Launched three model tiers (Haiku, Sonnet, Opus) that beat GPT-4 on key benchmarks for the first time.
Product AnnouncementLaunched three models: Haiku (fast/cheap), Sonnet (balanced), Opus (most intelligent). Opus outperformed peers on MMLU, GPQA, GSM8K. Introduced vision capabilities. Available in 159 countries. First model family with character training.
Instead of competing with a single model, Anthropic released three variants optimized for different tradeoffs: Haiku (fast, cheap, 3B-scale), Sonnet (balanced capability and speed), and Opus (most capable, higher cost). This tiered approach covered the entire market spectrum and became the industry standard, with OpenAI, Google, and others adopting similar three-tier strategies.
Claude 3 Opus topped major academic benchmarks including MMLU (multiple-choice reasoning), GPQA (PhD-level questions), and GSM8K (mathematical reasoning). This performance established Anthropic as a genuine frontier capability competitor, not just a safety-focused niche player. However, margins were thin and benchmark-specific — real-world superiority was less clear.
The Claude 3 family introduced vision capabilities enabling the model to process and analyze images. This opened new use cases and made Claude competitive with GPT-4V. Vision combined with frontier language understanding created a more versatile tool, though vision performance initially lagged on some specialized tasks.
The fundamental tension between model capability and inference cost. Haiku is fast and cheap but limited; Opus is powerful but expensive. This tradeoff is not easily resolved — you generally cannot have both maximum capability and minimum cost. The three-tier strategy acknowledges this by offering choices rather than solving it.