Claude Haiku 4.5

Haiku 4.5 is the volume tier — the model for work measured in millions of calls rather than hundreds. Classification, extraction, routing, tagging, first-pass filtering: tasks where the job is well-defined and the model does not need to be clever, just consistent and fast.

It is easy to underrate because it is not the model you would use to write something you care about. That is not what it is for.

Speed and cost, and they compound. At Haiku prices a pipeline can afford to call the model at every stage rather than batching to save money, which usually produces a better system than a smarter model used sparingly.

It is also well suited to the unglamorous middle of a pipeline: deciding which of five downstream paths a request should take, extracting structured fields from messy text, or checking whether a document is relevant before an expensive model reads it. Used as a filter in front of a flagship, Haiku often pays for the whole system.

It is not a reasoning model. Give it a problem with several interacting constraints and it will produce something plausible and wrong. It is also weaker at long context — it will accept a large document and answer confidently from the parts it attended to.

Writing quality is functional rather than good. Anything a customer reads should be drafted by something else.

Use Haiku 4.5 for high-volume, well-specified tasks, and as the routing and filtering layer in front of larger models. The architecture that works is a cheap model deciding what deserves an expensive one — not a single tier doing everything.