Claude 3.5 Haiku

Claude 3.5 Haiku (API name claude-3-5-haiku-20241022) was announced on 22 October 2024 and reached the API on 4 November. Anthropic's claim for it was blunt: at roughly the speed of the previous Haiku, it surpassed Claude 3 Opus — the largest model of the preceding generation — on many intelligence benchmarks. Its headline figure was 40.6% on SWE-bench Verified, a real-world software-engineering score that beat a number of agents built on the state-of-the-art models publicly available at the time. A model priced for high-volume background work was suddenly competitive on the benchmark the industry used to judge frontier coding. That capability had a visible price. Anthropic launched it at $1 per million input tokens and $5 per million output — four times the rate of Claude 3 Haiku — and said explicitly that the increase reflected the jump in intelligence. The market disagreed, and on 3 December 2024 the company revised the price down to $0.80 per million input and $4 per million output. The episode is worth recording because it is one of the few public instances of a frontier lab walking back a price increase within weeks of launch. The model shipped text-only, with image input announced as coming later — unusual in a generation where vision had become standard, and a reminder that Haiku-class models were positioned for latency and cost rather than feature completeness. Its natural uses were user-facing products with tight response budgets, specialised sub-agents inside larger systems, and grinding through large volumes of structured records. Anthropic notified developers on 19 December 2025 and retired the model on 19 February 2026. The recommended replacement is claude-haiku-4-5.

▸Videos

No videos yet.

▸Reviews