The fastest, cheapest Claude. Near Sonnet-4 quality for high-volume, latency-sensitive work.
Every model family needs a workhorse, and Haiku 4.5 is Anthropic's. At $1 input and $5 output per 1M tokens it costs a fifth of Opus 4.8 and a third of Sonnet 4.6, while keeping the same 200K context window — which is the unusual part. Most budget models cut the window first; Haiku keeps it, so the cheap tier can still read the same long documents the expensive tiers do.
The quality positioning is honest: near Sonnet-4 level — the 2025 Sonnet, not the current 4.6. For classification, extraction, routing, summarization, and templated generation, that's more than enough, and the speed is the actual product.
Haiku 4.5 handles structured tasks reliably: tagging and classification at volume, JSON extraction from messy text, first-pass document triage, chat where sub-second response matters more than depth. It supports vision and the standard Claude tool-use surface, so it slots into agent stacks as the fast inner-loop model while a bigger sibling does the planning.
The honest weakness: hard, multi-step reasoning. When a task needs the model to be right in a way that takes deliberation — legal nuance, subtle code review, research synthesis — step up to Sonnet 4.6 or beyond. Haiku will give you a fast answer, not always a deep one.
| Model | Input / 1M | Output / 1M | Context |
|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | 200K |
| Claude Sonnet 4.6 | $3 | $15 | 200K (1M β) |
| GPT-5.4 mini | $0.25 | $2 | 272K |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | 1M |
| DeepSeek V4-Flash | $0.14 | $0.28 | 1M |
Cross-family, Haiku is the premium budget model: GPT-5.4 mini, Gemini Flash-Lite, and DeepSeek V4-Flash all undercut it on price, some by a lot. You pay Haiku's rate for the Claude qualities — instruction-following, long-context reliability, and drop-in compatibility with a Sonnet/Opus/Fable stack. If your pipeline is provider-agnostic and purely cost-driven, the rivals win; if it's a Claude shop, Haiku is the obvious cheap tier.