment. The deployment of Claude Sonnet 4.6 falls under the AI Safety Level 3 (ASL-3) Standard, indicating that it has passed rigorous safety evaluations designed to mitigate potential risks associated with autonomous operations. The model was trained on a diverse dataset and incorporates unique features such as "extended thinking mode," allowing for adaptive reasoning depending on task complexity. These enhancements not only contribute to broader advancements in AI safety and alignment practices but also emphasize Anthropic's commitment to ethical development through responsible scalability in AI technologies.