Haiku 4.5 – you'd be amazed if you gave it a chance (barazany.dev)

Anthropic quietly released Haiku 4.5 — a smaller, cheaper model that claims Sonnet 4–level capabilities at roughly one-third the cost and 2–3x the speed. On SWE-bench it slightly outperformed Sonnet 4, and in real-world use the most noticeable win is latency: responses are fast enough to change developer workflows, removing the hesitation around tool use. The author reports Haiku 4.5 matches Sonnet 4’s quality for coding, reasoning, and tool use (with one notable gap in Hebrew), and has become the default in Claude Code for everyday work. Why this matters: Haiku 4.5 exemplifies a shifting frontier where smaller, cheaper models hit the practical performance bar for many tasks. Faster inference plus Sonnet-level competence means better UX, lower inference cost, and higher throughput for coding agents and interactive tools — often more impactful than marginal accuracy gains from flagship model upgrades. For ML practitioners and product teams, the takeaway is to re-evaluate model/agent trade-offs: optimized prompts and agent design still drive big gains, but cheaper, lower-latency models can materially change adoption and UX for real-world workflows, while larger models remain relevant for edge cases and specialized languages.