Anthropic ships Claude Haiku 5.5, cutting costs 約75％

Claude Haiku 5.5 arrives as a small-model play rather than a frontier-model one: Anthropic frames it as its fastest model, designed for high-volume, cost-sensitive tasks. Alongside the release, Anthropic revised pricing across its lineup, halving the cache-read rate for Claude Sonnet 5.5 — which it says lowers agent-processing costs by about 20% — and adding new monthly API credits for Claude Max and Team subscribers to help them build agents and applications on the Claude platform. Together these moves suggest Anthropic is competing on the cost of running agents at scale, not just on raw capability.

The model also lands in a specific niche. It is aimed at fast, repetitive work such as summaries, compression, database queries and classification requests, and can serve as a coding sub-agent alongside Opus 5.5 and Sonnet 5.5. The Haiku 5.5 benchmarks were measured with OSWorld 2.1, which tests how well an agent operates a real computer to finish long, multi-step tasks. It is also the first Haiku-class model with adjustable processing settings, letting users choose between cost optimization and intelligence optimization.

The company's claim of roughly 75% lower average operating costs versus Haiku 4.5 is where the practical stakes sit. Whether that translates into real savings appears to hinge on how closely a given workload matches the summary, compression, query and classification patterns Anthropic describes, since the figure is an average rather than a per-task guarantee.