Claude Sonnet 3.5 (upgraded)

par Anthropic

AI Model

Faible confiance

The upgraded Claude 3.5 Sonnet is an enhanced version of the original Claude 3.5 Sonnet, released by Anthropic. It introduces computer use capabilities, allowing the model to interpret screenshots of a GUI and generate appropriate tool calls to perform tasks. This model sets new state-of-the-art standards in agentic coding (SWE-bench Verified), agentic task completion (TAU-bench), and computer use from screenshots (OSWorld). The upgraded Claude 3.5 Sonnet was trained on a proprietary mix of publicly available internet data (up to April 2024), private datasets, and internally generated data. It underwent RLHF (Reinforcement Learning from Human Feedback) and Constitutional AI fine-tuning to align with principles like helpfulness, honesty, and harmlessness. The model is multimodal (text and vision) and multilingual. Safety evaluations show strong alignment, with improved harm reduction and enhanced resistance to prompt injection. It was deployed under AI Safety Level 2 (ASL-2) protections.