Claude 3

First Claude family with tiered model sizes: Opus (largest), Sonnet (balanced), and Haiku (fastest). All models support 200K token context and multimodal vision input. Architecture and parameters undisclosed.

Claude 3 Opus set new benchmarks for graduate-level reasoning (GPQA 50.4%), multilingual math (MGSM 90.7%), and was the first model to surpass GPT-4 on several evaluations. The tiered approach let users choose between intelligence, speed, and cost. AA Intelligence Index: 18 (Opus), 10 (Sonnet), 12 (Haiku). Proprietary.