Claude 3 Haiku

Claude 3 Haiku (API name claude-3-haiku-20240307) was announced on 4 March 2024 as the smallest of the three Claude 3 models and reached general availability on 13 March. Anthropic's pitch was speed at a price: at $0.25 per million input tokens and $1.25 per million output it was, by the company's account, the fastest and most cost-effective model in its intelligence class, and the illustration used was concrete — it could read an information-dense arXiv paper of roughly 10,000 tokens, charts and graphs included, in under three seconds. What made it matter beyond price was that Anthropic did not strip it down to get there. Haiku shipped with the same 200,000-token context window as Opus and Sonnet, and with the same image input, at a moment when cheap models from most vendors were text-only and short-context. That combination — vision, long context, near-instant answers, a quarter of a dollar per million tokens — made it the default engine for classification, extraction, moderation and routing work that never appears in a demo but accounts for most of the tokens a production system actually spends. The Claude 3 generation was also where Anthropic addressed a specific complaint about its earlier models: unnecessary refusals. All three models, Haiku included, were markedly less likely to decline prompts that merely brushed against the guardrails without being harmful. It ended up the longest-serving model Anthropic has shipped. Deprecation was announced on 19 February 2026 and the model was switched off on 20 April 2026 — a working life of just over two years, in a field where a year is now a full generation. The recommended replacement is claude-haiku-4-5.

▸Videos

No videos yet.

▸Reviews