Definition
Claude Haiku is Anthropic's model line aimed at high-volume tasks where latency and cost strongly constrain model selection. Typical uses include classification, extraction, summarization, and short subtasks delegated by an agent. Haiku identifies a line; a name such as Claude Haiku 5.5 identifies a release within it.
Origin and attribution
Anthropic announced Haiku with the Claude 3 family on March 4, 2024, positioning it for quick responses and inexpensive work. The announcement said that Haiku would become available after the launch-day release of Opus and Sonnet. Anthropic receives the product attribution; the source does not name one inventor of the line.
Claude Haiku 5.5 launched on October 7, 2026. Anthropic describes it as its first Haiku model with adjustable effort. Its Claude API identifier is claude-haiku-5-5. Older Haiku releases have different capabilities, so the name alone does not establish support for that setting.
Scope and limits
Fast and inexpensive are relative claims. They depend on prompt length, effort, provider, token use, and the need for retries. Anthropic qualifies its Haiku 5.5 speed claim by standard serving speed; Opus in Fast Mode can generate tokens faster. Haiku 5.5 also has different price bands for prompts above and below 100,000 tokens.
Haiku's task positioning does not establish that it is suitable for every simple-looking decision. A short request can require difficult judgment, and an inexpensive wrong answer can cause an expensive downstream action. Anthropic's launch guidance favors Sonnet or Opus for more complex agentic coding.
Operational significance
Evaluate bounded tasks before assigning them to Haiku. For extraction, check whether the returned value matches the source; for routing, measure the cost of a misrouted request. A subagent's result should pass the same acceptance checks as work from the lead model.
Use an escalation rule when a result is incomplete or fails validation. Include that extra work when calculating cost and latency.
Distinguish it from nearby terms
- Claude is the overall family. Haiku is its line for workloads that place stronger constraints on cost and latency.
- Sonnet and Opus have different product positions and can be escalation targets; their names do not establish the best choice for every task.
- A subagent is a delegated agent role. Haiku can power that role, but delegation requires a harness and task boundary.
- Classification is a task. Haiku is a general-purpose model line that can perform it, rather than a dedicated classifier by definition.
Check your understanding
A Haiku subagent extracts payment instructions quickly but occasionally substitutes an account number. What validation and escalation rules would make that workflow acceptable, and how would they affect its measured cost?