Claude Haiku (currently Haiku 4.5) is Anthropic's fastest model, built for high-volume, well-scoped work like classification, extraction and routing, at $1/$5 per million input/output tokens. Pick it for repetitive calls where speed and cost matter; step up to Sonnet the moment a task needs real reasoning.
Other good fits
An application that needs a fast reply and doesn't need deep reasoning on each individual call.
+ Pros
Read pros
Haiku is Anthropic's smallest and fastest current model, described as "the fastest model with near-frontier intelligence." It's built for calls you make a lot of: classifying an incoming message, extracting a field from text, routing a request, drafting something short, where speed and low cost per call matter more than deep reasoning on any single one.
Pick it when…
You're making the same kind of call repeatedly, at volume, and the task is genuinely simple and well-scoped.
You want the fastest response time in the current Claude lineup for a routine task.
You want image input and prompt caching on the cheapest current model, for something like tagging a stream of uploaded photos.
− Cons
Read cons
Haiku's knowledge cutoff (February 2025) is noticeably older than Sonnet's or Opus's, so it can miss more recent facts if you rely on its own knowledge instead of giving it source material. It also doesn't support the adjustable reasoning effort the larger models do, so a task that needs the model to actually think through a problem, rather than pattern-match quickly, will strain it.
Pick something else when…
The task needs real reasoning, not just fast pattern-matching.Claude Sonnet is the reliable everyday step up.
You need a larger context window. Sonnet and Opus both offer 1M tokens against Haiku's 200K.
A single, high-stakes answer matters more than call volume or speed.Claude Opus is built for that instead.
Avoid if
You need a 1M-token context window; Haiku's context is 200K tokens, smaller than the rest of the current Claude lineup.
Context
Haiku 4.5 has a 200K-token context window and a 64K-token maximum output, smaller than Sonnet's or Opus's 1M-token window. Its reliable knowledge cutoff is February 2025, earlier than the rest of the current lineup. Anthropic commits to keeping it available on its own platforms no sooner than 15 October 2026.
Requirements
Haiku is a closed, cloud-hosted model with no self-hosting or offline option. Use it through the Claude API, claude.ai, or a supported cloud platform (Amazon Bedrock, Google Cloud, Microsoft Foundry).
Licence & cost
Prices checked 23 September 2026 (pricing (opens in a new tab)): $1 per million input tokens, $5 per million output tokens, through the Claude API, the lowest rate in the current Claude lineup. Batch requests run at 50% off; prompt-cache reads cost about 10% of the base input price. At high call volume, even Haiku's low per-token rate adds up, so it's worth tracking total spend, not just the per-call price.
First task
Documentation-based suggestion — not run here.
Build a simple classifier: feed it fifty made-up support messages and ask it to label each one as billing, technical or other in a single word. Compare its labels against ones a person assigned to the same messages.
Done when Haiku agrees with the human labels on most of the fifty, and you've identified the specific pattern (a certain phrasing, an ambiguous message) behind any disagreement.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Anthropic's models overview and pricing pages, read on 2026-09-23. No model calls were made for this review; this describes the current Haiku generation (Haiku 4.5) and will need updating when Anthropic ships the next one.
Claude models overview · Sep 23, 2026 — Haiku 4.5 specs: 200K context, 64K max output, February 2025 knowledge cutoff, retirement not before October 2026.
Claude API pricing · Sep 23, 2026 — Haiku 4.5 at $1/$5 per million input/output tokens; batch and cache-read discounts.
Choosing a model · Sep 23, 2026 — Guidance on when Haiku is the right choice versus Sonnet or Opus.
Further review notes & alternatives
Alternatives and trade-offs
Claude Sonnet wins the moment a task needs real reasoning, not just fast classification.
Claude Opus wins for long, demanding agentic work where Haiku would need constant correction.
GPT-mini wins as the closest cross-vendor comparison for cheap, high-volume calls; worth benchmarking both on your actual workload.
Classify, extract, route, and draft short outputs at high request volume.
Choose it for well-scoped repetitive calls with image input and caching.
Its 200K context and batch support fit routine cloud processing.