Claude Haiku Glossary

The fast-and-cheap tier of Anthropic's *Claude* family. The default for high-volume routine tasks where latency and cost matter more than marginal capability.

Claude Haiku is the smallest and fastest tier of Anthropic’s Claude model family. Where Opus prioritises capability and Sonnet balances capability and cost, Haiku is optimised for speed and price — typical token costs are roughly one-twelfth of Opus, with latencies low enough for genuinely interactive use cases (chat, real-time tool selection, sub-second classification). Major versions include Haiku 3 (March 2024), Haiku 3.5 (November 2024), Haiku 4, 4.5, 4.6, and 4.7.

For this Dictionary, Haiku is the high-volume routine-work tier. Use cases that the operator’s workflows actually route to Haiku: image OCR triage, low-stakes classification, the first-pass annotation that gets refined by a higher tier later, vision tasks where the operator only needs a quick description, and sub-agent dispatch decisions where the model’s job is to pick a tool, not to think hard. The Dictionary’s F (Fast & Cheap) query-mode in AGENTS.md explicitly routes to Haiku for vision and reading tasks.

Haiku’s capability ceiling is meaningfully below the frontier — operators who try to do agentic writing or complex coding with Haiku as the primary engine will be disappointed. But for high-volume routine work, the cost differential makes it the rational choice.

See also

Return to Dictionary All Entries (A–Z) For Students Other Writing