Claude Haiku 5.5 launches at ~75% lower cost than Haiku 4.5, built for subagents

Anthropic launched Claude Haiku 5.5, the fastest and most efficient model in the Claude 5.5 family and the third 5.5 release in roughly a month, following Opus 5.5 and Sonnet 5.5. It costs about 75% less than Claude Haiku 4.5 on average, with savings reaching 90% on prompts under 100,000 tokens, and delivers up to 2.5x faster inference. It shipped simultaneously on Amazon Bedrock and Claude Platform on AWS.

Anthropic positions Haiku 5.5 for subagents and high-volume, cost-sensitive work: classification, summarization, extraction and live customer support. The subagent framing is key. Modern agent systems have an orchestrator model, often Opus, delegating searches, file reads and simple checks to many cheaper workers. When dozens of subagent calls happen per task, a 75% price cut on the worker model changes total costs far more than a cut to the orchestrator would.

The pricing structure has drawn mixed reactions. On Hacker News, where the launch hit 1,034 points and 482 comments, developers praised bundled API credits for Max and Team plans, but some called the 100K-token pricing cliff a problem for agent workloads, which frequently exceed that context size. Competitively, Haiku 5.5 squares off against OpenAI's Luna and Google's Flash-class models, and against open alternatives including Zhipu's GLM-5.3 and DeepSeek V4 Flash that now run on the same clouds. Reuters framed the release as part of Anthropic expanding its lineup ahead of a planned IPO.

Anthropic also expanded its Cyber Verification Program into Defense, Red Team and Specialized tiers in the same window, giving verified security professionals access to Opus 5.5, Sonnet 5.5 and Mythos 5.1. Watch for independent benchmarks comparing Haiku 5.5 to rivals at similar price points, real-world subagent cost reports, and whether Anthropic adjusts the pricing threshold in response to developer complaints.