Mar 2024
Claude 3 Family — Haiku, Sonnet, and Opus Set New Benchmarks
Anthropic launched the Claude 3 model family in three tiers — Haiku, Sonnet, and Opus — covering fast and cheap to most capable. Claude 3 Opus surpassed GPT-4 on most major benchmarks at launch, including MMLU, HumanEval coding, and GPQA graduate-level science, marking the first time a non-OpenAI model led broad academic evaluations across the board. All three Claude 3 models featured a 200,000-token context window as standard.
Why it mattersOffering three tiers at different speed-cost-quality tradeoffs let teams match model tier to task criticality, reducing inference costs by routing lightweight tasks to Haiku.
Try itRoute classification, summarization, and extraction tasks to Claude Haiku and reserve Opus-class models for reasoning-heavy steps to cut API costs by 5–10x on high-volume workloads.