Claude Haiku 4.5 vs Claude Sonnet 5

Haiku 4.5 costs exactly half of Sonnet 5 - $1 and $5 per million tokens against $2 and $10 - and the naming invites you to read it as the small model in the same family. It is not. Haiku 4.5 belongs to the previous generation and the gaps are wide: a 200K context window against Sonnet 5's 1M, a 64K maximum output against 128K, and a reliable knowledge cutoff of February 2025 against January 2026, nearly a year further back. It also uses the older extended-thinking mode rather than the adaptive thinking the Claude 5 models use, and it does not accept the effort parameter at all. Haiku 4.5 is the right choice for high-volume classification, extraction and routing, where latency and unit cost dominate and the knowledge gap is irrelevant. For anything reasoning over a long document or needing current knowledge, the 2x saving is a false economy. One planning note: Anthropic's retirement commitment for Haiku 4.5 runs only to 15 October 2026, the earliest date in the current lineup.