A small model designed for lots of calls
Anthropic introduced Claude Haiku 5.5 on October 7 as the fastest and most efficient model in its Claude 5.5 family. The company positions it for high-volume, cost-sensitive work such as summaries, classification, routing, live support and small coding changes. At prompts up to 100,000 tokens, the API price is $0.10 per million input tokens and $0.50 per million output tokens—90% below Haiku 4.5 at that prompt length. Anthropic says the average cost per task is about 75% lower after accounting for its updated tokenizer; longer prompts receive a 50% price reduction.
Haiku 5.5 is available in Claude, the API, Claude Code and major cloud platforms. Anthropic says it is the first Haiku with effort controls, letting teams trade some model effort for cost. Its product thesis is to use a small fast model for bounded subtasks while a stronger model plans or handles ambiguous work.
Customer evidence points to subagents
Anthropic’s customer examples test different things. AlphaSense says Haiku 5.5 improved on Haiku 4.5 in a 400-query document test, scoring 0.84 versus 0.76. Box reports an 11-point gain over Haiku 4.5 at about half the latency in its early tests. Asana reports over 30% lower task latency and up to 2.5× faster inference per agent turn in its own evaluation. These are customer-reported results, not one shared independent benchmark.
One concrete pattern is a larger model building a deck while Haiku extracts a revenue line from a long filing, or a quick subagent classifying live requests. Lower token prices can make parallel or background agents more practical when tasks have clear success criteria and their output can be checked.
Where to place the boundary
Anthropic describes Haiku 5.5 as strong for repetitive browser actions such as form filling and data entry, and for focused coding edits. That does not mean every task should be routed to it. Difficult reasoning, architecture and knowledge work may still need a larger model; unclear instructions can turn a cheap run into several retries.
A useful deployment evaluates cost per accepted task, latency and error rate—not token price alone. If a small model finishes a well-scoped step quickly and a more capable model reviews exceptions, the system can spend its budget where judgment matters most. Haiku 5.5’s launch is another sign that model choice is becoming part of agent design.
Source published 2026-10-07. Coverage is based on the maker’s announcement and demonstration.