Haiku 5.5

Claude Haiku — Anthropic API

$1 input / $5 output per million tokens (Anthropic API)

  • $1 / $5
  • 200K Context
  • Fastest
  • High volume

Try Model

Overview

Claude Haiku 5.5 (API id `claude-haiku-5-5`) is Anthropic’s fastest, lowest-cost Claude 5.x SKU: 200K context, low latency, and $1 / $5 list pricing. Use it for classification, short replies, cheap agent steps, and high-volume loops. Escalate to Sonnet 5.5 ($2 / $10) when coding quality or longer briefs matter.

Capabilities

  • Speed firstLowest latency in the Claude 5.x lineup—good for chat UIs, routers, and tight tool loops.
  • High-volume cost controlAt $1 / $5 per million tokens it is half Sonnet 5.5 ($2 / $10) and one-quarter Opus 5.5 ($4 / $20).
  • Prompt cacheCache reads are $0.10 / MTok (10% of input). Reuse a short system prompt across many cheap turns.
  • BatchBatch API is 50% off input and output ($0.50 / $2.50) when the job can wait.

Context Window

Haiku 5.5 is listed at 200K input tokens and up to 64K output—smaller than Sonnet 5.5’s 1M window. Keep briefs short; escalate long repos to Sonnet.

  • Prefer short system prompts and only the files you need.
  • Trim logs before you fill the 200K window.
  • Confirm the current window on Anthropic’s Haiku 5.5 model card.

Pricing

Standard API: $1 / MTok in · $5 / MTok out

These are Anthropic list prices for Claude Haiku 5.5. Hosts (Bedrock, Vertex, OpenRouter, IDEs) may add a markup. MEAGENT Agent Hub uses account credits and is not a Claude invoice.

Input
$1 / million tokens
Output
$5 / million tokens
5-minute cache write
$1.25 / MTok
1-hour cache write
$2 / MTok
Cache read
$0.10 / MTok (0.1× input)
Batch API
$0.50 / $2.50 (50% off in/out)
US-only inference
1.1× input and output where offered
  • Worked example: 20K input + 1K output ≈ $0.02 + $0.005 = $0.025 at list (before cache).
  • Cached follow-up on a 20K system prompt: cache read ≈ $0.002 plus new tokens.
  • Prices change. Re-check Anthropic’s pricing page before you budget.

Anthropic pricing

How to Use

  1. Pin the SKUUse `claude-haiku-5-5` (or your host’s equivalent) so you don’t accidentally bill Sonnet or Opus.
  2. Route the easy stepsUse Haiku for classify → extract → draft stubs; send hard coding and long briefs to Sonnet 5.5.
  3. Cache the static prefixPut stable instructions in a cacheable prefix. Reads are $0.10 / MTok.
  4. Escalate on purposeIf Haiku underperforms, move the task to Sonnet 5.5 ($2 / $10)—don’t leave quality-critical work on the cheapest SKU.

Alternatives