Claude Haiku 4.5 Review

Small models used to be a compromise you accepted for price, and Haiku 4.5 is the point where that stopped being true for most work. On the tasks that make up the bulk of real usage, classification, extraction, summarising, routing, transforming one format into another, it is indistinguishable from models several times its cost, and it answers in a fraction of the time. It handles tool use cleanly and follows structured output instructions without the drift that used to make small models unreliable in a pipeline.

Where it shows its size is on long chains of reasoning and on genuinely hard code: it will produce something plausible rather than pause, and a difficult debugging session exposes the difference against the frontier models quickly. The right way to use it is as the workhorse tier, with escalation to a larger model when a task proves it needs one. Priced per token rather than per seat, it is by far the cheapest capable model in the catalogue.